perf session: Move the global threads list to perf_session
[linux-2.6.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 #include "util/session.h"
10
11 #include "util/parse-options.h"
12 #include "util/trace-event.h"
13
14 #include "util/debug.h"
15
16 #include <sys/prctl.h>
17
18 #include <semaphore.h>
19 #include <pthread.h>
20 #include <math.h>
21
22 static char                     const *input_name = "perf.data";
23
24 static u64                      sample_type;
25
26 static char                     default_sort_order[] = "avg, max, switch, runtime";
27 static char                     *sort_order = default_sort_order;
28
29 static int                      profile_cpu = -1;
30
31 #define PR_SET_NAME             15               /* Set process name */
32 #define MAX_CPUS                4096
33
34 static u64                      run_measurement_overhead;
35 static u64                      sleep_measurement_overhead;
36
37 #define COMM_LEN                20
38 #define SYM_LEN                 129
39
40 #define MAX_PID                 65536
41
42 static unsigned long            nr_tasks;
43
44 struct sched_atom;
45
46 struct task_desc {
47         unsigned long           nr;
48         unsigned long           pid;
49         char                    comm[COMM_LEN];
50
51         unsigned long           nr_events;
52         unsigned long           curr_event;
53         struct sched_atom       **atoms;
54
55         pthread_t               thread;
56         sem_t                   sleep_sem;
57
58         sem_t                   ready_for_work;
59         sem_t                   work_done_sem;
60
61         u64                     cpu_usage;
62 };
63
64 enum sched_event_type {
65         SCHED_EVENT_RUN,
66         SCHED_EVENT_SLEEP,
67         SCHED_EVENT_WAKEUP,
68         SCHED_EVENT_MIGRATION,
69 };
70
71 struct sched_atom {
72         enum sched_event_type   type;
73         u64                     timestamp;
74         u64                     duration;
75         unsigned long           nr;
76         int                     specific_wait;
77         sem_t                   *wait_sem;
78         struct task_desc        *wakee;
79 };
80
81 static struct task_desc         *pid_to_task[MAX_PID];
82
83 static struct task_desc         **tasks;
84
85 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
86 static u64                      start_time;
87
88 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
89
90 static unsigned long            nr_run_events;
91 static unsigned long            nr_sleep_events;
92 static unsigned long            nr_wakeup_events;
93
94 static unsigned long            nr_sleep_corrections;
95 static unsigned long            nr_run_events_optimized;
96
97 static unsigned long            targetless_wakeups;
98 static unsigned long            multitarget_wakeups;
99
100 static u64                      cpu_usage;
101 static u64                      runavg_cpu_usage;
102 static u64                      parent_cpu_usage;
103 static u64                      runavg_parent_cpu_usage;
104
105 static unsigned long            nr_runs;
106 static u64                      sum_runtime;
107 static u64                      sum_fluct;
108 static u64                      run_avg;
109
110 static unsigned long            replay_repeat = 10;
111 static unsigned long            nr_timestamps;
112 static unsigned long            nr_unordered_timestamps;
113 static unsigned long            nr_state_machine_bugs;
114 static unsigned long            nr_context_switch_bugs;
115 static unsigned long            nr_events;
116 static unsigned long            nr_lost_chunks;
117 static unsigned long            nr_lost_events;
118
119 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
120
121 enum thread_state {
122         THREAD_SLEEPING = 0,
123         THREAD_WAIT_CPU,
124         THREAD_SCHED_IN,
125         THREAD_IGNORE
126 };
127
128 struct work_atom {
129         struct list_head        list;
130         enum thread_state       state;
131         u64                     sched_out_time;
132         u64                     wake_up_time;
133         u64                     sched_in_time;
134         u64                     runtime;
135 };
136
137 struct work_atoms {
138         struct list_head        work_list;
139         struct thread           *thread;
140         struct rb_node          node;
141         u64                     max_lat;
142         u64                     max_lat_at;
143         u64                     total_lat;
144         u64                     nb_atoms;
145         u64                     total_runtime;
146 };
147
148 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
149
150 static struct rb_root           atom_root, sorted_atom_root;
151
152 static u64                      all_runtime;
153 static u64                      all_count;
154
155
156 static u64 get_nsecs(void)
157 {
158         struct timespec ts;
159
160         clock_gettime(CLOCK_MONOTONIC, &ts);
161
162         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
163 }
164
165 static void burn_nsecs(u64 nsecs)
166 {
167         u64 T0 = get_nsecs(), T1;
168
169         do {
170                 T1 = get_nsecs();
171         } while (T1 + run_measurement_overhead < T0 + nsecs);
172 }
173
174 static void sleep_nsecs(u64 nsecs)
175 {
176         struct timespec ts;
177
178         ts.tv_nsec = nsecs % 999999999;
179         ts.tv_sec = nsecs / 999999999;
180
181         nanosleep(&ts, NULL);
182 }
183
184 static void calibrate_run_measurement_overhead(void)
185 {
186         u64 T0, T1, delta, min_delta = 1000000000ULL;
187         int i;
188
189         for (i = 0; i < 10; i++) {
190                 T0 = get_nsecs();
191                 burn_nsecs(0);
192                 T1 = get_nsecs();
193                 delta = T1-T0;
194                 min_delta = min(min_delta, delta);
195         }
196         run_measurement_overhead = min_delta;
197
198         printf("run measurement overhead: %Ld nsecs\n", min_delta);
199 }
200
201 static void calibrate_sleep_measurement_overhead(void)
202 {
203         u64 T0, T1, delta, min_delta = 1000000000ULL;
204         int i;
205
206         for (i = 0; i < 10; i++) {
207                 T0 = get_nsecs();
208                 sleep_nsecs(10000);
209                 T1 = get_nsecs();
210                 delta = T1-T0;
211                 min_delta = min(min_delta, delta);
212         }
213         min_delta -= 10000;
214         sleep_measurement_overhead = min_delta;
215
216         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
217 }
218
219 static struct sched_atom *
220 get_new_event(struct task_desc *task, u64 timestamp)
221 {
222         struct sched_atom *event = zalloc(sizeof(*event));
223         unsigned long idx = task->nr_events;
224         size_t size;
225
226         event->timestamp = timestamp;
227         event->nr = idx;
228
229         task->nr_events++;
230         size = sizeof(struct sched_atom *) * task->nr_events;
231         task->atoms = realloc(task->atoms, size);
232         BUG_ON(!task->atoms);
233
234         task->atoms[idx] = event;
235
236         return event;
237 }
238
