perf: Make common SAMPLE_EVENT parser
[linux-3.10.git] / tools / perf / builtin-sched.c
1 #include "builtin.h"
2 #include "perf.h"
3
4 #include "util/util.h"
5 #include "util/cache.h"
6 #include "util/symbol.h"
7 #include "util/thread.h"
8 #include "util/header.h"
9
10 #include "util/parse-options.h"
11 #include "util/trace-event.h"
12
13 #include "util/debug.h"
14 #include "util/data_map.h"
15
16 #include <sys/types.h>
17 #include <sys/prctl.h>
18
19 #include <semaphore.h>
20 #include <pthread.h>
21 #include <math.h>
22
23 static char                     const *input_name = "perf.data";
24
25 static struct perf_header       *header;
26 static u64                      sample_type;
27
28 static char                     default_sort_order[] = "avg, max, switch, runtime";
29 static char                     *sort_order = default_sort_order;
30
31 static int                      profile_cpu = -1;
32
33 #define PR_SET_NAME             15               /* Set process name */
34 #define MAX_CPUS                4096
35
36 static u64                      run_measurement_overhead;
37 static u64                      sleep_measurement_overhead;
38
39 #define COMM_LEN                20
40 #define SYM_LEN                 129
41
42 #define MAX_PID                 65536
43
44 static unsigned long            nr_tasks;
45
46 struct sched_atom;
47
48 struct task_desc {
49         unsigned long           nr;
50         unsigned long           pid;
51         char                    comm[COMM_LEN];
52
53         unsigned long           nr_events;
54         unsigned long           curr_event;
55         struct sched_atom       **atoms;
56
57         pthread_t               thread;
58         sem_t                   sleep_sem;
59
60         sem_t                   ready_for_work;
61         sem_t                   work_done_sem;
62
63         u64                     cpu_usage;
64 };
65
66 enum sched_event_type {
67         SCHED_EVENT_RUN,
68         SCHED_EVENT_SLEEP,
69         SCHED_EVENT_WAKEUP,
70         SCHED_EVENT_MIGRATION,
71 };
72
73 struct sched_atom {
74         enum sched_event_type   type;
75         u64                     timestamp;
76         u64                     duration;
77         unsigned long           nr;
78         int                     specific_wait;
79         sem_t                   *wait_sem;
80         struct task_desc        *wakee;
81 };
82
83 static struct task_desc         *pid_to_task[MAX_PID];
84
85 static struct task_desc         **tasks;
86
87 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
88 static u64                      start_time;
89
90 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
91
92 static unsigned long            nr_run_events;
93 static unsigned long            nr_sleep_events;
94 static unsigned long            nr_wakeup_events;
95
96 static unsigned long            nr_sleep_corrections;
97 static unsigned long            nr_run_events_optimized;
98
99 static unsigned long            targetless_wakeups;
100 static unsigned long            multitarget_wakeups;
101
102 static u64                      cpu_usage;
103 static u64                      runavg_cpu_usage;
104 static u64                      parent_cpu_usage;
105 static u64                      runavg_parent_cpu_usage;
106
107 static unsigned long            nr_runs;
108 static u64                      sum_runtime;
109 static u64                      sum_fluct;
110 static u64                      run_avg;
111
112 static unsigned long            replay_repeat = 10;
113 static unsigned long            nr_timestamps;
114 static unsigned long            nr_unordered_timestamps;
115 static unsigned long            nr_state_machine_bugs;
116 static unsigned long            nr_context_switch_bugs;
117 static unsigned long            nr_events;
118 static unsigned long            nr_lost_chunks;
119 static unsigned long            nr_lost_events;
120
121 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
122
123 enum thread_state {
124         THREAD_SLEEPING = 0,
125         THREAD_WAIT_CPU,
126         THREAD_SCHED_IN,
127         THREAD_IGNORE
128 };
129
130 struct work_atom {
131         struct list_head        list;
132         enum thread_state       state;
133         u64                     sched_out_time;
134         u64                     wake_up_time;
135         u64                     sched_in_time;
136         u64                     runtime;
137 };
138
139 struct work_atoms {
140         struct list_head        work_list;
141         struct thread           *thread;
142         struct rb_node          node;
143         u64                     max_lat;
144         u64                     total_lat;
145         u64                     nb_atoms;
146         u64                     total_runtime;
147 };
148
149 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
150
151 static struct rb_root           atom_root, sorted_atom_root;
152
153 static u64                      all_runtime;
154 static u64                      all_count;
155
156
157 static u64 get_nsecs(void)
158 {
159         struct timespec ts;
160
161         clock_gettime(CLOCK_MONOTONIC, &ts);
162
163         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
164 }
165
166 static void burn_nsecs(u64 nsecs)
167 {
168         u64 T0 = get_nsecs(), T1;
169
170         do {
171                 T1 = get_nsecs();
172         } while (T1 + run_measurement_overhead < T0 + nsecs);
173 }
174
175 static void sleep_nsecs(u64 nsecs)
176 {
177         struct timespec ts;
178
179         ts.tv_nsec = nsecs % 999999999;
180         ts.tv_sec = nsecs / 999999999;
181
182         nanosleep(&ts, NULL);
183 }
184
185 static void calibrate_run_measurement_overhead(void)
186 {
187         u64 T0, T1, delta, min_delta = 1000000000ULL;
188         int i;
189
190         for (i = 0; i < 10; i++) {
191                 T0 = get_nsecs();
192                 burn_nsecs(0);
193                 T1 = get_nsecs();
194                 delta = T1-T0;
195                 min_delta = min(min_delta, delta);
196         }
197         run_measurement_overhead = min_delta;
198
199         printf("run measurement overhead: %Ld nsecs\n", min_delta);
200 }
201
202 static void calibrate_sleep_measurement_overhead(void)
203 {
204         u64 T0, T1, delta, min_delta = 1000000000ULL;
205         int i;
206
207         for (i = 0; i < 10; i++) {
208                 T0 = get_nsecs();
209                 sleep_nsecs(10000);
210                 T1 = get_nsecs();
211                 delta = T1-T0;
212                 min_delta = min(min_delta, delta);
213         }
214         min_delta -= 10000;
215         sleep_measurement_overhead = min_delta;
216
217         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
218 }
219
220 static struct sched_atom *
221 get_new_event(struct task_desc *task, u64 timestamp)
222 {
223         struct sched_atom *event = zalloc(sizeof(*event));
224         unsigned long idx = task->nr_events;
225         size_t size;
226
227         event->timestamp = timestamp;
228         event->nr = idx;
229
230         task->nr_events++;
231         size = sizeof(struct sched_atom *) * task->nr_events;
232         task->atoms = realloc(task->atoms, size);
233         BUG_ON(!task->atoms);
234
235         task->atoms[idx] = event;
236
237         return event;
238 }
239
240 static struct sched_atom *last_event(struct task_desc *task)
241 {
242         if (!task->nr_events)
243                 return NULL;
244
245         return task->atoms[task->nr_events - 1];
246 }
247
248 static void
249 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
250 {
251         struct sched_atom *event, *curr_event = last_event(task);
252
253         /*
254          * optimize an existing RUN event by merging this one
255          * to it:
256          */
257         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
258                 nr_run_events_optimized++;
259                 curr_event->duration += duration;
