e1df7055ab823f718b02669037ff0b08251e244b
[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 unsigned long            total_comm = 0;
26
27 static struct rb_root           threads;
28 static struct thread            *last_match;
29
30 static struct perf_header       *header;
31 static u64                      sample_type;
32
33 static char                     default_sort_order[] = "avg, max, switch, runtime";
34 static char                     *sort_order = default_sort_order;
35
36 static char                     *cwd;
37 static int                      cwdlen;
38
39 #define PR_SET_NAME             15               /* Set process name */
40 #define MAX_CPUS                4096
41
42 #define BUG_ON(x)               assert(!(x))
43
44 static u64                      run_measurement_overhead;
45 static u64                      sleep_measurement_overhead;
46
47 #define COMM_LEN                20
48 #define SYM_LEN                 129
49
50 #define MAX_PID                 65536
51
52 static unsigned long            nr_tasks;
53
54 struct sched_atom;
55
56 struct task_desc {
57         unsigned long           nr;
58         unsigned long           pid;
59         char                    comm[COMM_LEN];
60
61         unsigned long           nr_events;
62         unsigned long           curr_event;
63         struct sched_atom       **atoms;
64
65         pthread_t               thread;
66         sem_t                   sleep_sem;
67
68         sem_t                   ready_for_work;
69         sem_t                   work_done_sem;
70
71         u64                     cpu_usage;
72 };
73
74 enum sched_event_type {
75         SCHED_EVENT_RUN,
76         SCHED_EVENT_SLEEP,
77         SCHED_EVENT_WAKEUP,
78 };
79
80 struct sched_atom {
81         enum sched_event_type   type;
82         u64                     timestamp;
83         u64                     duration;
84         unsigned long           nr;
85         int                     specific_wait;
86         sem_t                   *wait_sem;
87         struct task_desc        *wakee;
88 };
89
90 static struct task_desc         *pid_to_task[MAX_PID];
91
92 static struct task_desc         **tasks;
93
94 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
95 static u64                      start_time;
96
97 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
98
99 static unsigned long            nr_run_events;
100 static unsigned long            nr_sleep_events;
101 static unsigned long            nr_wakeup_events;
102
103 static unsigned long            nr_sleep_corrections;
104 static unsigned long            nr_run_events_optimized;
105
106 static unsigned long            targetless_wakeups;
107 static unsigned long            multitarget_wakeups;
108
109 static u64                      cpu_usage;
110 static u64                      runavg_cpu_usage;
111 static u64                      parent_cpu_usage;
112 static u64                      runavg_parent_cpu_usage;
113
114 static unsigned long            nr_runs;
115 static u64                      sum_runtime;
116 static u64                      sum_fluct;
117 static u64                      run_avg;
118
119 static unsigned long            replay_repeat = 10;
120 static unsigned long            nr_timestamps;
121 static unsigned long            nr_unordered_timestamps;
122 static unsigned long            nr_state_machine_bugs;
123 static unsigned long            nr_context_switch_bugs;
124 static unsigned long            nr_events;
125 static unsigned long            nr_lost_chunks;
126 static unsigned long            nr_lost_events;
127
128 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
129
130 enum thread_state {
131         THREAD_SLEEPING = 0,
132         THREAD_WAIT_CPU,
133         THREAD_SCHED_IN,
134         THREAD_IGNORE
135 };
136
137 struct work_atom {
138         struct list_head        list;
139         enum thread_state       state;
140         u64                     sched_out_time;
141         u64                     wake_up_time;
142         u64                     sched_in_time;
143         u64                     runtime;
144 };
145
146 struct work_atoms {
147         struct list_head        work_list;
148         struct thread           *thread;
149         struct rb_node          node;
150         u64                     max_lat;
151         u64                     total_lat;
152         u64                     nb_atoms;
153         u64                     total_runtime;
154 };
155
156 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
157
158 static struct rb_root           atom_root, sorted_atom_root;
159
160 static u64                      all_runtime;
161 static u64                      all_count;
162
163
164 static u64 get_nsecs(void)
165 {
166         struct timespec ts;
167
168         clock_gettime(CLOCK_MONOTONIC, &ts);
169
170         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
171 }
172
173 static void burn_nsecs(u64 nsecs)
174 {
175         u64 T0 = get_nsecs(), T1;
176
177         do {
178                 T1 = get_nsecs();
179         } while (T1 + run_measurement_overhead < T0 + nsecs);
180 }
181
182 static void sleep_nsecs(u64 nsecs)
183 {
184         struct timespec ts;
185
186         ts.tv_nsec = nsecs % 999999999;
187         ts.tv_sec = nsecs / 999999999;
188
189         nanosleep(&ts, NULL);
190 }
191
192 static void calibrate_run_measurement_overhead(void)
193 {
194         u64 T0, T1, delta, min_delta = 1000000000ULL;
195         int i;
196
197         for (i = 0; i < 10; i++) {
198                 T0 = get_nsecs();
199                 burn_nsecs(0);
200                 T1 = get_nsecs();
201                 delta = T1-T0;
202                 min_delta = min(min_delta, delta);
203         }
204         run_measurement_overhead = min_delta;
205
206         printf("run measurement overhead: %Ld nsecs\n", min_delta);
207 }
208
209 static void calibrate_sleep_measurement_overhead(void)
210 {
211         u64 T0, T1, delta, min_delta = 1000000000ULL;
212         int i;
213
214         for (i = 0; i < 10; i++) {
215                 T0 = get_nsecs();
216                 sleep_nsecs(10000);
217                 T1 = get_nsecs();
218                 delta = T1-T0;
219                 min_delta = min(min_delta, delta);
220         }
221         min_delta -= 10000;
222         sleep_measurement_overhead = min_delta;
223
224         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
225 }
226
227 static struct sched_atom *
228 get_new_event(struct task_desc *task, u64 timestamp)
229 {
230         struct sched_atom *event = calloc(1, sizeof(*event));
231         unsigned long idx = task->nr_events;
232         size_t size;
233
234         event->timestamp = timestamp;
235         event->nr = idx;
236
237         task->nr_events++;
238         size = sizeof(struct sched_atom *) * task->nr_events;
239         task->atoms = realloc(task->atoms, size);
240         BUG_ON(!task->atoms);
241
242         task->atoms[idx] = event;
243
244         return event;
245 }
246
247 static struct sched_atom *last_event(struct task_desc *task)
248 {
249         if (!task->nr_events)
250                 return NULL;
251
252         return task->atoms[task->nr_events - 1];
253 }
254
255 static void
256 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
257 {
258         struct sched_atom *event, *curr_event = last_event(task);
259
260         /*
261          * optimize an existing RUN event by merging this one
262          * to it:
263          */
264         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
265                 nr_run_events_optimized++;
266                 curr_event->duration += duration;
267                 return;