239 static struct sched_atom *last_event(struct task_desc *task)
240 {
241         if (!task->nr_events)
242                 return NULL;
243
244         return task->atoms[task->nr_events - 1];
245 }
246
247 static void
248 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
249 {
250         struct sched_atom *event, *curr_event = last_event(task);
251
252         /*
253          * optimize an existing RUN event by merging this one
254          * to it:
255          */
256         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
257                 nr_run_events_optimized++;
258                 curr_event->duration += duration;
259                 return;
260         }
261
262         event = get_new_event(task, timestamp);
263
264         event->type = SCHED_EVENT_RUN;
265         event->duration = duration;
266
267         nr_run_events++;
268 }
269
270 static void
271 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
272                        struct task_desc *wakee)
273 {
274         struct sched_atom *event, *wakee_event;
275
276         event = get_new_event(task, timestamp);
277         event->type = SCHED_EVENT_WAKEUP;
278         event->wakee = wakee;
279
280         wakee_event = last_event(wakee);
281         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
282                 targetless_wakeups++;
283                 return;
284         }
285         if (wakee_event->wait_sem) {
286                 multitarget_wakeups++;
287                 return;
288         }
289
290         wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
291         sem_init(wakee_event->wait_sem, 0, 0);
292         wakee_event->specific_wait = 1;
293         event->wait_sem = wakee_event->wait_sem;
294
295         nr_wakeup_events++;
296 }
297
298 static void
299 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
300                       u64 task_state __used)
301 {
302         struct sched_atom *event = get_new_event(task, timestamp);
303
304         event->type = SCHED_EVENT_SLEEP;
305
306         nr_sleep_events++;
307 }
308
309 static struct task_desc *register_pid(unsigned long pid, const char *comm)
310 {
311         struct task_desc *task;
312
313         BUG_ON(pid >= MAX_PID);
314
315         task = pid_to_task[pid];
316
317         if (task)
318                 return task;
319
320         task = zalloc(sizeof(*task));
321         task->pid = pid;
322         task->nr = nr_tasks;
323         strcpy(task->comm, comm);
324         /*
325          * every task starts in sleeping state - this gets ignored
326          * if there's no wakeup pointing to this sleep state:
327          */
328         add_sched_event_sleep(task, 0, 0);
329
330         pid_to_task[pid] = task;
331         nr_tasks++;
332         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
333         BUG_ON(!tasks);
334         tasks[task->nr] = task;
335
336         if (verbose)
337                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
338
339         return task;
340 }
341
342
343 static void print_task_traces(void)
344 {
345         struct task_desc *task;
346         unsigned long i;
347
348         for (i = 0; i < nr_tasks; i++) {
349                 task = tasks[i];
350                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
351                         task->nr, task->comm, task->pid, task->nr_events);
352         }
353 }
354
355 static void add_cross_task_wakeups(void)
356 {
357         struct task_desc *task1, *task2;
358         unsigned long i, j;
359
360         for (i = 0; i < nr_tasks; i++) {
361                 task1 = tasks[i];
362                 j = i + 1;
363                 if (j == nr_tasks)
364                         j = 0;
365                 task2 = tasks[j];
366                 add_sched_event_wakeup(task1, 0, task2);
367         }
368 }
369
370 static void
371 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
372 {
373         int ret = 0;
374         u64 now;
375         long long delta;
376
377         now = get_nsecs();
378         delta = start_time + atom->timestamp - now;
379
380         switch (atom->type) {
381                 case SCHED_EVENT_RUN:
382                         burn_nsecs(atom->duration);
383                         break;
384                 case SCHED_EVENT_SLEEP:
385                         if (atom->wait_sem)
386                                 ret = sem_wait(atom->wait_sem);
387                         BUG_ON(ret);
388                         break;
389                 case SCHED_EVENT_WAKEUP:
390                         if (atom->wait_sem)
391                                 ret = sem_post(atom->wait_sem);
392                         BUG_ON(ret);
393                         break;
394                 case SCHED_EVENT_MIGRATION:
395                         break;
396                 default:
397                         BUG_ON(1);
398         }
399 }
400
401 static u64 get_cpu_usage_nsec_parent(void)
402 {
403         struct rusage ru;
404         u64 sum;
405         int err;
406
407         err = getrusage(RUSAGE_SELF, &ru);
408         BUG_ON(err);
409
410         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
411         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
412
413         return sum;
414 }
415
416 static int self_open_counters(void)
417 {
418         struct perf_event_attr attr;
419         int fd;
420
421         memset(&attr, 0, sizeof(attr));
422
423         attr.type = PERF_TYPE_SOFTWARE;
424         attr.config = PERF_COUNT_SW_TASK_CLOCK;
425
426         fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
427
428         if (fd < 0)
429                 die("Error: sys_perf_event_open() syscall returned"
430                     "with %d (%s)\n", fd, strerror(errno));
431         return fd;
432 }
433
434 static u64 get_cpu_usage_nsec_self(int fd)
435 {
436         u64 runtime;
437         int ret;
438
439         ret = read(fd, &runtime, sizeof(runtime));
440         BUG_ON(ret != sizeof(runtime));
441
442         return runtime;
443 }
444
445 static void *thread_func(void *ctx)
446 {
447         struct task_desc *this_task = ctx;
448         u64 cpu_usage_0, cpu_usage_1;
449         unsigned long i, ret;
450         char comm2[22];
451         int fd;
452
453         sprintf(comm2, ":%s", this_task->comm);
454         prctl(PR_SET_NAME, comm2);
455         fd = self_open_counters();
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(fd);
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(fd);
473         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
474         ret = sem_post(&this_task->work_done_sem);
475         BUG_ON(ret);
476
477         ret = pthread_mutex_lock(&work_done_wait_mutex);
478         BUG_ON(ret);
479         ret = pthread_mutex_unlock(&work_done_wait_mutex);
480         BUG_ON(ret);
481
482         goto again;
483 }
484
485 static void create_tasks(void)
486 {
487         struct task_desc *task;
488         pthread_attr_t attr;
489         unsigned long i;
490         int err;
491
492         err = pthread_attr_init(&attr);
493         BUG_ON(err);
494         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
495         BUG_ON(err);