260                 return;
261         }
262
263         event = get_new_event(task, timestamp);
264
265         event->type = SCHED_EVENT_RUN;
266         event->duration = duration;
267
268         nr_run_events++;
269 }
270
271 static void
272 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
273                        struct task_desc *wakee)
274 {
275         struct sched_atom *event, *wakee_event;
276
277         event = get_new_event(task, timestamp);
278         event->type = SCHED_EVENT_WAKEUP;
279         event->wakee = wakee;
280
281         wakee_event = last_event(wakee);
282         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
283                 targetless_wakeups++;
284                 return;
285         }
286         if (wakee_event->wait_sem) {
287                 multitarget_wakeups++;
288                 return;
289         }
290
291         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
292         sem_init(wakee_event->wait_sem, 0, 0);
293         wakee_event->specific_wait = 1;
294         event->wait_sem = wakee_event->wait_sem;
295
296         nr_wakeup_events++;
297 }
298
299 static void
300 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
301                       u64 task_state __used)
302 {
303         struct sched_atom *event = get_new_event(task, timestamp);
304
305         event->type = SCHED_EVENT_SLEEP;
306
307         nr_sleep_events++;
308 }
309
310 static struct task_desc *register_pid(unsigned long pid, const char *comm)
311 {
312         struct task_desc *task;
313
314         BUG_ON(pid >= MAX_PID);
315
316         task = pid_to_task[pid];
317
318         if (task)
319                 return task;
320
321         task = zalloc(sizeof(*task));
322         task->pid = pid;
323         task->nr = nr_tasks;
324         strcpy(task->comm, comm);
325         /*
326          * every task starts in sleeping state - this gets ignored
327          * if there's no wakeup pointing to this sleep state:
328          */
329         add_sched_event_sleep(task, 0, 0);
330
331         pid_to_task[pid] = task;
332         nr_tasks++;
333         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
334         BUG_ON(!tasks);
335         tasks[task->nr] = task;
336
337         if (verbose)
338                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
339
340         return task;
341 }
342
343
344 static void print_task_traces(void)
345 {
346         struct task_desc *task;
347         unsigned long i;
348
349         for (i = 0; i < nr_tasks; i++) {
350                 task = tasks[i];
351                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
352                         task->nr, task->comm, task->pid, task->nr_events);
353         }
354 }
355
356 static void add_cross_task_wakeups(void)
357 {
358         struct task_desc *task1, *task2;
359         unsigned long i, j;
360
361         for (i = 0; i < nr_tasks; i++) {
362                 task1 = tasks[i];
363                 j = i + 1;
364                 if (j == nr_tasks)
365                         j = 0;
366                 task2 = tasks[j];
367                 add_sched_event_wakeup(task1, 0, task2);
368         }
369 }
370
371 static void
372 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
373 {
374         int ret = 0;
375         u64 now;
376         long long delta;
377
378         now = get_nsecs();
379         delta = start_time + atom->timestamp - now;
380
381         switch (atom->type) {
382                 case SCHED_EVENT_RUN:
383                         burn_nsecs(atom->duration);
384                         break;
385                 case SCHED_EVENT_SLEEP:
386                         if (atom->wait_sem)
387                                 ret = sem_wait(atom->wait_sem);
388                         BUG_ON(ret);
389                         break;
390                 case SCHED_EVENT_WAKEUP:
391                         if (atom->wait_sem)
392                                 ret = sem_post(atom->wait_sem);
393                         BUG_ON(ret);
394                         break;
395                 case SCHED_EVENT_MIGRATION:
396                         break;
397                 default:
398                         BUG_ON(1);
399         }
400 }
401
402 static u64 get_cpu_usage_nsec_parent(void)
403 {
404         struct rusage ru;
405         u64 sum;
406         int err;
407
408         err = getrusage(RUSAGE_SELF, &ru);
409         BUG_ON(err);
410
411         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
412         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
413
414         return sum;
415 }
416
417 static u64 get_cpu_usage_nsec_self(void)
418 {
419         char filename [] = "/proc/1234567890/sched";
420         unsigned long msecs, nsecs;
421         char *line = NULL;
422         u64 total = 0;
423         size_t len = 0;
424         ssize_t chars;
425         FILE *file;
426         int ret;
427
428         sprintf(filename, "/proc/%d/sched", getpid());
429         file = fopen(filename, "r");
430         BUG_ON(!file);
431
432         while ((chars = getline(&line, &len, file)) != -1) {
433                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
434                         &msecs, &nsecs);
435                 if (ret == 2) {
436                         total = msecs*1e6 + nsecs;
437                         break;
438                 }
439         }
440         if (line)
441                 free(line);
442         fclose(file);
443
444         return total;
445 }
446
447 static void *thread_func(void *ctx)
448 {
449         struct task_desc *this_task = ctx;
450         u64 cpu_usage_0, cpu_usage_1;
451         unsigned long i, ret;
452         char comm2[22];
453
454         sprintf(comm2, ":%s", this_task->comm);
455         prctl(PR_SET_NAME, comm2);
456
457 again:
458         ret = sem_post(&this_task->ready_for_work);
459         BUG_ON(ret);
460         ret = pthread_mutex_lock(&start_work_mutex);
461         BUG_ON(ret);
462         ret = pthread_mutex_unlock(&start_work_mutex);
463         BUG_ON(ret);
464
465         cpu_usage_0 = get_cpu_usage_nsec_self();
466
467         for (i = 0; i < this_task->nr_events; i++) {
468                 this_task->curr_event = i;
469                 process_sched_event(this_task, this_task->atoms[i]);
470         }
471
472         cpu_usage_1 = get_cpu_usage_nsec_self();
473         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
474
475         ret = sem_post(&this_task->work_done_sem);
476         BUG_ON(ret);
477
478         ret = pthread_mutex_lock(&work_done_wait_mutex);
479         BUG_ON(ret);
480         ret = pthread_mutex_unlock(&work_done_wait_mutex);
481         BUG_ON(ret);
482
483         goto again;
484 }
485
486 static void create_tasks(void)
487 {
488         struct task_desc *task;
489         pthread_attr_t attr;
490         unsigned long i;
491         int err;
492
493         err = pthread_attr_init(&attr);
494         BUG_ON(err);
495         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
496         BUG_ON(err);
497         err = pthread_mutex_lock(&start_work_mutex);
498         BUG_ON(err);
499         err = pthread_mutex_lock(&work_done_wait_mutex);
500         BUG_ON(err);
501         for (i = 0; i < nr_tasks; i++) {
502                 task = tasks[i];
503                 sem_init(&task->sleep_sem, 0, 0);
504                 sem_init(&task->ready_for_work, 0, 0);
505                 sem_init(&task->work_done_sem, 0, 0);
506                 task->curr_event = 0;
507                 err = pthread_create(&task->thread, &attr, thread_func, task);
508                 BUG_ON(err);
509         }
510 }
511
512 static void wait_for_tasks(void)
513 {
514         u64 cpu_usage_0, cpu_usage_1;
515         struct task_desc *task;
516         unsigned long i, ret;
517
518         start_time = get_nsecs();
519         cpu_usage = 0;
520         pthread_mutex_unlock(&work_done_wait_mutex);
521