268         }
269
270         event = get_new_event(task, timestamp);
271
272         event->type = SCHED_EVENT_RUN;
273         event->duration = duration;
274
275         nr_run_events++;
276 }
277
278 static void
279 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
280                        struct task_desc *wakee)
281 {
282         struct sched_atom *event, *wakee_event;
283
284         event = get_new_event(task, timestamp);
285         event->type = SCHED_EVENT_WAKEUP;
286         event->wakee = wakee;
287
288         wakee_event = last_event(wakee);
289         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
290                 targetless_wakeups++;
291                 return;
292         }
293         if (wakee_event->wait_sem) {
294                 multitarget_wakeups++;
295                 return;
296         }
297
298         wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
299         sem_init(wakee_event->wait_sem, 0, 0);
300         wakee_event->specific_wait = 1;
301         event->wait_sem = wakee_event->wait_sem;
302
303         nr_wakeup_events++;
304 }
305
306 static void
307 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
308                       u64 task_state __used)
309 {
310         struct sched_atom *event = get_new_event(task, timestamp);
311
312         event->type = SCHED_EVENT_SLEEP;
313
314         nr_sleep_events++;
315 }
316
317 static struct task_desc *register_pid(unsigned long pid, const char *comm)
318 {
319         struct task_desc *task;
320
321         BUG_ON(pid >= MAX_PID);
322
323         task = pid_to_task[pid];
324
325         if (task)
326                 return task;
327
328         task = calloc(1, sizeof(*task));
329         task->pid = pid;
330         task->nr = nr_tasks;
331         strcpy(task->comm, comm);
332         /*
333          * every task starts in sleeping state - this gets ignored
334          * if there's no wakeup pointing to this sleep state:
335          */
336         add_sched_event_sleep(task, 0, 0);
337
338         pid_to_task[pid] = task;
339         nr_tasks++;
340         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
341         BUG_ON(!tasks);
342         tasks[task->nr] = task;
343
344         if (verbose)
345                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
346
347         return task;
348 }
349
350
351 static void print_task_traces(void)
352 {
353         struct task_desc *task;
354         unsigned long i;
355
356         for (i = 0; i < nr_tasks; i++) {
357                 task = tasks[i];
358                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
359                         task->nr, task->comm, task->pid, task->nr_events);
360         }
361 }
362
363 static void add_cross_task_wakeups(void)
364 {
365         struct task_desc *task1, *task2;
366         unsigned long i, j;
367
368         for (i = 0; i < nr_tasks; i++) {
369                 task1 = tasks[i];
370                 j = i + 1;
371                 if (j == nr_tasks)
372                         j = 0;
373                 task2 = tasks[j];
374                 add_sched_event_wakeup(task1, 0, task2);
375         }
376 }
377
378 static void
379 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
380 {
381         int ret = 0;
382         u64 now;
383         long long delta;
384
385         now = get_nsecs();
386         delta = start_time + atom->timestamp - now;
387
388         switch (atom->type) {
389                 case SCHED_EVENT_RUN:
390                         burn_nsecs(atom->duration);
391                         break;
392                 case SCHED_EVENT_SLEEP:
393                         if (atom->wait_sem)
394                                 ret = sem_wait(atom->wait_sem);
395                         BUG_ON(ret);
396                         break;
397                 case SCHED_EVENT_WAKEUP:
398                         if (atom->wait_sem)
399                                 ret = sem_post(atom->wait_sem);
400                         BUG_ON(ret);
401                         break;
402                 default:
403                         BUG_ON(1);
404         }
405 }
406
407 static u64 get_cpu_usage_nsec_parent(void)
408 {
409         struct rusage ru;
410         u64 sum;
411         int err;
412
413         err = getrusage(RUSAGE_SELF, &ru);
414         BUG_ON(err);
415
416         sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
417         sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
418
419         return sum;
420 }
421
422 static u64 get_cpu_usage_nsec_self(void)
423 {
424         char filename [] = "/proc/1234567890/sched";
425         unsigned long msecs, nsecs;
426         char *line = NULL;
427         u64 total = 0;
428         size_t len = 0;
429         ssize_t chars;
430         FILE *file;
431         int ret;
432
433         sprintf(filename, "/proc/%d/sched", getpid());
434         file = fopen(filename, "r");
435         BUG_ON(!file);
436
437         while ((chars = getline(&line, &len, file)) != -1) {
438                 ret = sscanf(line, "se.sum_exec_runtime : %ld.%06ld\n",
439                         &msecs, &nsecs);
440                 if (ret == 2) {
441                         total = msecs*1e6 + nsecs;
442                         break;
443                 }
444         }
445         if (line)
446                 free(line);
447         fclose(file);
448
449         return total;
450 }
451
452 static void *thread_func(void *ctx)
453 {
454         struct task_desc *this_task = ctx;
455         u64 cpu_usage_0, cpu_usage_1;
456         unsigned long i, ret;
457         char comm2[22];
458
459         sprintf(comm2, ":%s", this_task->comm);
460         prctl(PR_SET_NAME, comm2);
461
462 again:
463         ret = sem_post(&this_task->ready_for_work);
464         BUG_ON(ret);
465         ret = pthread_mutex_lock(&start_work_mutex);
466         BUG_ON(ret);
467         ret = pthread_mutex_unlock(&start_work_mutex);
468         BUG_ON(ret);
469
470         cpu_usage_0 = get_cpu_usage_nsec_self();
471
472         for (i = 0; i < this_task->nr_events; i++) {
473                 this_task->curr_event = i;
474                 process_sched_event(this_task, this_task->atoms[i]);
475         }
476
477         cpu_usage_1 = get_cpu_usage_nsec_self();
478         this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
479
480         ret = sem_post(&this_task->work_done_sem);
481         BUG_ON(ret);
482
483         ret = pthread_mutex_lock(&work_done_wait_mutex);
484         BUG_ON(ret);
485         ret = pthread_mutex_unlock(&work_done_wait_mutex);
486         BUG_ON(ret);
487
488         goto again;
489 }
490
491 static void create_tasks(void)
492 {
493         struct task_desc *task;
494         pthread_attr_t attr;
495         unsigned long i;
496         int err;
497
498         err = pthread_attr_init(&attr);
499         BUG_ON(err);
500         err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
501         BUG_ON(err);
502         err = pthread_mutex_lock(&start_work_mutex);
503         BUG_ON(err);
504         err = pthread_mutex_lock(&work_done_wait_mutex);
505         BUG_ON(err);
506         for (i = 0; i < nr_tasks; i++) {
507                 task = tasks[i];
508                 sem_init(&task->sleep_sem, 0, 0);
509                 sem_init(&task->ready_for_work, 0, 0);
510                 sem_init(&task->work_done_sem, 0, 0);
511                 task->curr_event = 0;
512                 err = pthread_create(&task->thread, &attr, thread_func, task);
513                 BUG_ON(err);
514         }
515 }
516
517 static void wait_for_tasks(void)
518 {
519         u64 cpu_usage_0, cpu_usage_1;
520         struct task_desc *task;