496         err = pthread_mutex_lock(&start_work_mutex);
497         BUG_ON(err);
498         err = pthread_mutex_lock(&work_done_wait_mutex);
499         BUG_ON(err);
500         for (i = 0; i < nr_tasks; i++) {
501                 task = tasks[i];
502                 sem_init(&task->sleep_sem, 0, 0);
503                 sem_init(&task->ready_for_work, 0, 0);
504                 sem_init(&task->work_done_sem, 0, 0);
505                 task->curr_event = 0;
506                 err = pthread_create(&task->thread, &attr, thread_func, task);
507                 BUG_ON(err);
508         }
509 }
510
511 static void wait_for_tasks(void)
512 {
513         u64 cpu_usage_0, cpu_usage_1;
514         struct task_desc *task;
515         unsigned long i, ret;
516
517         start_time = get_nsecs();
518         cpu_usage = 0;
519         pthread_mutex_unlock(&work_done_wait_mutex);
520
521         for (i = 0; i < nr_tasks; i++) {
522                 task = tasks[i];
523                 ret = sem_wait(&task->ready_for_work);
524                 BUG_ON(ret);
525                 sem_init(&task->ready_for_work, 0, 0);
526         }
527         ret = pthread_mutex_lock(&work_done_wait_mutex);
528         BUG_ON(ret);
529
530         cpu_usage_0 = get_cpu_usage_nsec_parent();
531
532         pthread_mutex_unlock(&start_work_mutex);
533
534         for (i = 0; i < nr_tasks; i++) {
535                 task = tasks[i];
536                 ret = sem_wait(&task->work_done_sem);
537                 BUG_ON(ret);
538                 sem_init(&task->work_done_sem, 0, 0);
539                 cpu_usage += task->cpu_usage;
540                 task->cpu_usage = 0;
541         }
542
543         cpu_usage_1 = get_cpu_usage_nsec_parent();
544         if (!runavg_cpu_usage)
545                 runavg_cpu_usage = cpu_usage;
546         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
547
548         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
549         if (!runavg_parent_cpu_usage)
550                 runavg_parent_cpu_usage = parent_cpu_usage;
551         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
552                                    parent_cpu_usage)/10;
553
554         ret = pthread_mutex_lock(&start_work_mutex);
555         BUG_ON(ret);
556
557         for (i = 0; i < nr_tasks; i++) {
558                 task = tasks[i];
559                 sem_init(&task->sleep_sem, 0, 0);
560                 task->curr_event = 0;
561         }
562 }
563
564 static void run_one_test(void)
565 {
566         u64 T0, T1, delta, avg_delta, fluct, std_dev;
567
568         T0 = get_nsecs();
569         wait_for_tasks();
570         T1 = get_nsecs();
571
572         delta = T1 - T0;
573         sum_runtime += delta;
574         nr_runs++;
575
576         avg_delta = sum_runtime / nr_runs;
577         if (delta < avg_delta)
578                 fluct = avg_delta - delta;
579         else
580                 fluct = delta - avg_delta;
581         sum_fluct += fluct;
582         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
583         if (!run_avg)
584                 run_avg = delta;
585         run_avg = (run_avg*9 + delta)/10;
586
587         printf("#%-3ld: %0.3f, ",
588                 nr_runs, (double)delta/1000000.0);
589
590         printf("ravg: %0.2f, ",
591                 (double)run_avg/1e6);
592
593         printf("cpu: %0.2f / %0.2f",
594                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
595
596 #if 0
597         /*
598          * rusage statistics done by the parent, these are less
599          * accurate than the sum_exec_runtime based statistics:
600          */
601         printf(" [%0.2f / %0.2f]",
602                 (double)parent_cpu_usage/1e6,
603                 (double)runavg_parent_cpu_usage/1e6);
604 #endif
605
606         printf("\n");
607
608         if (nr_sleep_corrections)
609                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
610         nr_sleep_corrections = 0;
611 }
612
613 static void test_calibrations(void)
614 {
615         u64 T0, T1;
616
617         T0 = get_nsecs();
618         burn_nsecs(1e6);
619         T1 = get_nsecs();
620
621         printf("the run test took %Ld nsecs\n", T1-T0);
622
623         T0 = get_nsecs();
624         sleep_nsecs(1e6);
625         T1 = get_nsecs();
626
627         printf("the sleep test took %Ld nsecs\n", T1-T0);
628 }
629
630 #define FILL_FIELD(ptr, field, event, data)     \
631         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
632
633 #define FILL_ARRAY(ptr, array, event, data)                     \
634 do {                                                            \
635         void *__array = raw_field_ptr(event, #array, data);     \
636         memcpy(ptr.array, __array, sizeof(ptr.array));  \
637 } while(0)
638
639 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
640 do {                                                            \
641         FILL_FIELD(ptr, common_type, event, data);              \
642         FILL_FIELD(ptr, common_flags, event, data);             \
643         FILL_FIELD(ptr, common_preempt_count, event, data);     \
644         FILL_FIELD(ptr, common_pid, event, data);               \
645         FILL_FIELD(ptr, common_tgid, event, data);              \
646 } while (0)
647
648
649
650 struct trace_switch_event {
651         u32 size;
652
653         u16 common_type;
654         u8 common_flags;
655         u8 common_preempt_count;
656         u32 common_pid;
657         u32 common_tgid;
658
659         char prev_comm[16];
660         u32 prev_pid;
661         u32 prev_prio;
662         u64 prev_state;
663         char next_comm[16];
664         u32 next_pid;
665         u32 next_prio;
666 };
667
668 struct trace_runtime_event {
669         u32 size;
670
671         u16 common_type;
672         u8 common_flags;
673         u8 common_preempt_count;
674         u32 common_pid;
675         u32 common_tgid;
676
677         char comm[16];
678         u32 pid;
679         u64 runtime;
680         u64 vruntime;
681 };
682
683 struct trace_wakeup_event {
684         u32 size;
685
686         u16 common_type;
687         u8 common_flags;
688         u8 common_preempt_count;
689         u32 common_pid;
690         u32 common_tgid;
691
692         char comm[16];
693         u32 pid;
694
695         u32 prio;
696         u32 success;
697         u32 cpu;
698 };
699
700 struct trace_fork_event {
701         u32 size;
702
703         u16 common_type;
704         u8 common_flags;
705         u8 common_preempt_count;
706         u32 common_pid;
707         u32 common_tgid;
708
709         char parent_comm[16];
710         u32 parent_pid;
711         char child_comm[16];
712         u32 child_pid;
713 };
714
715 struct trace_migrate_task_event {
716         u32 size;
717
718         u16 common_type;
719         u8 common_flags;
720         u8 common_preempt_count;
721         u32 common_pid;
722         u32 common_tgid;
723
724         char comm[16];
725         u32 pid;
726
727         u32 prio;
728         u32 cpu;
729 };
730
731 struct trace_sched_handler {
732         void (*switch_event)(struct trace_switch_event *,
733                              struct perf_session *,
