perf sched: Account for lost events, increase default buffering
[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
15 #include <sys/types.h>
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 static int                      input;
24 static unsigned long            page_size;
25 static unsigned long            mmap_window = 32;
26
27 static unsigned long            total_comm = 0;
28
29 static struct rb_root           threads;
30 static struct thread            *last_match;
31
32 static struct perf_header       *header;
33 static u64                      sample_type;
34
35 static char                     default_sort_order[] = "avg, max, switch, runtime";
36 static char                     *sort_order = default_sort_order;
37
38 #define PR_SET_NAME             15               /* Set process name */
39 #define MAX_CPUS                4096
40
41 #define BUG_ON(x)               assert(!(x))
42
43 static u64                      run_measurement_overhead;
44 static u64                      sleep_measurement_overhead;
45
46 #define COMM_LEN                20
47 #define SYM_LEN                 129
48
49 #define MAX_PID                 65536
50
51 static unsigned long            nr_tasks;
52
53 struct sched_atom;
54
55 struct task_desc {
56         unsigned long           nr;
57         unsigned long           pid;
58         char                    comm[COMM_LEN];
59
60         unsigned long           nr_events;
61         unsigned long           curr_event;
62         struct sched_atom       **atoms;
63
64         pthread_t               thread;
65         sem_t                   sleep_sem;
66
67         sem_t                   ready_for_work;
68         sem_t                   work_done_sem;
69
70         u64                     cpu_usage;
71 };
72
73 enum sched_event_type {
74         SCHED_EVENT_RUN,
75         SCHED_EVENT_SLEEP,
76         SCHED_EVENT_WAKEUP,
77 };
78
79 struct sched_atom {
80         enum sched_event_type   type;
81         u64                     timestamp;
82         u64                     duration;
83         unsigned long           nr;
84         int                     specific_wait;
85         sem_t                   *wait_sem;
86         struct task_desc        *wakee;
87 };
88
89 static struct task_desc         *pid_to_task[MAX_PID];
90
91 static struct task_desc         **tasks;
92
93 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
94 static u64                      start_time;
95
96 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
97
98 static unsigned long            nr_run_events;
99 static unsigned long            nr_sleep_events;
100 static unsigned long            nr_wakeup_events;
101
102 static unsigned long            nr_sleep_corrections;
103 static unsigned long            nr_run_events_optimized;
104
105 static unsigned long            targetless_wakeups;
106 static unsigned long            multitarget_wakeups;
107
108 static u64                      cpu_usage;
109 static u64                      runavg_cpu_usage;
110 static u64                      parent_cpu_usage;
111 static u64                      runavg_parent_cpu_usage;
112
113 static unsigned long            nr_runs;
114 static u64                      sum_runtime;
115 static u64                      sum_fluct;
116 static u64                      run_avg;
117
118 static unsigned long            replay_repeat = 10;
119 static unsigned long            nr_timestamps;
120 static unsigned long            nr_unordered_timestamps;
121 static unsigned long            nr_state_machine_bugs;
122 static unsigned long            nr_events;
123 static unsigned long            nr_lost_chunks;
124 static unsigned long            nr_lost_events;
125
126 #define TASK_STATE_TO_CHAR_STR "RSDTtZX"
127
128 enum thread_state {
129         THREAD_SLEEPING = 0,
130         THREAD_WAIT_CPU,
131         THREAD_SCHED_IN,
132         THREAD_IGNORE
133 };
134
135 struct work_atom {
136         struct list_head        list;
137         enum thread_state       state;
138         u64                     sched_out_time;
139         u64                     wake_up_time;
140         u64                     sched_in_time;
141         u64                     runtime;
142 };
143
144 struct work_atoms {
145         struct list_head        work_list;
146         struct thread           *thread;
147         struct rb_node          node;
148         u64                     max_lat;
149         u64                     total_lat;
150         u64                     nb_atoms;
151         u64                     total_runtime;
152 };
153