521         unsigned long i, ret;
522
523         start_time = get_nsecs();
524         cpu_usage = 0;
525         pthread_mutex_unlock(&work_done_wait_mutex);
526
527         for (i = 0; i < nr_tasks; i++) {
528                 task = tasks[i];
529                 ret = sem_wait(&task->ready_for_work);
530                 BUG_ON(ret);
531                 sem_init(&task->ready_for_work, 0, 0);
532         }
533         ret = pthread_mutex_lock(&work_done_wait_mutex);
534         BUG_ON(ret);
535
536         cpu_usage_0 = get_cpu_usage_nsec_parent();
537
538         pthread_mutex_unlock(&start_work_mutex);
539
540         for (i = 0; i < nr_tasks; i++) {
541                 task = tasks[i];
542                 ret = sem_wait(&task->work_done_sem);
543                 BUG_ON(ret);
544                 sem_init(&task->work_done_sem, 0, 0);
545                 cpu_usage += task->cpu_usage;
546                 task->cpu_usage = 0;
547         }
548
549         cpu_usage_1 = get_cpu_usage_nsec_parent();
550         if (!runavg_cpu_usage)
551                 runavg_cpu_usage = cpu_usage;
552         runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
553
554         parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
555         if (!runavg_parent_cpu_usage)
556                 runavg_parent_cpu_usage = parent_cpu_usage;
557         runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
558                                    parent_cpu_usage)/10;
559
560         ret = pthread_mutex_lock(&start_work_mutex);
561         BUG_ON(ret);
562
563         for (i = 0; i < nr_tasks; i++) {
564                 task = tasks[i];
565                 sem_init(&task->sleep_sem, 0, 0);
566                 task->curr_event = 0;
567         }
568 }
569
570 static void run_one_test(void)
571 {
572         u64 T0, T1, delta, avg_delta, fluct, std_dev;
573
574         T0 = get_nsecs();
575         wait_for_tasks();
576         T1 = get_nsecs();
577
578         delta = T1 - T0;
579         sum_runtime += delta;
580         nr_runs++;
581
582         avg_delta = sum_runtime / nr_runs;
583         if (delta < avg_delta)
584                 fluct = avg_delta - delta;
585         else
586                 fluct = delta - avg_delta;
587         sum_fluct += fluct;
588         std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
589         if (!run_avg)
590                 run_avg = delta;
591         run_avg = (run_avg*9 + delta)/10;
592
593         printf("#%-3ld: %0.3f, ",
594                 nr_runs, (double)delta/1000000.0);
595
596         printf("ravg: %0.2f, ",
597                 (double)run_avg/1e6);
598
599         printf("cpu: %0.2f / %0.2f",
600                 (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
601
602 #if 0
603         /*
604          * rusage statistics done by the parent, these are less
605          * accurate than the sum_exec_runtime based statistics:
606          */
607         printf(" [%0.2f / %0.2f]",
608                 (double)parent_cpu_usage/1e6,
609                 (double)runavg_parent_cpu_usage/1e6);
610 #endif
611
612         printf("\n");
613
614         if (nr_sleep_corrections)
615                 printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
616         nr_sleep_corrections = 0;
617 }
618
619 static void test_calibrations(void)
620 {
621         u64 T0, T1;
622
623         T0 = get_nsecs();
624         burn_nsecs(1e6);
625         T1 = get_nsecs();
626
627         printf("the run test took %Ld nsecs\n", T1-T0);
628
629         T0 = get_nsecs();
630         sleep_nsecs(1e6);
631         T1 = get_nsecs();
632
633         printf("the sleep test took %Ld nsecs\n", T1-T0);
634 }
635
636 static int
637 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
638 {
639         struct thread *thread;
640
641         thread = threads__findnew(event->comm.pid, &threads, &last_match);
642
643         dump_printf("%p [%p]: perf_event_comm: %s:%d\n",
644                 (void *)(offset + head),
645                 (void *)(long)(event->header.size),
646                 event->comm.comm, event->comm.pid);
647
648         if (thread == NULL ||
649             thread__set_comm(thread, event->comm.comm)) {
650                 dump_printf("problem processing perf_event_comm, skipping event.\n");
651                 return -1;
652         }
653         total_comm++;
654
655         return 0;
656 }
657
658
659 struct raw_event_sample {
660         u32 size;
661         char data[0];
662 };
663
664 #define FILL_FIELD(ptr, field, event, data)     \
665         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
666
667 #define FILL_ARRAY(ptr, array, event, data)                     \
668 do {                                                            \
669         void *__array = raw_field_ptr(event, #array, data);     \
670         memcpy(ptr.array, __array, sizeof(ptr.array));  \
671 } while(0)
672
673 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
674 do {                                                            \
675         FILL_FIELD(ptr, common_type, event, data);              \
676         FILL_FIELD(ptr, common_flags, event, data);             \
677         FILL_FIELD(ptr, common_preempt_count, event, data);     \
678         FILL_FIELD(ptr, common_pid, event, data);               \
679         FILL_FIELD(ptr, common_tgid, event, data);              \
680 } while (0)
681
682
683
684 struct trace_switch_event {
685         u32 size;
686
687         u16 common_type;
688         u8 common_flags;
689         u8 common_preempt_count;
690         u32 common_pid;
691         u32 common_tgid;
692
693         char prev_comm[16];
694         u32 prev_pid;
695         u32 prev_prio;
696         u64 prev_state;
697         char next_comm[16];
698         u32 next_pid;
699         u32 next_prio;
700 };
701
702 struct trace_runtime_event {
703         u32 size;
704
705         u16 common_type;
706         u8 common_flags;
707         u8 common_preempt_count;
708         u32 common_pid;
709         u32 common_tgid;
710
711         char comm[16];
712         u32 pid;
713         u64 runtime;
714         u64 vruntime;
715 };
716
717 struct trace_wakeup_event {
718         u32 size;
719
720         u16 common_type;
721         u8 common_flags;
722         u8 common_preempt_count;
723         u32 common_pid;
724         u32 common_tgid;
725
726         char comm[16];
727         u32 pid;
728
729         u32 prio;
730         u32 success;
731         u32 cpu;
732 };
733
734 struct trace_fork_event {
735         u32 size;
736
737         u16 common_type;
738         u8 common_flags;
739         u8 common_preempt_count;
740         u32 common_pid;
741         u32 common_tgid;
742
743         char parent_comm[16];
744         u32 parent_pid;
745         char child_comm[16];
746         u32 child_pid;
747 };
748
749 struct trace_sched_handler {
750         void (*switch_event)(struct trace_switch_event *,
751                              struct event *,
752                              int cpu,
753                              u64 timestamp,
754                              struct thread *thread);
755
756         void (*runtime_event)(struct trace_runtime_event *,
757                               struct event *,
758                               int cpu,
759                               u64 timestamp,
760                               struct thread *thread);
761
762         void (*wakeup_event)(struct trace_wakeup_event *,
763                              struct event *,
764                              int cpu,
765                              u64 timestamp,
766                              struct thread *thread);
767
768         void (*fork_event)(struct trace_fork_event *,