522         for (i = 0; i < nr_tasks; i++) {
523                 task = tasks[i];
524                 ret = sem_wait(&task->ready_for_work);
525                 BUG_ON(ret);
526                 sem_init(&task->ready_for_work, 0, 0);
527         }
528         ret = pthread_mutex_lock(&work_done_wait_mutex);
529         BUG_ON(ret);
530
531         cpu_usage_0 = get_cpu_usage_nsec_parent();
532
533         pthread_mutex_unlock(&start_work_mutex);
534
535         for (i = 0; i < nr_tasks; i++) {
536                 task = tasks[i];
537                 ret = sem_wait(&task->work_done_sem);
538                 BUG_ON(ret);
539                 sem_init(&task->work_done_sem, 0, 0);
540                 cpu_usage += task->cpu_usage;
541                 task->cpu_usage = 0;
542         }
543
544         cpu_usage_1 = get_cpu_usage_nsec_parent();
545         if (!runavg_cpu_usage)
546                 runavg_cpu_usage = cpu_usage;
547         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
548
549         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
550         if (!runavg_parent_cpu_usage)
551                 runavg_parent_cpu_usage = parent_cpu_usage;
552         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
553                                    parent_cpu_usage)/10;
554
555         ret = pthread_mutex_lock(&start_work_mutex);
556         BUG_ON(ret);
557
558         for (i = 0; i < nr_tasks; i++) {
559                 task = tasks[i];
560                 sem_init(&task->sleep_sem, 0, 0);
561                 task->curr_event = 0;
562         }
563 }
564
565 static void run_one_test(void)
566 {
567         u64 T0, T1, delta, avg_delta, fluct, std_dev;
568
569         T0 = get_nsecs();
570         wait_for_tasks();
571         T1 = get_nsecs();
572
573         delta = T1 - T0;
574         sum_runtime += delta;
575         nr_runs++;
576
577         avg_delta = sum_runtime / nr_runs;
578         if (delta < avg_delta)
579                 fluct = avg_delta - delta;
580         else
581                 fluct = delta - avg_delta;
582         sum_fluct += fluct;
583         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
584         if (!run_avg)
585                 run_avg = delta;
586         run_avg = (run_avg*9 + delta)/10;
587
588         printf("#%-3ld: %0.3f, ",
589                 nr_runs, (double)delta/1000000.0);
590
591         printf("ravg: %0.2f, ",
592                 (double)run_avg/1e6);
593
594         printf("cpu: %0.2f / %0.2f",
595                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
596
597 #if 0
598         /*
599          * rusage statistics done by the parent, these are less
600          * accurate than the sum_exec_runtime based statistics:
601          */
602         printf(" [%0.2f / %0.2f]",
603                 (double)parent_cpu_usage/1e6,
604                 (double)runavg_parent_cpu_usage/1e6);
605 #endif
606
607         printf("\n");
608
609         if (nr_sleep_corrections)
610                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
611         nr_sleep_corrections = 0;
612 }
613
614 static void test_calibrations(void)
615 {
616         u64 T0, T1;
617
618         T0 = get_nsecs();
619         burn_nsecs(1e6);
620         T1 = get_nsecs();
621
622         printf("the run test took %Ld nsecs\n", T1-T0);
623
624         T0 = get_nsecs();
625         sleep_nsecs(1e6);
626         T1 = get_nsecs();
627
628         printf("the sleep test took %Ld nsecs\n", T1-T0);
629 }
630
631 struct raw_event_sample {
632         u32 size;
633         char data[0];
634 };
635
636 #define FILL_FIELD(ptr, field, event, data)     \
637         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
638
639 #define FILL_ARRAY(ptr, array, event, data)                     \
640 do {                                                            \
641         void *__array = raw_field_ptr(event, #array, data);     \
642         memcpy(ptr.array, __array, sizeof(ptr.array));  \
643 } while(0)
644
645 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
646 do {                                                            \
647         FILL_FIELD(ptr, common_type, event, data);              \
648         FILL_FIELD(ptr, common_flags, event, data);             \
649         FILL_FIELD(ptr, common_preempt_count, event, data);     \
650         FILL_FIELD(ptr, common_pid, event, data);               \
651         FILL_FIELD(ptr, common_tgid, event, data);              \
652 } while (0)
653
654
655
656 struct trace_switch_event {
657         u32 size;
658
659         u16 common_type;
660         u8 common_flags;
661         u8 common_preempt_count;
662         u32 common_pid;
663         u32 common_tgid;
664
665         char prev_comm[16];
666         u32 prev_pid;
667         u32 prev_prio;
668         u64 prev_state;
669         char next_comm[16];
670         u32 next_pid;
671         u32 next_prio;
672 };
673
674 struct trace_runtime_event {
675         u32 size;
676
677         u16 common_type;
678         u8 common_flags;
679         u8 common_preempt_count;
680         u32 common_pid;
681         u32 common_tgid;
682
683         char comm[16];
684         u32 pid;
685         u64 runtime;
686         u64 vruntime;
687 };
688
689 struct trace_wakeup_event {
690         u32 size;
691
692         u16 common_type;
693         u8 common_flags;
694         u8 common_preempt_count;
695         u32 common_pid;
696         u32 common_tgid;
697
698         char comm[16];
699         u32 pid;
700
701         u32 prio;
702         u32 success;
703         u32 cpu;
704 };
705
706 struct trace_fork_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 parent_comm[16];
716         u32 parent_pid;
717         char child_comm[16];
718         u32 child_pid;
719 };
720
721 struct trace_migrate_task_event {
722         u32 size;
723
724         u16 common_type;
725         u8 common_flags;
726         u8 common_preempt_count;
727         u32 common_pid;
728         u32 common_tgid;
729
730         char comm[16];
731         u32 pid;
732
733         u32 prio;
734         u32 cpu;
735 };
736
737 struct trace_sched_handler {
738         void (*switch_event)(struct trace_switch_event *,
739                              struct event *,
740                              int cpu,
741                              u64 timestamp,
742                              struct thread *thread);
743
744         void (*runtime_event)(struct trace_runtime_event *,
745                               struct event *,
746                               int cpu,
747                               u64 timestamp,
748                               struct thread *thread);
749
750         void (*wakeup_event)(struct trace_wakeup_event *,
751                              struct event *,
752                              int cpu,
753                              u64 timestamp,
754                              struct thread *thread);
755
756         void (*fork_event)(struct trace_fork_event *,
757                            struct event *,
758                            int cpu,
759                            u64 timestamp,
760                            struct thread *thread);
761
762         void (*migrate_task_event)(struct trace_migrate_task_event *,
763                            struct event *,
764                            int cpu,
765                            u64 timestamp,
766                            struct thread *thread);
767 };
768
769
770 static void
771 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
772                     struct event *event,
773                     int cpu __used,
774                     u64 timestamp __used,
775                     struct thread *thread __used)
776 {
777         struct task_desc *waker, *wakee;
778
779         if (verbose) {
780                 printf("sched_wakeup event %p\n", event);
781
782                 printf(" ... pid %d woke up %s/%d\n",
783                         wakeup_event->common_pid,
784                         wakeup_event->comm,