154 typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
155
156 static struct rb_root           atom_root, sorted_atom_root;
157
158 static u64                      all_runtime;
159 static u64                      all_count;
160
161 static int read_events(void);
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 void __cmd_replay(void)
637 {
638         unsigned long i;
639
640         calibrate_run_measurement_overhead();
641         calibrate_sleep_measurement_overhead();
642
643         test_calibrations();
644
645         read_events();
646
647         printf("nr_run_events:        %ld\n", nr_run_events);
648         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
649         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
650
651         if (targetless_wakeups)
652                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
653         if (multitarget_wakeups)
654                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
655         if (nr_run_events_optimized)
656                 printf("run atoms optimized: %ld\n",
657                         nr_run_events_optimized);
658
659         print_task_traces();
660         add_cross_task_wakeups();
661
662         create_tasks();
663         printf("------------------------------------------------------------\n");
664         for (i = 0; i < replay_repeat; i++)
665                 run_one_test();
666 }
667
668 static int
669 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
670 {
671         struct thread *thread;
672
673         thread = threads__findnew(event->comm.pid, &threads, &last_match);
674
675         dump_printf("%p [%p]: perf_event_comm: %s:%d\n",
676                 (void *)(offset + head),
677                 (void *)(long)(event->header.size),
678                 event->comm.comm, event->comm.pid);
679
680         if (thread == NULL ||
681             thread__set_comm(thread, event->comm.comm)) {
682                 dump_printf("problem processing perf_event_comm, skipping event.\n");
683                 return -1;
684         }
685         total_comm++;
686
687         return 0;
688 }
689
690
691 struct raw_event_sample {
692         u32 size;
693         char data[0];
694 };
695
696 #define FILL_FIELD(ptr, field, event, data)     \
697         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
698
699 #define FILL_ARRAY(ptr, array, event, data)                     \
700 do {                                                            \
701         void *__array = raw_field_ptr(event, #array, data);     \
702         memcpy(ptr.array, __array, sizeof(ptr.array));  \
703 } while(0)
704
705 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
706 do {                                                            \
707         FILL_FIELD(ptr, common_type, event, data);              \
708         FILL_FIELD(ptr, common_flags, event, data);             \
709         FILL_FIELD(ptr, common_preempt_count, event, data);     \
710         FILL_FIELD(ptr, common_pid, event, data);               \
711         FILL_FIELD(ptr, common_tgid, event, data);              \
712 } while (0)
713
714
715
716 struct trace_switch_event {
717         u32 size;
718
719         u16 common_type;
720         u8 common_flags;
721         u8 common_preempt_count;
722         u32 common_pid;
723         u32 common_tgid;
724
725         char prev_comm[16];
726         u32 prev_pid;
727         u32 prev_prio;
728         u64 prev_state;
729         char next_comm[16];
730         u32 next_pid;
731         u32 next_prio;
732 };
733
734 struct trace_runtime_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 comm[16];
744         u32 pid;
745         u64 runtime;
746         u64 vruntime;
747 };
748
749 struct trace_wakeup_event {
750         u32 size;
751
752         u16 common_type;
753         u8 common_flags;
754         u8 common_preempt_count;
755         u32 common_pid;
756         u32 common_tgid;
757
758         char comm[16];
759         u32 pid;
760
761         u32 prio;
762         u32 success;
763         u32 cpu;
764 };
765
766 struct trace_fork_event {
767         u32 size;
768
769         u16 common_type;
770         u8 common_flags;
771         u8 common_preempt_count;
772         u32 common_pid;
773         u32 common_tgid;
774
775         char parent_comm[16];
776         u32 parent_pid;
777         char child_comm[16];
778         u32 child_pid;
779 };
780
781 struct trace_sched_handler {
782         void (*switch_event)(struct trace_switch_event *,
783                              struct event *,
784                              int cpu,