769                            struct event *,
770                            int cpu,
771                            u64 timestamp,
772                            struct thread *thread);
773 };
774
775
776 static void
777 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
778                     struct event *event,
779                     int cpu __used,
780                     u64 timestamp __used,
781                     struct thread *thread __used)
782 {
783         struct task_desc *waker, *wakee;
784
785         if (verbose) {
786                 printf("sched_wakeup event %p\n", event);
787
788                 printf(" ... pid %d woke up %s/%d\n",
789                         wakeup_event->common_pid,
790                         wakeup_event->comm,
791                         wakeup_event->pid);
792         }
793
794         waker = register_pid(wakeup_event->common_pid, "<unknown>");
795         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
796
797         add_sched_event_wakeup(waker, timestamp, wakee);
798 }
799
800 static u64 cpu_last_switched[MAX_CPUS];
801
802 static void
803 replay_switch_event(struct trace_switch_event *switch_event,
804                     struct event *event,
805                     int cpu,
806                     u64 timestamp,
807                     struct thread *thread __used)
808 {
809         struct task_desc *prev, *next;
810         u64 timestamp0;
811         s64 delta;
812
813         if (verbose)
814                 printf("sched_switch event %p\n", event);
815
816         if (cpu >= MAX_CPUS || cpu < 0)
817                 return;
818
819         timestamp0 = cpu_last_switched[cpu];
820         if (timestamp0)
821                 delta = timestamp - timestamp0;
822         else
823                 delta = 0;
824
825         if (delta < 0)
826                 die("hm, delta: %Ld < 0 ?\n", delta);
827
828         if (verbose) {
829                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
830                         switch_event->prev_comm, switch_event->prev_pid,
831                         switch_event->next_comm, switch_event->next_pid,
832                         delta);
833         }
834
835         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
836         next = register_pid(switch_event->next_pid, switch_event->next_comm);
837
838         cpu_last_switched[cpu] = timestamp;
839
840         add_sched_event_run(prev, timestamp, delta);
841         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
842 }
843
844
845 static void
846 replay_fork_event(struct trace_fork_event *fork_event,
847                   struct event *event,
848                   int cpu __used,
849                   u64 timestamp __used,
850                   struct thread *thread __used)
851 {
852         if (verbose) {
853                 printf("sched_fork event %p\n", event);
854                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
855                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
856         }
857         register_pid(fork_event->parent_pid, fork_event->parent_comm);
858         register_pid(fork_event->child_pid, fork_event->child_comm);
859 }
860
861 static struct trace_sched_handler replay_ops  = {
862         .wakeup_event           = replay_wakeup_event,
863         .switch_event           = replay_switch_event,
864         .fork_event             = replay_fork_event,
865 };
866
867 struct sort_dimension {
868         const char              *name;
869         sort_fn_t               cmp;
870         struct list_head        list;
871 };
872
873 static LIST_HEAD(cmp_pid);
874
875 static int
876 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
877 {
878         struct sort_dimension *sort;
879         int ret = 0;
880
881         BUG_ON(list_empty(list));
882
883         list_for_each_entry(sort, list, list) {
884                 ret = sort->cmp(l, r);
885                 if (ret)
886                         return ret;
887         }
888
889         return ret;
890 }
891
892 static struct work_atoms *
893 thread_atoms_search(struct rb_root *root, struct thread *thread,
894                          struct list_head *sort_list)
895 {
896         struct rb_node *node = root->rb_node;
897         struct work_atoms key = { .thread = thread };
898
899         while (node) {
900                 struct work_atoms *atoms;
901                 int cmp;
902
903                 atoms = container_of(node, struct work_atoms, node);
904
905                 cmp = thread_lat_cmp(sort_list, &key, atoms);
906                 if (cmp > 0)
907                         node = node->rb_left;
908                 else if (cmp < 0)
909                         node = node->rb_right;
910                 else {
911                         BUG_ON(thread != atoms->thread);
912                         return atoms;
913                 }
914         }
915         return NULL;
916 }
917
918 static void
919 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
920                          struct list_head *sort_list)
921 {
922         struct rb_node **new = &(root->rb_node), *parent = NULL;
923
924         while (*new) {
925                 struct work_atoms *this;
926                 int cmp;
927
928                 this = container_of(*new, struct work_atoms, node);
929                 parent = *new;
930
931                 cmp = thread_lat_cmp(sort_list, data, this);
932
933                 if (cmp > 0)
934                         new = &((*new)->rb_left);
935                 else
936                         new = &((*new)->rb_right);
937         }
938
939         rb_link_node(&data->node, parent, new);
940         rb_insert_color(&data->node, root);
941 }
942
943 static void thread_atoms_insert(struct thread *thread)
944 {
945         struct work_atoms *atoms;
946
947         atoms = calloc(sizeof(*atoms), 1);
948         if (!atoms)
949                 die("No memory");
950
951         atoms->thread = thread;
952         INIT_LIST_HEAD(&atoms->work_list);
953         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
954 }
955
956 static void
957 latency_fork_event(struct trace_fork_event *fork_event __used,
958                    struct event *event __used,
959                    int cpu __used,
960                    u64 timestamp __used,
961                    struct thread *thread __used)
962 {
963         /* should insert the newcomer */
964 }
965
966 __used
967 static char sched_out_state(struct trace_switch_event *switch_event)
968 {
969         const char *str = TASK_STATE_TO_CHAR_STR;
970
971         return str[switch_event->prev_state];
972 }
973
974 static void
975 add_sched_out_event(struct work_atoms *atoms,
976                     char run_state,
977                     u64 timestamp)
978 {
979         struct work_atom *atom;
980
981         atom = calloc(sizeof(*atom), 1);
982         if (!atom)
983                 die("Non memory");
984
985         atom->sched_out_time = timestamp;
986
987         if (run_state == 'R') {
988                 atom->state = THREAD_WAIT_CPU;
989                 atom->wake_up_time = atom->sched_out_time;
990         }
991
992         list_add_tail(&atom->list, &atoms->work_list);
993 }
994
995 static void
996 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
997 {
998         struct work_atom *atom;
999
1000         BUG_ON(list_empty(&atoms->work_list));
1001
1002         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1003
1004         atom->runtime += delta;
1005         atoms->total_runtime += delta;
1006 }
1007
1008 static void
1009 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1010 {
1011         struct work_atom *atom;
1012         u64 delta;