734                              struct event *,
735                              int cpu,
736                              u64 timestamp,
737                              struct thread *thread);
738
739         void (*runtime_event)(struct trace_runtime_event *,
740                               struct perf_session *,
741                               struct event *,
742                               int cpu,
743                               u64 timestamp,
744                               struct thread *thread);
745
746         void (*wakeup_event)(struct trace_wakeup_event *,
747                              struct perf_session *,
748                              struct event *,
749                              int cpu,
750                              u64 timestamp,
751                              struct thread *thread);
752
753         void (*fork_event)(struct trace_fork_event *,
754                            struct event *,
755                            int cpu,
756                            u64 timestamp,
757                            struct thread *thread);
758
759         void (*migrate_task_event)(struct trace_migrate_task_event *,
760                            struct perf_session *session,
761                            struct event *,
762                            int cpu,
763                            u64 timestamp,
764                            struct thread *thread);
765 };
766
767
768 static void
769 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
770                     struct perf_session *session __used,
771                     struct event *event,
772                     int cpu __used,
773                     u64 timestamp __used,
774                     struct thread *thread __used)
775 {
776         struct task_desc *waker, *wakee;
777
778         if (verbose) {
779                 printf("sched_wakeup event %p\n", event);
780
781                 printf(" ... pid %d woke up %s/%d\n",
782                         wakeup_event->common_pid,
783                         wakeup_event->comm,
784                         wakeup_event->pid);
785         }
786
787         waker = register_pid(wakeup_event->common_pid, "<unknown>");
788         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
789
790         add_sched_event_wakeup(waker, timestamp, wakee);
791 }
792
793 static u64 cpu_last_switched[MAX_CPUS];
794
795 static void
796 replay_switch_event(struct trace_switch_event *switch_event,
797                     struct perf_session *session __used,
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->max_lat_at = timestamp;
1025         }
1026         atoms->nb_atoms++;
1027 }
1028
1029 static void
1030 latency_switch_event(struct trace_switch_event *switch_event,
1031                      struct perf_session *session,
1032                      struct event *event __used,
1033                      int cpu,
1034                      u64 timestamp,
1035                      struct thread *thread __used)
1036 {
1037         struct work_atoms *out_events, *in_events;
1038         struct thread *sched_out, *sched_in;
1039         u64 timestamp0;
1040         s64 delta;
1041
1042         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1043
1044         timestamp0 = cpu_last_switched[cpu];
1045         cpu_last_switched[cpu] = timestamp;
1046         if (timestamp0)
1047                 delta = timestamp - timestamp0;
1048         else
1049                 delta = 0;
1050
1051         if (delta < 0)
1052                 die("hm, delta: %Ld < 0 ?\n", delta);
1053
1054
1055         sched_out = perf_session__findnew(session, switch_event->prev_pid);
1056         sched_in = perf_session__findnew(session, switch_event->next_pid);
1057
1058         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1059         if (!out_events) {
1060                 thread_atoms_insert(sched_out);
1061                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1062                 if (!out_events)
1063                         die("out-event: Internal tree error");
1064         }
1065         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1066
1067         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1068         if (!in_events) {
1069                 thread_atoms_insert(sched_in);
1070                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1071                 if (!in_events)
1072                         die("in-event: Internal tree error");
1073                 /*
1074                  * Take came in we have not heard about yet,
1075                  * add in an initial atom in runnable state:
1076                  */
1077                 add_sched_out_event(in_events, 'R', timestamp);
1078         }
1079         add_sched_in_event(in_events, timestamp);
1080 }
1081
1082 static void
1083 latency_runtime_event(struct trace_runtime_event *runtime_event,
1084                      struct perf_session *session,
1085                      struct event *event __used,
1086                      int cpu,
1087                      u64 timestamp,
1088                      struct thread *this_thread __used)
1089 {
1090         struct thread *thread = perf_session__findnew(session, runtime_event->pid);
1091         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1092
1093         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1094         if (!atoms) {
1095                 thread_atoms_insert(thread);
1096                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1097                 if (!atoms)
1098                         die("in-event: Internal tree error");
1099                 add_sched_out_event(atoms, 'R', timestamp);
1100         }
1101
1102         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1103 }
1104
1105 static void
1106 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1107                      struct perf_session *session,
1108                      struct event *__event __used,
1109                      int cpu __used,
1110                      u64 timestamp,
1111                      struct thread *thread __used)
1112 {
1113         struct work_atoms *atoms;
1114         struct work_atom *atom;
1115         struct thread *wakee;
1116
1117         /* Note for later, it may be interesting to observe the failing cases */
1118         if (!wakeup_event->success)
1119                 return;
1120
1121         wakee = perf_session__findnew(session, wakeup_event->pid);
1122         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1123         if (!atoms) {
1124                 thread_atoms_insert(wakee);
1125                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1126                 if (!atoms)
1127                         die("wakeup-event: Internal tree error");
1128                 add_sched_out_event(atoms, 'S', timestamp);
1129         }
1130
1131         BUG_ON(list_empty(&atoms->work_list));
1132
1133         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1134
1135         /*
1136          * You WILL be missing events if you've recorded only
1137          * one CPU, or are only looking at only one, so don't
1138          * make useless noise.