785                         wakeup_event->pid);
786         }
787
788         waker = register_pid(wakeup_event->common_pid, "<unknown>");
789         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
790
791         add_sched_event_wakeup(waker, timestamp, wakee);
792 }
793
794 static u64 cpu_last_switched[MAX_CPUS];
795
796 static void
797 replay_switch_event(struct trace_switch_event *switch_event,
798                     struct event *event,
799                     int cpu,
800                     u64 timestamp,
801                     struct thread *thread __used)
802 {
803         struct task_desc *prev, *next;
804         u64 timestamp0;
805         s64 delta;
806
807         if (verbose)
808                 printf("sched_switch event %p\n", event);
809
810         if (cpu >= MAX_CPUS || cpu < 0)
811                 return;
812
813         timestamp0 = cpu_last_switched[cpu];
814         if (timestamp0)
815                 delta = timestamp - timestamp0;
816         else
817                 delta = 0;
818
819         if (delta < 0)
820                 die("hm, delta: %Ld < 0 ?\n", delta);
821
822         if (verbose) {
823                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
824                         switch_event->prev_comm, switch_event->prev_pid,
825                         switch_event->next_comm, switch_event->next_pid,
826                         delta);
827         }
828
829         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
830         next = register_pid(switch_event->next_pid, switch_event->next_comm);
831
832         cpu_last_switched[cpu] = timestamp;
833
834         add_sched_event_run(prev, timestamp, delta);
835         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
836 }
837
838
839 static void
840 replay_fork_event(struct trace_fork_event *fork_event,
841                   struct event *event,
842                   int cpu __used,
843                   u64 timestamp __used,
844                   struct thread *thread __used)
845 {
846         if (verbose) {
847                 printf("sched_fork event %p\n", event);
848                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
849                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
850         }
851         register_pid(fork_event->parent_pid, fork_event->parent_comm);
852         register_pid(fork_event->child_pid, fork_event->child_comm);
853 }
854
855 static struct trace_sched_handler replay_ops  = {
856         .wakeup_event           = replay_wakeup_event,
857         .switch_event           = replay_switch_event,
858         .fork_event             = replay_fork_event,
859 };
860
861 struct sort_dimension {
862         const char              *name;
863         sort_fn_t               cmp;
864         struct list_head        list;
865 };
866
867 static LIST_HEAD(cmp_pid);
868
869 static int
870 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
871 {
872         struct sort_dimension *sort;
873         int ret = 0;
874
875         BUG_ON(list_empty(list));
876
877         list_for_each_entry(sort, list, list) {
878                 ret = sort->cmp(l, r);
879                 if (ret)
880                         return ret;
881         }
882
883         return ret;
884 }
885
886 static struct work_atoms *
887 thread_atoms_search(struct rb_root *root, struct thread *thread,
888                          struct list_head *sort_list)
889 {
890         struct rb_node *node = root->rb_node;
891         struct work_atoms key = { .thread = thread };
892
893         while (node) {
894                 struct work_atoms *atoms;
895                 int cmp;
896
897                 atoms = container_of(node, struct work_atoms, node);
898
899                 cmp = thread_lat_cmp(sort_list, &key, atoms);
900                 if (cmp > 0)
901                         node = node->rb_left;
902                 else if (cmp < 0)
903                         node = node->rb_right;
904                 else {
905                         BUG_ON(thread != atoms->thread);
906                         return atoms;
907                 }
908         }
909         return NULL;
910 }
911
912 static void
913 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
914                          struct list_head *sort_list)
915 {
916         struct rb_node **new = &(root->rb_node), *parent = NULL;
917
918         while (*new) {
919                 struct work_atoms *this;
920                 int cmp;
921
922                 this = container_of(*new, struct work_atoms, node);
923                 parent = *new;
924
925                 cmp = thread_lat_cmp(sort_list, data, this);
926
927                 if (cmp > 0)
928                         new = &((*new)->rb_left);
929                 else
930                         new = &((*new)->rb_right);
931         }
932
933         rb_link_node(&data->node, parent, new);
934         rb_insert_color(&data->node, root);
935 }
936
937 static void thread_atoms_insert(struct thread *thread)
938 {
939         struct work_atoms *atoms = zalloc(sizeof(*atoms));
940         if (!atoms)
941                 die("No memory");
942
943         atoms->thread = thread;
944         INIT_LIST_HEAD(&atoms->work_list);
945         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
946 }
947
948 static void
949 latency_fork_event(struct trace_fork_event *fork_event __used,
950                    struct event *event __used,
951                    int cpu __used,
952                    u64 timestamp __used,
953                    struct thread *thread __used)
954 {
955         /* should insert the newcomer */
956 }
957
958 __used
959 static char sched_out_state(struct trace_switch_event *switch_event)
960 {
961         const char *str = TASK_STATE_TO_CHAR_STR;
962
963         return str[switch_event->prev_state];
964 }
965
966 static void
967 add_sched_out_event(struct work_atoms *atoms,
968                     char run_state,
969                     u64 timestamp)
970 {
971         struct work_atom *atom = zalloc(sizeof(*atom));
972         if (!atom)
973                 die("Non memory");
974
975         atom->sched_out_time = timestamp;
976
977         if (run_state == 'R') {
978                 atom->state = THREAD_WAIT_CPU;
979                 atom->wake_up_time = atom->sched_out_time;
980         }
981
982         list_add_tail(&atom->list, &atoms->work_list);
983 }
984
985 static void
986 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
987 {
988         struct work_atom *atom;
989
990         BUG_ON(list_empty(&atoms->work_list));
991
992         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
993
994         atom->runtime += delta;
995         atoms->total_runtime += delta;
996 }
997
998 static void
999 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1000 {
1001         struct work_atom *atom;
1002         u64 delta;
1003
1004         if (list_empty(&atoms->work_list))
1005                 return;
1006
1007         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1008
1009         if (atom->state != THREAD_WAIT_CPU)
1010                 return;
1011
1012         if (timestamp < atom->wake_up_time) {
1013                 atom->state = THREAD_IGNORE;
1014                 return;
1015         }
1016
1017         atom->state = THREAD_SCHED_IN;
1018         atom->sched_in_time = timestamp;
1019
1020         delta = atom->sched_in_time - atom->wake_up_time;
1021         atoms->total_lat += delta;
1022         if (delta > atoms->max_lat)
1023                 atoms->max_lat = delta;
1024         atoms->nb_atoms++;
1025 }
1026
1027 static void
1028 latency_switch_event(struct trace_switch_event *switch_event,
1029                      struct event *event __used,
1030                      int cpu,
1031                      u64 timestamp,
1032                      struct thread *thread __used)
1033 {
1034         struct work_atoms *out_events, *in_events;
1035         struct thread *sched_out, *sched_in;