1013
1014         if (list_empty(&atoms->work_list))
1015                 return;
1016
1017         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1018
1019         if (atom->state != THREAD_WAIT_CPU)
1020                 return;
1021
1022         if (timestamp < atom->wake_up_time) {
1023                 atom->state = THREAD_IGNORE;
1024                 return;
1025         }
1026
1027         atom->state = THREAD_SCHED_IN;
1028         atom->sched_in_time = timestamp;
1029
1030         delta = atom->sched_in_time - atom->wake_up_time;
1031         atoms->total_lat += delta;
1032         if (delta > atoms->max_lat)
1033                 atoms->max_lat = delta;
1034         atoms->nb_atoms++;
1035 }
1036
1037 static void
1038 latency_switch_event(struct trace_switch_event *switch_event,
1039                      struct event *event __used,
1040                      int cpu,
1041                      u64 timestamp,
1042                      struct thread *thread __used)
1043 {
1044         struct work_atoms *out_events, *in_events;
1045         struct thread *sched_out, *sched_in;
1046         u64 timestamp0;
1047         s64 delta;
1048
1049         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1050
1051         timestamp0 = cpu_last_switched[cpu];
1052         cpu_last_switched[cpu] = timestamp;
1053         if (timestamp0)
1054                 delta = timestamp - timestamp0;
1055         else
1056                 delta = 0;
1057
1058         if (delta < 0)
1059                 die("hm, delta: %Ld < 0 ?\n", delta);
1060
1061
1062         sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1063         sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1064
1065         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1066         if (!out_events) {
1067                 thread_atoms_insert(sched_out);
1068                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1069                 if (!out_events)
1070                         die("out-event: Internal tree error");
1071         }
1072         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1073
1074         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1075         if (!in_events) {
1076                 thread_atoms_insert(sched_in);
1077                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1078                 if (!in_events)
1079                         die("in-event: Internal tree error");
1080                 /*
1081                  * Take came in we have not heard about yet,
1082                  * add in an initial atom in runnable state:
1083                  */
1084                 add_sched_out_event(in_events, 'R', timestamp);
1085         }
1086         add_sched_in_event(in_events, timestamp);
1087 }
1088
1089 static void
1090 latency_runtime_event(struct trace_runtime_event *runtime_event,
1091                      struct event *event __used,
1092                      int cpu,
1093                      u64 timestamp,
1094                      struct thread *this_thread __used)
1095 {
1096         struct work_atoms *atoms;
1097         struct thread *thread;
1098
1099         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1100
1101         thread = threads__findnew(runtime_event->pid, &threads, &last_match);
1102         atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1103         if (!atoms) {
1104                 thread_atoms_insert(thread);
1105                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1106                 if (!atoms)
1107                         die("in-event: Internal tree error");
1108                 add_sched_out_event(atoms, 'R', timestamp);
1109         }
1110
1111         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1112 }
1113
1114 static void
1115 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1116                      struct event *__event __used,
1117                      int cpu __used,
1118                      u64 timestamp,
1119                      struct thread *thread __used)
1120 {
1121         struct work_atoms *atoms;
1122         struct work_atom *atom;
1123         struct thread *wakee;
1124
1125         /* Note for later, it may be interesting to observe the failing cases */
1126         if (!wakeup_event->success)
1127                 return;
1128
1129         wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
1130         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1131         if (!atoms) {
1132                 thread_atoms_insert(wakee);
1133                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1134                 if (!atoms)
1135                         die("wakeup-event: Internal tree error");
1136                 add_sched_out_event(atoms, 'S', timestamp);
1137         }
1138
1139         BUG_ON(list_empty(&atoms->work_list));
1140
1141         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1142
1143         if (atom->state != THREAD_SLEEPING)
1144                 nr_state_machine_bugs++;
1145
1146         nr_timestamps++;
1147         if (atom->sched_out_time > timestamp) {
1148                 nr_unordered_timestamps++;
1149                 return;
1150         }
1151
1152         atom->state = THREAD_WAIT_CPU;
1153         atom->wake_up_time = timestamp;
1154 }
1155
1156 static struct trace_sched_handler lat_ops  = {
1157         .wakeup_event           = latency_wakeup_event,
1158         .switch_event           = latency_switch_event,
1159         .runtime_event          = latency_runtime_event,
1160         .fork_event             = latency_fork_event,
1161 };
1162
1163 static void output_lat_thread(struct work_atoms *work_list)
1164 {
1165         int i;
1166         int ret;
1167         u64 avg;
1168
1169         if (!work_list->nb_atoms)
1170                 return;
1171         /*
1172          * Ignore idle threads:
1173          */
1174         if (!strcmp(work_list->thread->comm, "swapper"))
1175                 return;
1176
1177         all_runtime += work_list->total_runtime;
1178         all_count += work_list->nb_atoms;
1179
1180         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1181
1182         for (i = 0; i < 24 - ret; i++)
1183                 printf(" ");
1184
1185         avg = work_list->total_lat / work_list->nb_atoms;
1186
1187         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1188               (double)work_list->total_runtime / 1e6,
1189                  work_list->nb_atoms, (double)avg / 1e6,
1190                  (double)work_list->max_lat / 1e6);
1191 }
1192
1193 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1194 {
1195         if (l->thread->pid < r->thread->pid)
1196                 return -1;
1197         if (l->thread->pid > r->thread->pid)
1198                 return 1;
1199
1200         return 0;
1201 }
1202
1203 static struct sort_dimension pid_sort_dimension = {
1204         .name                   = "pid",
1205         .cmp                    = pid_cmp,
1206 };
1207
1208 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1209 {
1210         u64 avgl, avgr;
1211
1212         if (!l->nb_atoms)
1213                 return -1;
1214
1215         if (!r->nb_atoms)
1216                 return 1;
1217
1218         avgl = l->total_lat / l->nb_atoms;
1219         avgr = r->total_lat / r->nb_atoms;
1220
1221         if (avgl < avgr)
1222                 return -1;
1223         if (avgl > avgr)
1224                 return 1;
1225
1226         return 0;
1227 }
1228
1229 static struct sort_dimension avg_sort_dimension = {
1230         .name                   = "avg",
1231         .cmp                    = avg_cmp,
1232 };
1233
1234 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1235 {