1139          */
1140         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1141                 nr_state_machine_bugs++;
1142
1143         nr_timestamps++;
1144         if (atom->sched_out_time > timestamp) {
1145                 nr_unordered_timestamps++;
1146                 return;
1147         }
1148
1149         atom->state = THREAD_WAIT_CPU;
1150         atom->wake_up_time = timestamp;
1151 }
1152
1153 static void
1154 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1155                      struct perf_session *session,
1156                      struct event *__event __used,
1157                      int cpu __used,
1158                      u64 timestamp,
1159                      struct thread *thread __used)
1160 {
1161         struct work_atoms *atoms;
1162         struct work_atom *atom;
1163         struct thread *migrant;
1164
1165         /*
1166          * Only need to worry about migration when profiling one CPU.
1167          */
1168         if (profile_cpu == -1)
1169                 return;
1170
1171         migrant = perf_session__findnew(session, migrate_task_event->pid);
1172         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1173         if (!atoms) {
1174                 thread_atoms_insert(migrant);
1175                 register_pid(migrant->pid, migrant->comm);
1176                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1177                 if (!atoms)
1178                         die("migration-event: Internal tree error");
1179                 add_sched_out_event(atoms, 'R', timestamp);
1180         }
1181
1182         BUG_ON(list_empty(&atoms->work_list));
1183
1184         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1185         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1186
1187         nr_timestamps++;
1188
1189         if (atom->sched_out_time > timestamp)
1190                 nr_unordered_timestamps++;
1191 }
1192
1193 static struct trace_sched_handler lat_ops  = {
1194         .wakeup_event           = latency_wakeup_event,
1195         .switch_event           = latency_switch_event,
1196         .runtime_event          = latency_runtime_event,
1197         .fork_event             = latency_fork_event,
1198         .migrate_task_event     = latency_migrate_task_event,
1199 };
1200
1201 static void output_lat_thread(struct work_atoms *work_list)
1202 {
1203         int i;
1204         int ret;
1205         u64 avg;
1206
1207         if (!work_list->nb_atoms)
1208                 return;
1209         /*
1210          * Ignore idle threads:
1211          */
1212         if (!strcmp(work_list->thread->comm, "swapper"))
1213                 return;
1214
1215         all_runtime += work_list->total_runtime;
1216         all_count += work_list->nb_atoms;
1217
1218         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1219
1220         for (i = 0; i < 24 - ret; i++)
1221                 printf(" ");
1222
1223         avg = work_list->total_lat / work_list->nb_atoms;
1224
1225         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1226               (double)work_list->total_runtime / 1e6,
1227                  work_list->nb_atoms, (double)avg / 1e6,
1228                  (double)work_list->max_lat / 1e6,
1229                  (double)work_list->max_lat_at / 1e9);
1230 }
1231
1232 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1233 {
1234         if (l->thread->pid < r->thread->pid)
1235                 return -1;
1236         if (l->thread->pid > r->thread->pid)
1237                 return 1;
1238
1239         return 0;
1240 }
1241
1242 static struct sort_dimension pid_sort_dimension = {
1243         .name                   = "pid",
1244         .cmp                    = pid_cmp,
1245 };
1246
1247 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1248 {
1249         u64 avgl, avgr;
1250
1251         if (!l->nb_atoms)
1252                 return -1;
1253
1254         if (!r->nb_atoms)
1255                 return 1;
1256
1257         avgl = l->total_lat / l->nb_atoms;
1258         avgr = r->total_lat / r->nb_atoms;
1259
1260         if (avgl < avgr)
1261                 return -1;
1262         if (avgl > avgr)
1263                 return 1;
1264
1265         return 0;
1266 }
1267
1268 static struct sort_dimension avg_sort_dimension = {
1269         .name                   = "avg",
1270         .cmp                    = avg_cmp,
1271 };
1272
1273 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1274 {
1275         if (l->max_lat < r->max_lat)
1276                 return -1;
1277         if (l->max_lat > r->max_lat)
1278                 return 1;
1279
1280         return 0;
1281 }
1282
1283 static struct sort_dimension max_sort_dimension = {
1284         .name                   = "max",
1285         .cmp                    = max_cmp,
1286 };
1287
1288 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1289 {
1290         if (l->nb_atoms < r->nb_atoms)
1291                 return -1;
1292         if (l->nb_atoms > r->nb_atoms)
1293                 return 1;
1294
1295         return 0;
1296 }
1297
1298 static struct sort_dimension switch_sort_dimension = {
1299         .name                   = "switch",
1300         .cmp                    = switch_cmp,
1301 };
1302
1303 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1304 {
1305         if (l->total_runtime < r->total_runtime)
1306                 return -1;
1307         if (l->total_runtime > r->total_runtime)
1308                 return 1;
1309
1310         return 0;
1311 }
1312
1313 static struct sort_dimension runtime_sort_dimension = {
1314         .name                   = "runtime",
1315         .cmp                    = runtime_cmp,
1316 };
1317
1318 static struct sort_dimension *available_sorts[] = {
1319         &pid_sort_dimension,
1320         &avg_sort_dimension,
1321         &max_sort_dimension,
1322         &switch_sort_dimension,
1323         &runtime_sort_dimension,
1324 };
1325
1326 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1327
1328 static LIST_HEAD(sort_list);
1329
1330 static int sort_dimension__add(const char *tok, struct list_head *list)
1331 {
1332         int i;
1333
1334         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1335                 if (!strcmp(available_sorts[i]->name, tok)) {
1336                         list_add_tail(&available_sorts[i]->list, list);
1337
1338                         return 0;
1339                 }
1340         }
1341
1342         return -1;
1343 }
1344
1345 static void setup_sorting(void);
1346
1347 static void sort_lat(void)
1348 {
1349         struct rb_node *node;
1350
1351         for (;;) {
1352                 struct work_atoms *data;
1353                 node = rb_first(&atom_root);
1354                 if (!node)
1355                         break;
1356
1357                 rb_erase(node, &atom_root);
1358                 data = rb_entry(node, struct work_atoms, node);
1359                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1360         }
1361 }
1362
1363 static struct trace_sched_handler *trace_handler;
1364
1365 static void
1366 process_sched_wakeup_event(void *data, struct perf_session *session,
1367                            struct event *event,
1368                            int cpu __used,
1369                            u64 timestamp __used,
1370                            struct thread *thread __used)
1371 {
1372         struct trace_wakeup_event wakeup_event;
1373
1374         FILL_COMMON_FIELDS(wakeup_event, event, data);
1375
1376         FILL_ARRAY(wakeup_event, comm, event, data);
1377         FILL_FIELD(wakeup_event, pid, event, data);
1378         FILL_FIELD(wakeup_event, prio, event, data);
1379         FILL_FIELD(wakeup_event, success, event, data);
1380         FILL_FIELD(wakeup_event, cpu, event, data);
1381
1382         if (trace_handler->wakeup_event)
1383                 trace_handler->wakeup_event(&wakeup_event, session, event,
1384                                             cpu, timestamp, thread);
1385 }
1386
1387 /*
1388  * Track the current task - that way we can know whether there's any
1389  * weird events, such as a task being switched away that is not current.