785                              u64 timestamp,
786                              struct thread *thread);
787
788         void (*runtime_event)(struct trace_runtime_event *,
789                               struct event *,
790                               int cpu,
791                               u64 timestamp,
792                               struct thread *thread);
793
794         void (*wakeup_event)(struct trace_wakeup_event *,
795                              struct event *,
796                              int cpu,
797                              u64 timestamp,
798                              struct thread *thread);
799
800         void (*fork_event)(struct trace_fork_event *,
801                            struct event *,
802                            int cpu,
803                            u64 timestamp,
804                            struct thread *thread);
805 };
806
807
808 static void
809 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
810                     struct event *event,
811                     int cpu __used,
812                     u64 timestamp __used,
813                     struct thread *thread __used)
814 {
815         struct task_desc *waker, *wakee;
816
817         if (verbose) {
818                 printf("sched_wakeup event %p\n", event);
819
820                 printf(" ... pid %d woke up %s/%d\n",
821                         wakeup_event->common_pid,
822                         wakeup_event->comm,
823                         wakeup_event->pid);
824         }
825
826         waker = register_pid(wakeup_event->common_pid, "<unknown>");
827         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
828
829         add_sched_event_wakeup(waker, timestamp, wakee);
830 }
831
832 static u64 cpu_last_switched[MAX_CPUS];
833
834 static void
835 replay_switch_event(struct trace_switch_event *switch_event,
836                     struct event *event,
837                     int cpu,
838                     u64 timestamp,
839                     struct thread *thread __used)
840 {
841         struct task_desc *prev, *next;
842         u64 timestamp0;
843         s64 delta;
844
845         if (verbose)
846                 printf("sched_switch event %p\n", event);
847
848         if (cpu >= MAX_CPUS || cpu < 0)
849                 return;
850
851         timestamp0 = cpu_last_switched[cpu];
852         if (timestamp0)
853                 delta = timestamp - timestamp0;
854         else
855                 delta = 0;
856
857         if (delta < 0)
858                 die("hm, delta: %Ld < 0 ?\n", delta);
859
860         if (verbose) {
861                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
862                         switch_event->prev_comm, switch_event->prev_pid,
863                         switch_event->next_comm, switch_event->next_pid,
864                         delta);
865         }
866
867         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
868         next = register_pid(switch_event->next_pid, switch_event->next_comm);
869
870         cpu_last_switched[cpu] = timestamp;
871
872         add_sched_event_run(prev, timestamp, delta);
873         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
874 }
875
876
877 static void
878 replay_fork_event(struct trace_fork_event *fork_event,
879                   struct event *event,
880                   int cpu __used,
881                   u64 timestamp __used,
882                   struct thread *thread __used)
883 {
884         if (verbose) {
885                 printf("sched_fork event %p\n", event);
886                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
887                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
888         }
889         register_pid(fork_event->parent_pid, fork_event->parent_comm);
890         register_pid(fork_event->child_pid, fork_event->child_comm);
891 }
892
893 static struct trace_sched_handler replay_ops  = {
894         .wakeup_event           = replay_wakeup_event,
895         .switch_event           = replay_switch_event,
896         .fork_event             = replay_fork_event,
897 };
898
899 struct sort_dimension {
900         const char              *name;
901         sort_fn_t               cmp;
902         struct list_head        list;
903 };
904
905 static LIST_HEAD(cmp_pid);
906
907 static int
908 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
909 {
910         struct sort_dimension *sort;
911         int ret = 0;
912
913         BUG_ON(list_empty(list));
914
915         list_for_each_entry(sort, list, list) {
916                 ret = sort->cmp(l, r);