1236         if (l->max_lat < r->max_lat)
1237                 return -1;
1238         if (l->max_lat > r->max_lat)
1239                 return 1;
1240
1241         return 0;
1242 }
1243
1244 static struct sort_dimension max_sort_dimension = {
1245         .name                   = "max",
1246         .cmp                    = max_cmp,
1247 };
1248
1249 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1250 {
1251         if (l->nb_atoms < r->nb_atoms)
1252                 return -1;
1253         if (l->nb_atoms > r->nb_atoms)
1254                 return 1;
1255
1256         return 0;
1257 }
1258
1259 static struct sort_dimension switch_sort_dimension = {
1260         .name                   = "switch",
1261         .cmp                    = switch_cmp,
1262 };
1263
1264 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1265 {
1266         if (l->total_runtime < r->total_runtime)
1267                 return -1;
1268         if (l->total_runtime > r->total_runtime)
1269                 return 1;
1270
1271         return 0;
1272 }
1273
1274 static struct sort_dimension runtime_sort_dimension = {
1275         .name                   = "runtime",
1276         .cmp                    = runtime_cmp,
1277 };
1278
1279 static struct sort_dimension *available_sorts[] = {
1280         &pid_sort_dimension,
1281         &avg_sort_dimension,
1282         &max_sort_dimension,
1283         &switch_sort_dimension,
1284         &runtime_sort_dimension,
1285 };
1286
1287 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1288
1289 static LIST_HEAD(sort_list);
1290
1291 static int sort_dimension__add(char *tok, struct list_head *list)
1292 {
1293         int i;
1294
1295         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1296                 if (!strcmp(available_sorts[i]->name, tok)) {
1297                         list_add_tail(&available_sorts[i]->list, list);
1298
1299                         return 0;
1300                 }
1301         }
1302
1303         return -1;
1304 }
1305
1306 static void setup_sorting(void);
1307
1308 static void sort_lat(void)
1309 {
1310         struct rb_node *node;
1311
1312         for (;;) {
1313                 struct work_atoms *data;
1314                 node = rb_first(&atom_root);
1315                 if (!node)
1316                         break;
1317
1318                 rb_erase(node, &atom_root);
1319                 data = rb_entry(node, struct work_atoms, node);
1320                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1321         }
1322 }
1323
1324 static struct trace_sched_handler *trace_handler;
1325
1326 static void
1327 process_sched_wakeup_event(struct raw_event_sample *raw,
1328                            struct event *event,
1329                            int cpu __used,
1330                            u64 timestamp __used,
1331                            struct thread *thread __used)
1332 {
1333         struct trace_wakeup_event wakeup_event;
1334
1335         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1336
1337         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1338         FILL_FIELD(wakeup_event, pid, event, raw->data);
1339         FILL_FIELD(wakeup_event, prio, event, raw->data);
1340         FILL_FIELD(wakeup_event, success, event, raw->data);
1341         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1342
1343         if (trace_handler->wakeup_event)
1344                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1345 }
1346
1347 /*
1348  * Track the current task - that way we can know whether there's any
1349  * weird events, such as a task being switched away that is not current.
1350  */
1351 static int max_cpu;
1352
1353 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1354
1355 static struct thread *curr_thread[MAX_CPUS];
1356
1357 static char next_shortname1 = 'A';
1358 static char next_shortname2 = '0';
1359
1360 static void
1361 map_switch_event(struct trace_switch_event *switch_event,
1362                  struct event *event __used,
1363                  int this_cpu,
1364                  u64 timestamp,
1365                  struct thread *thread __used)
1366 {
1367         struct thread *sched_out, *sched_in;
1368         int new_shortname;
1369         u64 timestamp0;
1370         s64 delta;
1371         int cpu;
1372
1373         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1374
1375         if (this_cpu > max_cpu)
1376                 max_cpu = this_cpu;
1377
1378         timestamp0 = cpu_last_switched[this_cpu];
1379         cpu_last_switched[this_cpu] = timestamp;
1380         if (timestamp0)
1381                 delta = timestamp - timestamp0;
1382         else
1383                 delta = 0;
1384
1385         if (delta < 0)
1386                 die("hm, delta: %Ld < 0 ?\n", delta);
1387
1388
1389         sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1390         sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1391
1392         curr_thread[this_cpu] = sched_in;
1393
1394         printf("  ");
1395
1396         new_shortname = 0;
1397         if (!sched_in->shortname[0]) {
1398                 sched_in->shortname[0] = next_shortname1;
1399                 sched_in->shortname[1] = next_shortname2;
1400
1401                 if (next_shortname1 < 'Z') {
1402                         next_shortname1++;
1403                 } else {
1404                         next_shortname1='A';
1405                         if (next_shortname2 < '9') {
1406                                 next_shortname2++;
1407                         } else {
1408                                 next_shortname2='0';
1409                         }
1410                 }
1411                 new_shortname = 1;
1412         }
1413
1414         for (cpu = 0; cpu <= max_cpu; cpu++) {
1415                 if (cpu != this_cpu)
1416                         printf(" ");
1417                 else
1418                         printf("*");
1419
1420                 if (curr_thread[cpu]) {
1421                         if (curr_thread[cpu]->pid)
1422                                 printf("%2s ", curr_thread[cpu]->shortname);
1423                         else
1424                                 printf(".  ");
1425                 } else
1426                         printf("   ");
1427         }
1428
1429         printf("  %12.6f secs ", (double)timestamp/1e9);
1430         if (new_shortname) {
1431                 printf("%s => %s:%d\n",
1432                         sched_in->shortname, sched_in->comm, sched_in->pid);
1433         } else {
1434                 printf("\n");
1435         }
1436 }
1437
1438
1439 static void
1440 process_sched_switch_event(struct raw_event_sample *raw,
1441                            struct event *event,
1442                            int this_cpu,
1443                            u64 timestamp __used,
1444                            struct thread *thread __used)
1445 {
1446         struct trace_switch_event switch_event;
1447
1448         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1449
1450         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1451         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1452         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1453         FILL_FIELD(switch_event, prev_state, event, raw->data);
1454         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1455         FILL_FIELD(switch_event, next_pid, event, raw->data);
1456         FILL_FIELD(switch_event, next_prio, event, raw->data);
1457
1458         if (curr_pid[this_cpu] != (u32)-1) {
1459                 /*
1460                  * Are we trying to switch away a PID that is
1461                  * not current?