1036         u64 timestamp0;
1037         s64 delta;
1038
1039         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1040
1041         timestamp0 = cpu_last_switched[cpu];
1042         cpu_last_switched[cpu] = timestamp;
1043         if (timestamp0)
1044                 delta = timestamp - timestamp0;
1045         else
1046                 delta = 0;
1047
1048         if (delta < 0)
1049                 die("hm, delta: %Ld < 0 ?\n", delta);
1050
1051
1052         sched_out = threads__findnew(switch_event->prev_pid);
1053         sched_in = threads__findnew(switch_event->next_pid);
1054
1055         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1056         if (!out_events) {
1057                 thread_atoms_insert(sched_out);
1058                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1059                 if (!out_events)
1060                         die("out-event: Internal tree error");
1061         }
1062         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1063
1064         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1065         if (!in_events) {
1066                 thread_atoms_insert(sched_in);
1067                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1068                 if (!in_events)
1069                         die("in-event: Internal tree error");
1070                 /*
1071                  * Take came in we have not heard about yet,
1072                  * add in an initial atom in runnable state:
1073                  */
1074                 add_sched_out_event(in_events, 'R', timestamp);
1075         }
1076         add_sched_in_event(in_events, timestamp);
1077 }
1078
1079 static void
1080 latency_runtime_event(struct trace_runtime_event *runtime_event,
1081                      struct event *event __used,
1082                      int cpu,
1083                      u64 timestamp,
1084                      struct thread *this_thread __used)
1085 {
1086         struct thread *thread = threads__findnew(runtime_event->pid);
1087         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1088
1089         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1090         if (!atoms) {
1091                 thread_atoms_insert(thread);
1092                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1093                 if (!atoms)
1094                         die("in-event: Internal tree error");
1095                 add_sched_out_event(atoms, 'R', timestamp);
1096         }
1097
1098         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1099 }
1100
1101 static void
1102 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1103                      struct event *__event __used,
1104                      int cpu __used,
1105                      u64 timestamp,
1106                      struct thread *thread __used)
1107 {
1108         struct work_atoms *atoms;
1109         struct work_atom *atom;
1110         struct thread *wakee;
1111
1112         /* Note for later, it may be interesting to observe the failing cases */
1113         if (!wakeup_event->success)
1114                 return;
1115
1116         wakee = threads__findnew(wakeup_event->pid);
1117         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1118         if (!atoms) {
1119                 thread_atoms_insert(wakee);
1120                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1121                 if (!atoms)
1122                         die("wakeup-event: Internal tree error");
1123                 add_sched_out_event(atoms, 'S', timestamp);
1124         }
1125
1126         BUG_ON(list_empty(&atoms->work_list));
1127
1128         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1129
1130         /*
1131          * You WILL be missing events if you've recorded only
1132          * one CPU, or are only looking at only one, so don't
1133          * make useless noise.
1134          */
1135         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1136                 nr_state_machine_bugs++;
1137
1138         nr_timestamps++;
1139         if (atom->sched_out_time > timestamp) {
1140                 nr_unordered_timestamps++;
1141                 return;
1142         }
1143
1144         atom->state = THREAD_WAIT_CPU;
1145         atom->wake_up_time = timestamp;
1146 }
1147
1148 static void
1149 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1150                      struct event *__event __used,
1151                      int cpu __used,
1152                      u64 timestamp,
1153                      struct thread *thread __used)
1154 {
1155         struct work_atoms *atoms;
1156         struct work_atom *atom;
1157         struct thread *migrant;
1158
1159         /*
1160          * Only need to worry about migration when profiling one CPU.
1161          */
1162         if (profile_cpu == -1)
1163                 return;
1164
1165         migrant = threads__findnew(migrate_task_event->pid);
1166         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1167         if (!atoms) {
1168                 thread_atoms_insert(migrant);
1169                 register_pid(migrant->pid, migrant->comm);
1170                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1171                 if (!atoms)
1172                         die("migration-event: Internal tree error");
1173                 add_sched_out_event(atoms, 'R', timestamp);
1174         }
1175
1176         BUG_ON(list_empty(&atoms->work_list));
1177
1178         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1179         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1180
1181         nr_timestamps++;
1182
1183         if (atom->sched_out_time > timestamp)
1184                 nr_unordered_timestamps++;
1185 }
1186
1187 static struct trace_sched_handler lat_ops  = {
1188         .wakeup_event           = latency_wakeup_event,
1189         .switch_event           = latency_switch_event,
1190         .runtime_event          = latency_runtime_event,
1191         .fork_event             = latency_fork_event,
1192         .migrate_task_event     = latency_migrate_task_event,
1193 };
1194
1195 static void output_lat_thread(struct work_atoms *work_list)
1196 {
1197         int i;
1198         int ret;
1199         u64 avg;
1200
1201         if (!work_list->nb_atoms)
1202                 return;
1203         /*
1204          * Ignore idle threads:
1205          */
1206         if (!strcmp(work_list->thread->comm, "swapper"))
1207                 return;
1208
1209         all_runtime += work_list->total_runtime;
1210         all_count += work_list->nb_atoms;
1211
1212         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1213
1214         for (i = 0; i < 24 - ret; i++)
1215                 printf(" ");
1216
1217         avg = work_list->total_lat / work_list->nb_atoms;
1218
1219         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1220               (double)work_list->total_runtime / 1e6,
1221                  work_list->nb_atoms, (double)avg / 1e6,
1222                  (double)work_list->max_lat / 1e6);
1223 }
1224
1225 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1226 {
1227         if (l->thread->pid < r->thread->pid)
1228                 return -1;
1229         if (l->thread->pid > r->thread->pid)
1230                 return 1;
1231
1232         return 0;
1233 }
1234
1235 static struct sort_dimension pid_sort_dimension = {
1236         .name                   = "pid",
1237         .cmp                    = pid_cmp,
1238 };
1239
1240 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1241 {
1242         u64 avgl, avgr;
1243
1244         if (!l->nb_atoms)
1245                 return -1;
1246
1247         if (!r->nb_atoms)
1248                 return 1;
1249
1250         avgl = l->total_lat / l->nb_atoms;
1251         avgr = r->total_lat / r->nb_atoms;
1252
1253         if (avgl < avgr)
1254                 return -1;
1255         if (avgl > avgr)
1256                 return 1;