1390  */
1391 static int max_cpu;
1392
1393 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1394
1395 static struct thread *curr_thread[MAX_CPUS];
1396
1397 static char next_shortname1 = 'A';
1398 static char next_shortname2 = '0';
1399
1400 static void
1401 map_switch_event(struct trace_switch_event *switch_event,
1402                  struct perf_session *session,
1403                  struct event *event __used,
1404                  int this_cpu,
1405                  u64 timestamp,
1406                  struct thread *thread __used)
1407 {
1408         struct thread *sched_out, *sched_in;
1409         int new_shortname;
1410         u64 timestamp0;
1411         s64 delta;
1412         int cpu;
1413
1414         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1415
1416         if (this_cpu > max_cpu)
1417                 max_cpu = this_cpu;
1418
1419         timestamp0 = cpu_last_switched[this_cpu];
1420         cpu_last_switched[this_cpu] = timestamp;
1421         if (timestamp0)
1422                 delta = timestamp - timestamp0;
1423         else
1424                 delta = 0;
1425
1426         if (delta < 0)
1427                 die("hm, delta: %Ld < 0 ?\n", delta);
1428
1429
1430         sched_out = perf_session__findnew(session, switch_event->prev_pid);
1431         sched_in = perf_session__findnew(session, switch_event->next_pid);
1432
1433         curr_thread[this_cpu] = sched_in;
1434
1435         printf("  ");
1436
1437         new_shortname = 0;
1438         if (!sched_in->shortname[0]) {
1439                 sched_in->shortname[0] = next_shortname1;
1440                 sched_in->shortname[1] = next_shortname2;
1441
1442                 if (next_shortname1 < 'Z') {
1443                         next_shortname1++;
1444                 } else {
1445                         next_shortname1='A';
1446                         if (next_shortname2 < '9') {
1447                                 next_shortname2++;
1448                         } else {
1449                                 next_shortname2='0';
1450                         }
1451                 }
1452                 new_shortname = 1;
1453         }
1454
1455         for (cpu = 0; cpu <= max_cpu; cpu++) {
1456                 if (cpu != this_cpu)
1457                         printf(" ");
1458                 else
1459                         printf("*");
1460
1461                 if (curr_thread[cpu]) {
1462                         if (curr_thread[cpu]->pid)
1463                                 printf("%2s ", curr_thread[cpu]->shortname);
1464                         else
1465                                 printf(".  ");
1466                 } else
1467                         printf("   ");
1468         }
1469
1470         printf("  %12.6f secs ", (double)timestamp/1e9);
1471         if (new_shortname) {
1472                 printf("%s => %s:%d\n",
1473                         sched_in->shortname, sched_in->comm, sched_in->pid);
1474         } else {
1475                 printf("\n");
1476         }
1477 }
1478
1479
1480 static void
1481 process_sched_switch_event(void *data, struct perf_session *session,
1482                            struct event *event,
1483                            int this_cpu,
1484                            u64 timestamp __used,
1485                            struct thread *thread __used)
1486 {
1487         struct trace_switch_event switch_event;
1488
1489         FILL_COMMON_FIELDS(switch_event, event, data);
1490
1491         FILL_ARRAY(switch_event, prev_comm, event, data);
1492         FILL_FIELD(switch_event, prev_pid, event, data);
1493         FILL_FIELD(switch_event, prev_prio, event, data);
1494         FILL_FIELD(switch_event, prev_state, event, data);
1495         FILL_ARRAY(switch_event, next_comm, event, data);
1496         FILL_FIELD(switch_event, next_pid, event, data);
1497         FILL_FIELD(switch_event, next_prio, event, data);
1498
1499         if (curr_pid[this_cpu] != (u32)-1) {
1500                 /*
1501                  * Are we trying to switch away a PID that is
1502                  * not current?