917                 if (ret)
918                         return ret;
919         }
920
921         return ret;
922 }
923
924 static struct work_atoms *
925 thread_atoms_search(struct rb_root *root, struct thread *thread,
926                          struct list_head *sort_list)
927 {
928         struct rb_node *node = root->rb_node;
929         struct work_atoms key = { .thread = thread };
930
931         while (node) {
932                 struct work_atoms *atoms;
933                 int cmp;
934
935                 atoms = container_of(node, struct work_atoms, node);
936
937                 cmp = thread_lat_cmp(sort_list, &key, atoms);
938                 if (cmp > 0)
939                         node = node->rb_left;
940                 else if (cmp < 0)
941                         node = node->rb_right;
942                 else {
943                         BUG_ON(thread != atoms->thread);
944                         return atoms;
945                 }
946         }
947         return NULL;
948 }
949
950 static void
951 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
952                          struct list_head *sort_list)
953 {
954         struct rb_node **new = &(root->rb_node), *parent = NULL;
955
956         while (*new) {
957                 struct work_atoms *this;
958                 int cmp;
959
960                 this = container_of(*new, struct work_atoms, node);
961                 parent = *new;
962
963                 cmp = thread_lat_cmp(sort_list, data, this);
964
965                 if (cmp > 0)
966                         new = &((*new)->rb_left);
967                 else
968                         new = &((*new)->rb_right);
969         }
970
971         rb_link_node(&data->node, parent, new);
972         rb_insert_color(&data->node, root);
973 }
974
975 static void thread_atoms_insert(struct thread *thread)
976 {
977         struct work_atoms *atoms;
978
979         atoms = calloc(sizeof(*atoms), 1);
980         if (!atoms)
981                 die("No memory");
982
983         atoms->thread = thread;
984         INIT_LIST_HEAD(&atoms->work_list);
985         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
986 }
987
988 static void
989 latency_fork_event(struct trace_fork_event *fork_event __used,
990                    struct event *event __used,
991                    int cpu __used,
992                    u64 timestamp __used,
993                    struct thread *thread __used)
994 {
995         /* should insert the newcomer */
996 }
997
998 __used
999 static char sched_out_state(struct trace_switch_event *switch_event)
1000 {
1001         const char *str = TASK_STATE_TO_CHAR_STR;
1002
1003         return str[switch_event->prev_state];
1004 }
1005
1006 static void
1007 add_sched_out_event(struct work_atoms *atoms,
1008                     char run_state,
1009                     u64 timestamp)
1010 {
1011         struct work_atom *atom;
1012
1013         atom = calloc(sizeof(*atom), 1);
1014         if (!atom)
1015                 die("Non memory");
1016
1017         atom->sched_out_time = timestamp;
1018
1019         if (run_state == 'R') {
1020                 atom->state = THREAD_WAIT_CPU;
1021                 atom->wake_up_time = atom->sched_out_time;
1022         }
1023
1024         list_add_tail(&atom->list, &atoms->work_list);
1025 }
1026
1027 static void
1028 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
1029 {
1030         struct work_atom *atom;
1031
1032         BUG_ON(list_empty(&atoms->work_list));
1033
1034         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1035
1036         atom->runtime += delta;
1037         atoms->total_runtime += delta;
1038 }
1039
1040 static void
1041 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1042 {
1043         struct work_atom *atom;
1044         u64 delta;
1045
1046         if (list_empty(&atoms->work_list))
1047                 return;
1048
1049         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1050
1051         if (atom->state != THREAD_WAIT_CPU)
1052                 return;
1053
1054         if (timestamp < atom->wake_up_time) {
1055                 atom->state = THREAD_IGNORE;
1056                 return;
1057         }
1058
1059         atom->state = THREAD_SCHED_IN;
1060         atom->sched_in_time = timestamp;
1061
1062         delta = atom->sched_in_time - atom->wake_up_time;
1063         atoms->total_lat += delta;
1064         if (delta > atoms->max_lat)
1065                 atoms->max_lat = delta;
1066         atoms->nb_atoms++;
1067 }
1068
1069 static void
1070 latency_switch_event(struct trace_switch_event *switch_event,