1462                  */
1463                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1464                         nr_context_switch_bugs++;
1465         }
1466         if (trace_handler->switch_event)
1467                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1468
1469         curr_pid[this_cpu] = switch_event.next_pid;
1470 }
1471
1472 static void
1473 process_sched_runtime_event(struct raw_event_sample *raw,
1474                            struct event *event,
1475                            int cpu __used,
1476                            u64 timestamp __used,
1477                            struct thread *thread __used)
1478 {
1479         struct trace_runtime_event runtime_event;
1480
1481         FILL_ARRAY(runtime_event, comm, event, raw->data);
1482         FILL_FIELD(runtime_event, pid, event, raw->data);
1483         FILL_FIELD(runtime_event, runtime, event, raw->data);
1484         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1485
1486         if (trace_handler->runtime_event)
1487                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1488 }
1489
1490 static void
1491 process_sched_fork_event(struct raw_event_sample *raw,
1492                          struct event *event,
1493                          int cpu __used,
1494                          u64 timestamp __used,
1495                          struct thread *thread __used)
1496 {
1497         struct trace_fork_event fork_event;
1498
1499         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1500
1501         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1502         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1503         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1504         FILL_FIELD(fork_event, child_pid, event, raw->data);
1505
1506         if (trace_handler->fork_event)
1507                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1508 }
1509
1510 static void
1511 process_sched_exit_event(struct event *event,
1512                          int cpu __used,
1513                          u64 timestamp __used,
1514                          struct thread *thread __used)
1515 {
1516         if (verbose)
1517                 printf("sched_exit event %p\n", event);
1518 }
1519
1520 static void
1521 process_raw_event(event_t *raw_event __used, void *more_data,
1522                   int cpu, u64 timestamp, struct thread *thread)
1523 {
1524         struct raw_event_sample *raw = more_data;
1525         struct event *event;
1526         int type;
1527
1528         type = trace_parse_common_type(raw->data);
1529         event = trace_find_event(type);
1530
1531         if (!strcmp(event->name, "sched_switch"))
1532                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1533         if (!strcmp(event->name, "sched_stat_runtime"))
1534                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1535         if (!strcmp(event->name, "sched_wakeup"))
1536                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1537         if (!strcmp(event->name, "sched_wakeup_new"))
1538                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1539         if (!strcmp(event->name, "sched_process_fork"))
1540                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1541         if (!strcmp(event->name, "sched_process_exit"))
1542                 process_sched_exit_event(event, cpu, timestamp, thread);
1543 }
1544
1545 static int
1546 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1547 {
1548         struct thread *thread;
1549         u64 ip = event->ip.ip;
1550         u64 timestamp = -1;
1551         u32 cpu = -1;
1552         u64 period = 1;
1553         void *more_data = event->ip.__more_data;
1554
1555         if (!(sample_type & PERF_SAMPLE_RAW))
1556                 return 0;
1557
1558         thread = threads__findnew(event->ip.pid, &threads, &last_match);
1559
1560         if (sample_type & PERF_SAMPLE_TIME) {
1561                 timestamp = *(u64 *)more_data;
1562                 more_data += sizeof(u64);
1563         }
1564
1565         if (sample_type & PERF_SAMPLE_CPU) {
1566                 cpu = *(u32 *)more_data;
1567                 more_data += sizeof(u32);
1568                 more_data += sizeof(u32); /* reserved */
1569         }
1570
1571         if (sample_type & PERF_SAMPLE_PERIOD) {
1572                 period = *(u64 *)more_data;
1573                 more_data += sizeof(u64);
1574         }
1575
1576         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1577                 (void *)(offset + head),
1578                 (void *)(long)(event->header.size),
1579                 event->header.misc,
1580                 event->ip.pid, event->ip.tid,
1581                 (void *)(long)ip,
1582                 (long long)period);
1583
1584         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1585
1586         if (thread == NULL) {
1587                 eprintf("problem processing %d event, skipping it.\n",
1588                         event->header.type);
1589                 return -1;
1590         }
1591
1592         process_raw_event(event, more_data, cpu, timestamp, thread);
1593
1594         return 0;
1595 }
1596
1597 static int
1598 process_lost_event(event_t *event __used,
1599                    unsigned long offset __used,
1600                    unsigned long head __used)
1601 {
1602         nr_lost_chunks++;
1603         nr_lost_events += event->lost.lost;
1604
1605         return 0;
1606 }
1607
1608 static int sample_type_check(u64 type)
1609 {
1610         sample_type = type;
1611
1612         if (!(sample_type & PERF_SAMPLE_RAW)) {
1613                 fprintf(stderr,
1614                         "No trace sample to read. Did you call perf record "
1615                         "without -R?");
1616                 return -1;
1617         }
1618
1619         return 0;
1620 }
1621
1622 static struct perf_file_handler file_handler = {
1623         .process_sample_event   = process_sample_event,
1624         .process_comm_event     = process_comm_event,
1625         .process_lost_event     = process_lost_event,
1626         .sample_type_check      = sample_type_check,
1627 };
1628
1629 static int read_events(void)
1630 {
1631         register_idle_thread(&threads, &last_match);
1632         register_perf_file_handler(&file_handler);
1633
1634         return mmap_dispatch_perf_file(&header, input_name, 0, 0, &cwdlen, &cwd);
1635 }
1636
1637 static void print_bad_events(void)
1638 {
1639         if (nr_unordered_timestamps && nr_timestamps) {
1640                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1641                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1642                         nr_unordered_timestamps, nr_timestamps);
1643         }
1644         if (nr_lost_events && nr_events) {
1645                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1646                         (double)nr_lost_events/(double)nr_events*100.0,
1647                         nr_lost_events, nr_events, nr_lost_chunks);
1648         }
1649         if (nr_state_machine_bugs && nr_timestamps) {
1650                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1651                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1652                         nr_state_machine_bugs, nr_timestamps);
1653                 if (nr_lost_events)
1654                         printf(" (due to lost events?)");
1655                 printf("\n");
1656         }
1657         if (nr_context_switch_bugs && nr_timestamps) {
1658                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1659                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1660                         nr_context_switch_bugs, nr_timestamps);
1661                 if (nr_lost_events)
1662                         printf(" (due to lost events?)");
1663                 printf("\n");
1664         }
1665 }
1666
1667 static void __cmd_lat(void)
1668 {
1669         struct rb_node *next;
1670
1671         setup_pager();
1672         read_events();