1503                  */
1504                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1505                         nr_context_switch_bugs++;
1506         }
1507         if (trace_handler->switch_event)
1508                 trace_handler->switch_event(&switch_event, session, event,
1509                                             this_cpu, timestamp, thread);
1510
1511         curr_pid[this_cpu] = switch_event.next_pid;
1512 }
1513
1514 static void
1515 process_sched_runtime_event(void *data, struct perf_session *session,
1516                            struct event *event,
1517                            int cpu __used,
1518                            u64 timestamp __used,
1519                            struct thread *thread __used)
1520 {
1521         struct trace_runtime_event runtime_event;
1522
1523         FILL_ARRAY(runtime_event, comm, event, data);
1524         FILL_FIELD(runtime_event, pid, event, data);
1525         FILL_FIELD(runtime_event, runtime, event, data);
1526         FILL_FIELD(runtime_event, vruntime, event, data);
1527
1528         if (trace_handler->runtime_event)
1529                 trace_handler->runtime_event(&runtime_event, session, event, cpu, timestamp, thread);
1530 }
1531
1532 static void
1533 process_sched_fork_event(void *data,
1534                          struct event *event,
1535                          int cpu __used,
1536                          u64 timestamp __used,
1537                          struct thread *thread __used)
1538 {
1539         struct trace_fork_event fork_event;
1540
1541         FILL_COMMON_FIELDS(fork_event, event, data);
1542
1543         FILL_ARRAY(fork_event, parent_comm, event, data);
1544         FILL_FIELD(fork_event, parent_pid, event, data);
1545         FILL_ARRAY(fork_event, child_comm, event, data);
1546         FILL_FIELD(fork_event, child_pid, event, data);
1547
1548         if (trace_handler->fork_event)
1549                 trace_handler->fork_event(&fork_event, event,
1550                                           cpu, timestamp, thread);
1551 }
1552
1553 static void
1554 process_sched_exit_event(struct event *event,
1555                          int cpu __used,
1556                          u64 timestamp __used,
1557                          struct thread *thread __used)
1558 {
1559         if (verbose)
1560                 printf("sched_exit event %p\n", event);
1561 }
1562
1563 static void
1564 process_sched_migrate_task_event(void *data, struct perf_session *session,
1565                            struct event *event,
1566                            int cpu __used,
1567                            u64 timestamp __used,
1568                            struct thread *thread __used)
1569 {
1570         struct trace_migrate_task_event migrate_task_event;
1571
1572         FILL_COMMON_FIELDS(migrate_task_event, event, data);
1573
1574         FILL_ARRAY(migrate_task_event, comm, event, data);
1575         FILL_FIELD(migrate_task_event, pid, event, data);
1576         FILL_FIELD(migrate_task_event, prio, event, data);
1577         FILL_FIELD(migrate_task_event, cpu, event, data);
1578
1579         if (trace_handler->migrate_task_event)
1580                 trace_handler->migrate_task_event(&migrate_task_event, session,
1581                                                  event, cpu, timestamp, thread);
1582 }
1583
1584 static void
1585 process_raw_event(event_t *raw_event __used, struct perf_session *session,
1586                   void *data, int cpu, u64 timestamp, struct thread *thread)
1587 {
1588         struct event *event;
1589         int type;
1590
1591
1592         type = trace_parse_common_type(data);
1593         event = trace_find_event(type);
1594
1595         if (!strcmp(event->name, "sched_switch"))
1596                 process_sched_switch_event(data, session, event, cpu, timestamp, thread);
1597         if (!strcmp(event->name, "sched_stat_runtime"))
1598                 process_sched_runtime_event(data, session, event, cpu, timestamp, thread);
1599         if (!strcmp(event->name, "sched_wakeup"))
1600                 process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1601         if (!strcmp(event->name, "sched_wakeup_new"))
1602                 process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1603         if (!strcmp(event->name, "sched_process_fork"))
1604                 process_sched_fork_event(data, event, cpu, timestamp, thread);
1605         if (!strcmp(event->name, "sched_process_exit"))
1606                 process_sched_exit_event(event, cpu, timestamp, thread);
1607         if (!strcmp(event->name, "sched_migrate_task"))
1608                 process_sched_migrate_task_event(data, session, event, cpu, timestamp, thread);
1609 }
1610
1611 static int process_sample_event(event_t *event, struct perf_session *session)
1612 {
1613         struct sample_data data;
1614         struct thread *thread;
1615
1616         if (!(sample_type & PERF_SAMPLE_RAW))
1617                 return 0;
1618
1619         memset(&data, 0, sizeof(data));
1620         data.time = -1;
1621         data.cpu = -1;
1622         data.period = -1;
1623
1624         event__parse_sample(event, sample_type, &data);
1625
1626         dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
1627                 event->header.misc,
1628                 data.pid, data.tid,
1629                 (void *)(long)data.ip,
1630                 (long long)data.period);
1631
1632         thread = perf_session__findnew(session, data.pid);
1633         if (thread == NULL) {
1634                 pr_debug("problem processing %d event, skipping it.\n",
1635                          event->header.type);
1636                 return -1;
1637         }
1638
1639         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1640
1641         if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
1642                 return 0;
1643
1644         process_raw_event(event, session, data.raw_data, data.cpu, data.time, thread);
1645
1646         return 0;
1647 }
1648
1649 static int process_lost_event(event_t *event __used,
1650                               struct perf_session *session __used)
1651 {
1652         nr_lost_chunks++;
1653         nr_lost_events += event->lost.lost;
1654
1655         return 0;
1656 }
1657
1658 static int sample_type_check(u64 type)
1659 {
1660         sample_type = type;
1661
1662         if (!(sample_type & PERF_SAMPLE_RAW)) {
1663                 fprintf(stderr,
1664                         "No trace sample to read. Did you call perf record "
1665                         "without -R?");
1666                 return -1;
1667         }
1668
1669         return 0;
1670 }
1671
1672 static struct perf_event_ops event_ops = {
1673         .process_sample_event   = process_sample_event,
1674         .process_comm_event     = event__process_comm,
1675         .process_lost_event     = process_lost_event,
1676         .sample_type_check      = sample_type_check,
1677 };
1678
1679 static int read_events(void)
1680 {
1681         int err;
1682         struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
1683
1684         if (session == NULL)
1685                 return -ENOMEM;
1686
1687         err = perf_session__process_events(session, &event_ops);
1688         perf_session__delete(session);
1689         return err;
1690 }
1691
1692 static void print_bad_events(void)
1693 {
1694         if (nr_unordered_timestamps && nr_timestamps) {
1695                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1696                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1697                         nr_unordered_timestamps, nr_timestamps);
1698         }
1699         if (nr_lost_events && nr_events) {
1700                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1701                         (double)nr_lost_events/(double)nr_events*100.0,
1702                         nr_lost_events, nr_events, nr_lost_chunks);
1703         }
1704         if (nr_state_machine_bugs && nr_timestamps) {
1705                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1706                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1707                         nr_state_machine_bugs, nr_timestamps);
1708                 if (nr_lost_events)
1709                         printf(" (due to lost events?)");
1710                 printf("\n");
1711         }
1712         if (nr_context_switch_bugs && nr_timestamps) {