1071                      struct event *event __used,
1072                      int cpu,
1073                      u64 timestamp,
1074                      struct thread *thread __used)
1075 {
1076         struct work_atoms *out_events, *in_events;
1077         struct thread *sched_out, *sched_in;
1078         u64 timestamp0;
1079         s64 delta;
1080
1081         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1082
1083         timestamp0 = cpu_last_switched[cpu];
1084         cpu_last_switched[cpu] = timestamp;
1085         if (timestamp0)
1086                 delta = timestamp - timestamp0;
1087         else
1088                 delta = 0;
1089
1090         if (delta < 0)
1091                 die("hm, delta: %Ld < 0 ?\n", delta);
1092
1093
1094         sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
1095         sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
1096
1097         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1098         if (!out_events) {
1099                 thread_atoms_insert(sched_out);
1100                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1101                 if (!out_events)
1102                         die("out-event: Internal tree error");
1103         }
1104         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1105
1106         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1107         if (!in_events) {
1108                 thread_atoms_insert(sched_in);
1109                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1110                 if (!in_events)
1111                         die("in-event: Internal tree error");
1112                 /*
1113                  * Take came in we have not heard about yet,
1114                  * add in an initial atom in runnable state:
1115                  */
1116                 add_sched_out_event(in_events, 'R', timestamp);
1117         }
1118         add_sched_in_event(in_events, timestamp);
1119 }
1120
1121 static void
1122 latency_runtime_event(struct trace_runtime_event *runtime_event,
1123                      struct event *event __used,
1124                      int cpu,
1125                      u64 timestamp,
1126                      struct thread *this_thread __used)
1127 {
1128         struct work_atoms *atoms;
1129         struct thread *thread;
1130
1131         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1132
1133         thread = threads__findnew(runtime_event->pid, &threads, &last_match);
1134         atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1135         if (!atoms) {
1136                 thread_atoms_insert(thread);
1137                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1138                 if (!atoms)
1139                         die("in-event: Internal tree error");
1140                 add_sched_out_event(atoms, 'R', timestamp);
1141         }
1142
1143         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1144 }
1145
1146 static void
1147 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1148                      struct event *__event __used,
1149                      int cpu __used,
1150                      u64 timestamp,
1151                      struct thread *thread __used)
1152 {
1153         struct work_atoms *atoms;
1154         struct work_atom *atom;
1155         struct thread *wakee;
1156
1157         /* Note for later, it may be interesting to observe the failing cases */
1158         if (!wakeup_event->success)
1159                 return;
1160
1161         wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
1162         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1163         if (!atoms) {
1164                 thread_atoms_insert(wakee);
1165                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1166                 if (!atoms)
1167                         die("wakeup-event: Internal tree error");
1168                 add_sched_out_event(atoms, 'S', timestamp);
1169         }
1170
1171         BUG_ON(list_empty(&atoms->work_list));
1172
1173         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1174
1175         if (atom->state != THREAD_SLEEPING)
1176                 nr_state_machine_bugs++;
1177
1178         nr_timestamps++;
1179         if (atom->sched_out_time > timestamp) {
1180                 nr_unordered_timestamps++;
1181                 return;
1182         }
1183
1184         atom->state = THREAD_WAIT_CPU;
1185         atom->wake_up_time = timestamp;
1186 }
1187
1188 static struct trace_sched_handler lat_ops  = {
1189         .wakeup_event           = latency_wakeup_event,