1257
1258         return 0;
1259 }
1260
1261 static struct sort_dimension avg_sort_dimension = {
1262         .name                   = "avg",
1263         .cmp                    = avg_cmp,
1264 };
1265
1266 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1267 {
1268         if (l->max_lat < r->max_lat)
1269                 return -1;
1270         if (l->max_lat > r->max_lat)
1271                 return 1;
1272
1273         return 0;
1274 }
1275
1276 static struct sort_dimension max_sort_dimension = {
1277         .name                   = "max",
1278         .cmp                    = max_cmp,
1279 };
1280
1281 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1282 {
1283         if (l->nb_atoms < r->nb_atoms)
1284                 return -1;
1285         if (l->nb_atoms > r->nb_atoms)
1286                 return 1;
1287
1288         return 0;
1289 }
1290
1291 static struct sort_dimension switch_sort_dimension = {
1292         .name                   = "switch",
1293         .cmp                    = switch_cmp,
1294 };
1295
1296 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1297 {
1298         if (l->total_runtime < r->total_runtime)
1299                 return -1;
1300         if (l->total_runtime > r->total_runtime)
1301                 return 1;
1302
1303         return 0;
1304 }
1305
1306 static struct sort_dimension runtime_sort_dimension = {
1307         .name                   = "runtime",
1308         .cmp                    = runtime_cmp,
1309 };
1310
1311 static struct sort_dimension *available_sorts[] = {
1312         &pid_sort_dimension,
1313         &avg_sort_dimension,
1314         &max_sort_dimension,
1315         &switch_sort_dimension,
1316         &runtime_sort_dimension,
1317 };
1318
1319 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1320
1321 static LIST_HEAD(sort_list);
1322
1323 static int sort_dimension__add(const char *tok, struct list_head *list)
1324 {
1325         int i;
1326
1327         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1328                 if (!strcmp(available_sorts[i]->name, tok)) {
1329                         list_add_tail(&available_sorts[i]->list, list);
1330
1331                         return 0;
1332                 }
1333         }
1334
1335         return -1;
1336 }
1337
1338 static void setup_sorting(void);
1339
1340 static void sort_lat(void)
1341 {
1342         struct rb_node *node;
1343
1344         for (;;) {
1345                 struct work_atoms *data;
1346                 node = rb_first(&atom_root);
1347                 if (!node)
1348                         break;
1349
1350                 rb_erase(node, &atom_root);
1351                 data = rb_entry(node, struct work_atoms, node);
1352                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1353         }
1354 }
1355
1356 static struct trace_sched_handler *trace_handler;
1357
1358 static void
1359 process_sched_wakeup_event(struct raw_event_sample *raw,
1360                            struct event *event,
1361                            int cpu __used,
1362                            u64 timestamp __used,
1363                            struct thread *thread __used)
1364 {
1365         struct trace_wakeup_event wakeup_event;
1366
1367         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1368
1369         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1370         FILL_FIELD(wakeup_event, pid, event, raw->data);
1371         FILL_FIELD(wakeup_event, prio, event, raw->data);
1372         FILL_FIELD(wakeup_event, success, event, raw->data);
1373         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1374
1375         if (trace_handler->wakeup_event)
1376                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1377 }
1378
1379 /*
1380  * Track the current task - that way we can know whether there's any
1381  * weird events, such as a task being switched away that is not current.
1382  */
1383 static int max_cpu;
1384
1385 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1386
1387 static struct thread *curr_thread[MAX_CPUS];
1388
1389 static char next_shortname1 = 'A';
1390 static char next_shortname2 = '0';
1391
1392 static void
1393 map_switch_event(struct trace_switch_event *switch_event,
1394                  struct event *event __used,
1395                  int this_cpu,
1396                  u64 timestamp,
1397                  struct thread *thread __used)
1398 {
1399         struct thread *sched_out, *sched_in;
1400         int new_shortname;
1401         u64 timestamp0;
1402         s64 delta;
1403         int cpu;
1404
1405         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1406
1407         if (this_cpu > max_cpu)
1408                 max_cpu = this_cpu;
1409
1410         timestamp0 = cpu_last_switched[this_cpu];
1411         cpu_last_switched[this_cpu] = timestamp;
1412         if (timestamp0)
1413                 delta = timestamp - timestamp0;
1414         else
1415                 delta = 0;
1416
1417         if (delta < 0)
1418                 die("hm, delta: %Ld < 0 ?\n", delta);
1419
1420
1421         sched_out = threads__findnew(switch_event->prev_pid);
1422         sched_in = threads__findnew(switch_event->next_pid);
1423
1424         curr_thread[this_cpu] = sched_in;
1425
1426         printf("  ");
1427
1428         new_shortname = 0;
1429         if (!sched_in->shortname[0]) {
1430                 sched_in->shortname[0] = next_shortname1;
1431                 sched_in->shortname[1] = next_shortname2;
1432
1433                 if (next_shortname1 < 'Z') {
1434                         next_shortname1++;
1435                 } else {
1436                         next_shortname1='A';
1437                         if (next_shortname2 < '9') {
1438                                 next_shortname2++;
1439                         } else {
1440                                 next_shortname2='0';
1441                         }
1442                 }
1443                 new_shortname = 1;
1444         }
1445
1446         for (cpu = 0; cpu <= max_cpu; cpu++) {
1447                 if (cpu != this_cpu)
1448                         printf(" ");
1449                 else
1450                         printf("*");
1451
1452                 if (curr_thread[cpu]) {
1453                         if (curr_thread[cpu]->pid)
1454                                 printf("%2s ", curr_thread[cpu]->shortname);
1455                         else
1456                                 printf(".  ");
1457                 } else
1458                         printf("   ");
1459         }
1460
1461         printf("  %12.6f secs ", (double)timestamp/1e9);
1462         if (new_shortname) {
1463                 printf("%s => %s:%d\n",
1464                         sched_in->shortname, sched_in->comm, sched_in->pid);
1465         } else {
1466                 printf("\n");
1467         }
1468 }
1469
1470
1471 static void
1472 process_sched_switch_event(struct raw_event_sample *raw,
1473                            struct event *event,
1474                            int this_cpu,
1475                            u64 timestamp __used,
1476                            struct thread *thread __used)
1477 {
1478         struct trace_switch_event switch_event;
1479
1480         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1481
1482         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1483         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1484         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1485         FILL_FIELD(switch_event, prev_state, event, raw->data);
1486         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1487         FILL_FIELD(switch_event, next_pid, event, raw->data);
1488         FILL_FIELD(switch_event, next_prio, event, raw->data);
1489
1490         if (curr_pid[this_cpu] != (u32)-1) {
1491                 /*
1492                  * Are we trying to switch away a PID that is
1493                  * not current?