1713                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1714                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1715                         nr_context_switch_bugs, nr_timestamps);
1716                 if (nr_lost_events)
1717                         printf(" (due to lost events?)");
1718                 printf("\n");
1719         }
1720 }
1721
1722 static void __cmd_lat(void)
1723 {
1724         struct rb_node *next;
1725
1726         setup_pager();
1727         read_events();
1728         sort_lat();
1729
1730         printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1731         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
1732         printf(" ---------------------------------------------------------------------------------------------------------------\n");
1733
1734         next = rb_first(&sorted_atom_root);
1735
1736         while (next) {
1737                 struct work_atoms *work_list;
1738
1739                 work_list = rb_entry(next, struct work_atoms, node);
1740                 output_lat_thread(work_list);
1741                 next = rb_next(next);
1742         }
1743
1744         printf(" -----------------------------------------------------------------------------------------\n");
1745         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1746                 (double)all_runtime/1e6, all_count);
1747
1748         printf(" ---------------------------------------------------\n");
1749
1750         print_bad_events();
1751         printf("\n");
1752
1753 }
1754
1755 static struct trace_sched_handler map_ops  = {
1756         .wakeup_event           = NULL,
1757         .switch_event           = map_switch_event,
1758         .runtime_event          = NULL,
1759         .fork_event             = NULL,
1760 };
1761
1762 static void __cmd_map(void)
1763 {
1764         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1765
1766         setup_pager();
1767         read_events();
1768         print_bad_events();
1769 }
1770
1771 static void __cmd_replay(void)
1772 {
1773         unsigned long i;
1774
1775         calibrate_run_measurement_overhead();
1776         calibrate_sleep_measurement_overhead();
1777
1778         test_calibrations();
1779
1780         read_events();
1781
1782         printf("nr_run_events:        %ld\n", nr_run_events);
1783         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1784         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1785
1786         if (targetless_wakeups)
1787                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1788         if (multitarget_wakeups)
1789                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1790         if (nr_run_events_optimized)
1791                 printf("run atoms optimized: %ld\n",
1792                         nr_run_events_optimized);
1793
1794         print_task_traces();
1795         add_cross_task_wakeups();
1796
1797         create_tasks();
1798         printf("------------------------------------------------------------\n");
1799         for (i = 0; i < replay_repeat; i++)
1800                 run_one_test();
1801 }
1802
1803
1804 static const char * const sched_usage[] = {
1805         "perf sched [<options>] {record|latency|map|replay|trace}",
1806         NULL
1807 };
1808
1809 static const struct option sched_options[] = {
1810         OPT_STRING('i', "input", &input_name, "file",
1811                     "input file name"),
1812         OPT_BOOLEAN('v', "verbose", &verbose,
1813                     "be more verbose (show symbol address, etc)"),
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 latency_usage[] = {
1820         "perf sched latency [<options>]",
1821         NULL
1822 };
1823
1824 static const struct option latency_options[] = {
1825         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1826                    "sort by key(s): runtime, switch, avg, max"),
1827         OPT_BOOLEAN('v', "verbose", &verbose,
1828                     "be more verbose (show symbol address, etc)"),
1829         OPT_INTEGER('C', "CPU", &profile_cpu,
1830                     "CPU to profile on"),
1831         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1832                     "dump raw trace in ASCII"),
1833         OPT_END()
1834 };
1835
1836 static const char * const replay_usage[] = {
1837         "perf sched replay [<options>]",
1838         NULL
1839 };
1840
1841 static const struct option replay_options[] = {
1842         OPT_INTEGER('r', "repeat", &replay_repeat,
1843                     "repeat the workload replay N times (-1: infinite)"),
1844         OPT_BOOLEAN('v', "verbose", &verbose,
1845                     "be more verbose (show symbol address, etc)"),
1846         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1847                     "dump raw trace in ASCII"),
1848         OPT_END()
1849 };
1850
1851 static void setup_sorting(void)
1852 {
1853         char *tmp, *tok, *str = strdup(sort_order);
1854
1855         for (tok = strtok_r(str, ", ", &tmp);
1856                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1857                 if (sort_dimension__add(tok, &sort_list) < 0) {
1858                         error("Unknown --sort key: `%s'", tok);
1859                         usage_with_options(latency_usage, latency_options);
1860                 }
1861         }
1862
1863         free(str);
1864
1865         sort_dimension__add("pid", &cmp_pid);
1866 }
1867
1868 static const char *record_args[] = {
1869         "record",
1870         "-a",
1871         "-R",
1872         "-M",
1873         "-f",
1874         "-m", "1024",
1875         "-c", "1",
1876         "-e", "sched:sched_switch:r",
1877         "-e", "sched:sched_stat_wait:r",
1878         "-e", "sched:sched_stat_sleep:r",
1879         "-e", "sched:sched_stat_iowait:r",
1880         "-e", "sched:sched_stat_runtime:r",
1881         "-e", "sched:sched_process_exit:r",
1882         "-e", "sched:sched_process_fork:r",
1883         "-e", "sched:sched_wakeup:r",
1884         "-e", "sched:sched_migrate_task:r",
1885 };
1886
1887 static int __cmd_record(int argc, const char **argv)
1888 {
1889         unsigned int rec_argc, i, j;
1890         const char **rec_argv;
1891
1892         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1893         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1894
1895         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1896                 rec_argv[i] = strdup(record_args[i]);
1897
1898         for (j = 1; j < (unsigned int)argc; j++, i++)
1899                 rec_argv[i] = argv[j];
1900
1901         BUG_ON(i != rec_argc);
1902
1903         return cmd_record(i, rec_argv, NULL);
1904 }
1905
1906 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1907 {
1908         argc = parse_options(argc, argv, sched_options, sched_usage,
1909                              PARSE_OPT_STOP_AT_NON_OPTION);
1910         if (!argc)
1911                 usage_with_options(sched_usage, sched_options);
1912
1913         /*
1914          * Aliased to 'perf trace' for now:
1915          */
1916         if (!strcmp(argv[0], "trace"))
1917                 return cmd_trace(argc, argv, prefix);
1918
1919         symbol__init(0);
1920         if (!strncmp(argv[0], "rec", 3)) {
1921                 return __cmd_record(argc, argv);
1922         } else if (!strncmp(argv[0], "lat", 3)) {
1923                 trace_handler = &lat_ops;
1924                 if (argc > 1) {
1925                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1926                         if (argc)
1927                                 usage_with_options(latency_usage, latency_options);
1928                 }
1929                 setup_sorting();
1930                 __cmd_lat();
1931         } else if (!strcmp(argv[0], "map")) {
1932                 trace_handler = &map_ops;
1933                 setup_sorting();
1934                 __cmd_map();
1935         } else if (!strncmp(argv[0], "rep", 3)) {
1936                 trace_handler = &replay_ops;
1937                 if (argc) {
1938                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1939                         if (argc)
1940                                 usage_with_options(replay_usage, replay_options);
1941                 }
1942                 __cmd_replay();
1943         } else {
1944                 usage_with_options(sched_usage, sched_options);
1945         }
1946
1947         return 0;
1948 }