1673         sort_lat();
1674
1675         printf("\n -----------------------------------------------------------------------------------------\n");
1676         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1677         printf(" -----------------------------------------------------------------------------------------\n");
1678
1679         next = rb_first(&sorted_atom_root);
1680
1681         while (next) {
1682                 struct work_atoms *work_list;
1683
1684                 work_list = rb_entry(next, struct work_atoms, node);
1685                 output_lat_thread(work_list);
1686                 next = rb_next(next);
1687         }
1688
1689         printf(" -----------------------------------------------------------------------------------------\n");
1690         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1691                 (double)all_runtime/1e6, all_count);
1692
1693         printf(" ---------------------------------------------------\n");
1694
1695         print_bad_events();
1696         printf("\n");
1697
1698 }
1699
1700 static struct trace_sched_handler map_ops  = {
1701         .wakeup_event           = NULL,
1702         .switch_event           = map_switch_event,
1703         .runtime_event          = NULL,
1704         .fork_event             = NULL,
1705 };
1706
1707 static void __cmd_map(void)
1708 {
1709         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1710
1711         setup_pager();
1712         read_events();
1713         print_bad_events();
1714 }
1715
1716 static void __cmd_replay(void)
1717 {
1718         unsigned long i;
1719
1720         calibrate_run_measurement_overhead();
1721         calibrate_sleep_measurement_overhead();
1722
1723         test_calibrations();
1724
1725         read_events();
1726
1727         printf("nr_run_events:        %ld\n", nr_run_events);
1728         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1729         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1730
1731         if (targetless_wakeups)
1732                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1733         if (multitarget_wakeups)
1734                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1735         if (nr_run_events_optimized)
1736                 printf("run atoms optimized: %ld\n",
1737                         nr_run_events_optimized);
1738
1739         print_task_traces();
1740         add_cross_task_wakeups();
1741
1742         create_tasks();
1743         printf("------------------------------------------------------------\n");
1744         for (i = 0; i < replay_repeat; i++)
1745                 run_one_test();
1746 }
1747
1748
1749 static const char * const sched_usage[] = {
1750         "perf sched [<options>] {record|latency|map|replay|trace}",
1751         NULL
1752 };
1753
1754 static const struct option sched_options[] = {
1755         OPT_STRING('i', "input", &input_name, "file",
1756                     "input file name"),
1757         OPT_BOOLEAN('v', "verbose", &verbose,
1758                     "be more verbose (show symbol address, etc)"),
1759         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1760                     "dump raw trace in ASCII"),
1761         OPT_END()
1762 };
1763
1764 static const char * const latency_usage[] = {
1765         "perf sched latency [<options>]",
1766         NULL
1767 };
1768
1769 static const struct option latency_options[] = {
1770         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1771                    "sort by key(s): runtime, switch, avg, max"),
1772         OPT_BOOLEAN('v', "verbose", &verbose,
1773                     "be more verbose (show symbol address, etc)"),
1774         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1775                     "dump raw trace in ASCII"),
1776         OPT_END()
1777 };
1778
1779 static const char * const replay_usage[] = {
1780         "perf sched replay [<options>]",
1781         NULL
1782 };
1783
1784 static const struct option replay_options[] = {
1785         OPT_INTEGER('r', "repeat", &replay_repeat,
1786                     "repeat the workload replay N times (-1: infinite)"),
1787         OPT_BOOLEAN('v', "verbose", &verbose,
1788                     "be more verbose (show symbol address, etc)"),
1789         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1790                     "dump raw trace in ASCII"),
1791         OPT_END()
1792 };
1793
1794 static void setup_sorting(void)
1795 {
1796         char *tmp, *tok, *str = strdup(sort_order);
1797
1798         for (tok = strtok_r(str, ", ", &tmp);
1799                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1800                 if (sort_dimension__add(tok, &sort_list) < 0) {
1801                         error("Unknown --sort key: `%s'", tok);
1802                         usage_with_options(latency_usage, latency_options);
1803                 }
1804         }
1805
1806         free(str);
1807
1808         sort_dimension__add((char *)"pid", &cmp_pid);
1809 }
1810
1811 static const char *record_args[] = {
1812         "record",
1813         "-a",
1814         "-R",
1815         "-M",
1816         "-f",
1817         "-m", "1024",
1818         "-c", "1",
1819         "-e", "sched:sched_switch:r",
1820         "-e", "sched:sched_stat_wait:r",
1821         "-e", "sched:sched_stat_sleep:r",
1822         "-e", "sched:sched_stat_iowait:r",
1823         "-e", "sched:sched_stat_runtime:r",
1824         "-e", "sched:sched_process_exit:r",
1825         "-e", "sched:sched_process_fork:r",
1826         "-e", "sched:sched_wakeup:r",
1827         "-e", "sched:sched_migrate_task:r",
1828 };
1829
1830 static int __cmd_record(int argc, const char **argv)
1831 {
1832         unsigned int rec_argc, i, j;
1833         const char **rec_argv;
1834
1835         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1836         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1837
1838         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1839                 rec_argv[i] = strdup(record_args[i]);
1840
1841         for (j = 1; j < (unsigned int)argc; j++, i++)
1842                 rec_argv[i] = argv[j];
1843
1844         BUG_ON(i != rec_argc);
1845
1846         return cmd_record(i, rec_argv, NULL);
1847 }
1848
1849 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1850 {
1851         symbol__init();
1852
1853         argc = parse_options(argc, argv, sched_options, sched_usage,
1854                              PARSE_OPT_STOP_AT_NON_OPTION);
1855         if (!argc)
1856                 usage_with_options(sched_usage, sched_options);
1857
1858         if (!strncmp(argv[0], "rec", 3)) {
1859                 return __cmd_record(argc, argv);
1860         } else if (!strncmp(argv[0], "lat", 3)) {
1861                 trace_handler = &lat_ops;
1862                 if (argc > 1) {
1863                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1864                         if (argc)
1865                                 usage_with_options(latency_usage, latency_options);
1866                 }
1867                 setup_sorting();
1868                 __cmd_lat();
1869         } else if (!strcmp(argv[0], "map")) {
1870                 trace_handler = &map_ops;
1871                 setup_sorting();
1872                 __cmd_map();
1873         } else if (!strncmp(argv[0], "rep", 3)) {
1874                 trace_handler = &replay_ops;
1875                 if (argc) {
1876                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1877                         if (argc)
1878                                 usage_with_options(replay_usage, replay_options);
1879                 }
1880                 __cmd_replay();
1881         } else if (!strcmp(argv[0], "trace")) {
1882                 /*
1883                  * Aliased to 'perf trace' for now:
1884                  */
1885                 return cmd_trace(argc, argv, prefix);
1886         } else {
1887                 usage_with_options(sched_usage, sched_options);
1888         }
1889
1890         return 0;
1891 }