1494                  */
1495                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1496                         nr_context_switch_bugs++;
1497         }
1498         if (trace_handler->switch_event)
1499                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1500
1501         curr_pid[this_cpu] = switch_event.next_pid;
1502 }
1503
1504 static void
1505 process_sched_runtime_event(struct raw_event_sample *raw,
1506                            struct event *event,
1507                            int cpu __used,
1508                            u64 timestamp __used,
1509                            struct thread *thread __used)
1510 {
1511         struct trace_runtime_event runtime_event;
1512
1513         FILL_ARRAY(runtime_event, comm, event, raw->data);
1514         FILL_FIELD(runtime_event, pid, event, raw->data);
1515         FILL_FIELD(runtime_event, runtime, event, raw->data);
1516         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1517
1518         if (trace_handler->runtime_event)
1519                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1520 }
1521
1522 static void
1523 process_sched_fork_event(struct raw_event_sample *raw,
1524                          struct event *event,
1525                          int cpu __used,
1526                          u64 timestamp __used,
1527                          struct thread *thread __used)
1528 {
1529         struct trace_fork_event fork_event;
1530
1531         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1532
1533         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1534         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1535         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1536         FILL_FIELD(fork_event, child_pid, event, raw->data);
1537
1538         if (trace_handler->fork_event)
1539                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1540 }
1541
1542 static void
1543 process_sched_exit_event(struct event *event,
1544                          int cpu __used,
1545                          u64 timestamp __used,
1546                          struct thread *thread __used)
1547 {
1548         if (verbose)
1549                 printf("sched_exit event %p\n", event);
1550 }
1551
1552 static void
1553 process_sched_migrate_task_event(struct raw_event_sample *raw,
1554                            struct event *event,
1555                            int cpu __used,
1556                            u64 timestamp __used,
1557                            struct thread *thread __used)
1558 {
1559         struct trace_migrate_task_event migrate_task_event;
1560
1561         FILL_COMMON_FIELDS(migrate_task_event, event, raw->data);
1562
1563         FILL_ARRAY(migrate_task_event, comm, event, raw->data);
1564         FILL_FIELD(migrate_task_event, pid, event, raw->data);
1565         FILL_FIELD(migrate_task_event, prio, event, raw->data);
1566         FILL_FIELD(migrate_task_event, cpu, event, raw->data);
1567
1568         if (trace_handler->migrate_task_event)
1569                 trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
1570 }
1571
1572 static void
1573 process_raw_event(event_t *raw_event __used, void *more_data,
1574                   int cpu, u64 timestamp, struct thread *thread)
1575 {
1576         struct raw_event_sample *raw = more_data;
1577         struct event *event;
1578         int type;
1579
1580         type = trace_parse_common_type(raw->data);
1581         event = trace_find_event(type);
1582
1583         if (!strcmp(event->name, "sched_switch"))
1584                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1585         if (!strcmp(event->name, "sched_stat_runtime"))
1586                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1587         if (!strcmp(event->name, "sched_wakeup"))
1588                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1589         if (!strcmp(event->name, "sched_wakeup_new"))
1590                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1591         if (!strcmp(event->name, "sched_process_fork"))
1592                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1593         if (!strcmp(event->name, "sched_process_exit"))
1594                 process_sched_exit_event(event, cpu, timestamp, thread);
1595         if (!strcmp(event->name, "sched_migrate_task"))
1596                 process_sched_migrate_task_event(raw, event, cpu, timestamp, thread);
1597 }
1598
1599 static int process_sample_event(event_t *event)
1600 {
1601         struct sample_data data;
1602         struct thread *thread;
1603
1604         if (!(sample_type & PERF_SAMPLE_RAW))
1605                 return 0;
1606
1607         memset(&data, 0, sizeof(data));
1608         data.time = -1;
1609         data.cpu = -1;
1610         data.period = -1;
1611
1612         event__parse_sample(event, sample_type, &data);
1613
1614         dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
1615                 event->header.misc,
1616                 data.pid, data.tid,
1617                 (void *)(long)data.ip,
1618                 (long long)data.period);
1619
1620         thread = threads__findnew(data.pid);
1621         if (thread == NULL) {
1622                 pr_debug("problem processing %d event, skipping it.\n",
1623                          event->header.type);
1624                 return -1;
1625         }
1626
1627         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1628
1629         if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
1630                 return 0;
1631
1632         process_raw_event(event, data.raw_data, data.cpu, data.time, thread);
1633
1634         return 0;
1635 }
1636
1637 static int process_lost_event(event_t *event __used)
1638 {
1639         nr_lost_chunks++;
1640         nr_lost_events += event->lost.lost;
1641
1642         return 0;
1643 }
1644
1645 static int sample_type_check(u64 type)
1646 {
1647         sample_type = type;
1648
1649         if (!(sample_type & PERF_SAMPLE_RAW)) {
1650                 fprintf(stderr,
1651                         "No trace sample to read. Did you call perf record "
1652                         "without -R?");
1653                 return -1;
1654         }
1655
1656         return 0;
1657 }
1658
1659 static struct perf_file_handler file_handler = {
1660         .process_sample_event   = process_sample_event,
1661         .process_comm_event     = event__process_comm,
1662         .process_lost_event     = process_lost_event,
1663         .sample_type_check      = sample_type_check,
1664 };
1665
1666 static int read_events(void)
1667 {
1668         register_idle_thread();
1669         register_perf_file_handler(&file_handler);
1670
1671         return mmap_dispatch_perf_file(&header, input_name, 0, 0,
1672                                        &event__cwdlen, &event__cwd);
1673 }
1674
1675 static void print_bad_events(void)
1676 {
1677         if (nr_unordered_timestamps && nr_timestamps) {
1678                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1679                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1680                         nr_unordered_timestamps, nr_timestamps);
1681         }
1682         if (nr_lost_events && nr_events) {
1683                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1684                         (double)nr_lost_events/(double)nr_events*100.0,
1685                         nr_lost_events, nr_events, nr_lost_chunks);
1686         }
1687         if (nr_state_machine_bugs && nr_timestamps) {
1688                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1689                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1690                         nr_state_machine_bugs, nr_timestamps);
1691                 if (nr_lost_events)
1692                         printf(" (due to lost events?)");
1693                 printf("\n");
1694         }
1695         if (nr_context_switch_bugs && nr_timestamps) {
1696                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1697                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1698                         nr_context_switch_bugs, nr_timestamps);
1699                 if (nr_lost_events)
1700                         printf(" (due to lost events?)");
1701                 printf("\n");
1702         }
1703 }
1704
1705 static void __cmd_lat(void)
1706 {
1707         struct rb_node *next;
1708
1709         setup_pager();
1710         read_events();
1711         sort_lat();
1712
1713         printf("\n -----------------------------------------------------------------------------------------\n");
1714         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1715         printf(" -----------------------------------------------------------------------------------------\n");
1716
1717         next = rb_first(&sorted_atom_root);
1718
1719         while (next) {
1720                 struct work_atoms *work_list;
1721
1722                 work_list = rb_entry(next, struct work_atoms, node);
1723                 output_lat_thread(work_list);
1724                 next = rb_next(next);
1725         }
1726
1727         printf(" -----------------------------------------------------------------------------------------\n");
1728         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1729                 (double)all_runtime/1e6, all_count);
1730
1731         printf(" ---------------------------------------------------\n");
1732
1733         print_bad_events();
1734         printf("\n");
1735
1736 }
1737
1738 static struct trace_sched_handler map_ops  = {
1739         .wakeup_event           = NULL,
1740         .switch_event           = map_switch_event,
1741         .runtime_event          = NULL,
1742         .fork_event             = NULL,
1743 };
1744
1745 static void __cmd_map(void)
1746 {
1747         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1748
1749         setup_pager();
1750         read_events();
1751         print_bad_events();
1752 }
1753
1754 static void __cmd_replay(void)
1755 {
1756         unsigned long i;
1757
1758         calibrate_run_measurement_overhead();
1759         calibrate_sleep_measurement_overhead();
1760
1761         test_calibrations();
1762
1763         read_events();
1764
1765         printf("nr_run_events:        %ld\n", nr_run_events);
1766         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1767         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1768
1769         if (targetless_wakeups)
1770                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1771         if (multitarget_wakeups)
1772                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1773         if (nr_run_events_optimized)
1774                 printf("run atoms optimized: %ld\n",
1775                         nr_run_events_optimized);
1776
1777         print_task_traces();
1778         add_cross_task_wakeups();
1779
1780         create_tasks();
1781         printf("------------------------------------------------------------\n");
1782         for (i = 0; i < replay_repeat; i++)
1783                 run_one_test();
1784 }
1785
1786
1787 static const char * const sched_usage[] = {
1788         "perf sched [<options>] {record|latency|map|replay|trace}",
1789         NULL
1790 };
1791
1792 static const struct option sched_options[] = {
1793         OPT_STRING('i', "input", &input_name, "file",
1794                     "input file name"),
1795         OPT_BOOLEAN('v', "verbose", &verbose,
1796                     "be more verbose (show symbol address, etc)"),
1797         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1798                     "dump raw trace in ASCII"),
1799         OPT_END()
1800 };
1801
1802 static const char * const latency_usage[] = {
1803         "perf sched latency [<options>]",
1804         NULL
1805 };
1806
1807 static const struct option latency_options[] = {
1808         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1809                    "sort by key(s): runtime, switch, avg, max"),
1810         OPT_BOOLEAN('v', "verbose", &verbose,
1811                     "be more verbose (show symbol address, etc)"),
1812         OPT_INTEGER('C', "CPU", &profile_cpu,
1813                     "CPU to profile on"),
1814         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1815                     "dump raw trace in ASCII"),
1816         OPT_END()
1817 };
1818
1819 static const char * const replay_usage[] = {
1820         "perf sched replay [<options>]",
1821         NULL
1822 };
1823
1824 static const struct option replay_options[] = {
1825         OPT_INTEGER('r', "repeat", &replay_repeat,
1826                     "repeat the workload replay N times (-1: infinite)"),
1827         OPT_BOOLEAN('v', "verbose", &verbose,
1828                     "be more verbose (show symbol address, etc)"),
1829         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1830                     "dump raw trace in ASCII"),
1831         OPT_END()
1832 };
1833
1834 static void setup_sorting(void)
1835 {
1836         char *tmp, *tok, *str = strdup(sort_order);
1837
1838         for (tok = strtok_r(str, ", ", &tmp);
1839                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1840                 if (sort_dimension__add(tok, &sort_list) < 0) {
1841                         error("Unknown --sort key: `%s'", tok);
1842                         usage_with_options(latency_usage, latency_options);
1843                 }
1844         }
1845
1846         free(str);
1847
1848         sort_dimension__add("pid", &cmp_pid);
1849 }
1850
1851 static const char *record_args[] = {
1852         "record",
1853         "-a",
1854         "-R",
1855         "-M",
1856         "-f",
1857         "-m", "1024",
1858         "-c", "1",
1859         "-e", "sched:sched_switch:r",
1860         "-e", "sched:sched_stat_wait:r",
1861         "-e", "sched:sched_stat_sleep:r",
1862         "-e", "sched:sched_stat_iowait:r",
1863         "-e", "sched:sched_stat_runtime:r",
1864         "-e", "sched:sched_process_exit:r",
1865         "-e", "sched:sched_process_fork:r",
1866         "-e", "sched:sched_wakeup:r",
1867         "-e", "sched:sched_migrate_task:r",
1868 };
1869
1870 static int __cmd_record(int argc, const char **argv)
1871 {
1872         unsigned int rec_argc, i, j;
1873         const char **rec_argv;
1874
1875         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1876         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1877
1878         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1879                 rec_argv[i] = strdup(record_args[i]);
1880
1881         for (j = 1; j < (unsigned int)argc; j++, i++)
1882                 rec_argv[i] = argv[j];
1883
1884         BUG_ON(i != rec_argc);
1885
1886         return cmd_record(i, rec_argv, NULL);
1887 }
1888
1889 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1890 {
1891         symbol__init(0);
1892
1893         argc = parse_options(argc, argv, sched_options, sched_usage,
1894                              PARSE_OPT_STOP_AT_NON_OPTION);
1895         if (!argc)
1896                 usage_with_options(sched_usage, sched_options);
1897
1898         if (!strncmp(argv[0], "rec", 3)) {
1899                 return __cmd_record(argc, argv);
1900         } else if (!strncmp(argv[0], "lat", 3)) {
1901                 trace_handler = &lat_ops;
1902                 if (argc > 1) {
1903                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1904                         if (argc)
1905                                 usage_with_options(latency_usage, latency_options);
1906                 }
1907                 setup_sorting();
1908                 __cmd_lat();
1909         } else if (!strcmp(argv[0], "map")) {
1910                 trace_handler = &map_ops;
1911                 setup_sorting();
1912                 __cmd_map();
1913         } else if (!strncmp(argv[0], "rep", 3)) {
1914                 trace_handler = &replay_ops;
1915                 if (argc) {
1916                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1917                         if (argc)
1918                                 usage_with_options(replay_usage, replay_options);
1919                 }
1920                 __cmd_replay();
1921         } else if (!strcmp(argv[0], "trace")) {
1922                 /*
1923                  * Aliased to 'perf trace' for now:
1924                  */
1925                 return cmd_trace(argc, argv, prefix);
1926         } else {
1927                 usage_with_options(sched_usage, sched_options);
1928         }
1929
1930         return 0;
1931 }