1190         .switch_event           = latency_switch_event,
1191         .runtime_event          = latency_runtime_event,
1192         .fork_event             = latency_fork_event,
1193 };
1194
1195 static void output_lat_thread(struct work_atoms *work_list)
1196 {
1197         int i;
1198         int ret;
1199         u64 avg;
1200
1201         if (!work_list->nb_atoms)
1202                 return;
1203         /*
1204          * Ignore idle threads:
1205          */
1206         if (!work_list->thread->pid)
1207                 return;
1208
1209         all_runtime += work_list->total_runtime;
1210         all_count += work_list->nb_atoms;
1211
1212         ret = printf("  %s-%d ", work_list->thread->comm, work_list->thread->pid);
1213
1214         for (i = 0; i < 24 - ret; i++)
1215                 printf(" ");
1216
1217         avg = work_list->total_lat / work_list->nb_atoms;
1218
1219         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1220               (double)work_list->total_runtime / 1e6,
1221                  work_list->nb_atoms, (double)avg / 1e6,
1222                  (double)work_list->max_lat / 1e6);
1223 }
1224
1225 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1226 {
1227         if (l->thread->pid < r->thread->pid)
1228                 return -1;
1229         if (l->thread->pid > r->thread->pid)
1230                 return 1;
1231
1232         return 0;
1233 }
1234
1235 static struct sort_dimension pid_sort_dimension = {
1236         .name                   = "pid",
1237         .cmp                    = pid_cmp,
1238 };
1239
1240 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1241 {
1242         u64 avgl, avgr;
1243
1244         if (!l->nb_atoms)
1245                 return -1;
1246
1247         if (!r->nb_atoms)
1248                 return 1;
1249
1250         avgl = l->total_lat / l->nb_atoms;
1251         avgr = r->total_lat / r->nb_atoms;
1252
1253         if (avgl < avgr)
1254                 return -1;
1255         if (avgl > avgr)
1256                 return 1;
1257
1258         return 0;
1259 }
1260
1261 static struct sort_dimension avg_sort_dimension = {
1262         .name                   = "avg",
1263         .cmp                    = avg_cmp,
1264 };
1265
1266 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1267 {
1268         if (l->max_lat < r->max_lat)
1269                 return -1;
1270         if (l->max_lat > r->max_lat)
1271                 return 1;
1272
1273         return 0;
1274 }
1275
1276 static struct sort_dimension max_sort_dimension = {
1277         .name                   = "max",
1278         .cmp                    = max_cmp,
1279 };
1280
1281 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1282 {
1283         if (l->nb_atoms < r->nb_atoms)
1284                 return -1;
1285         if (l->nb_atoms > r->nb_atoms)
1286                 return 1;
1287
1288         return 0;
1289 }
1290
1291 static struct sort_dimension switch_sort_dimension = {
1292         .name                   = "switch",
1293         .cmp                    = switch_cmp,
1294 };
1295
1296 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1297 {
1298         if (l->total_runtime < r->total_runtime)
1299                 return -1;
1300         if (l->total_runtime > r->total_runtime)
1301                 return 1;
1302
1303         return 0;
1304 }
1305
1306 static struct sort_dimension runtime_sort_dimension = {
1307         .name                   = "runtime",
1308         .cmp                    = runtime_cmp,
1309 };
1310
1311 static struct sort_dimension *available_sorts[] = {
1312         &pid_sort_dimension,
1313         &avg_sort_dimension,
1314         &max_sort_dimension,
1315         &switch_sort_dimension,
1316         &runtime_sort_dimension,
1317 };
1318
1319 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1320
1321 static LIST_HEAD(sort_list);
1322
1323 static int sort_dimension__add(char *tok, struct list_head *list)
1324 {
1325         int i;
1326
1327         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1328                 if (!strcmp(available_sorts[i]->name, tok)) {
1329                         list_add_tail(&available_sorts[i]->list, list);
1330
1331                         return 0;
1332                 }
1333         }
1334
1335         return -1;
1336 }
1337
1338 static void setup_sorting(void);
1339
1340 static void sort_lat(void)
1341 {
1342         struct rb_node *node;
1343
1344         for (;;) {
1345                 struct work_atoms *data;
1346                 node = rb_first(&atom_root);
1347                 if (!node)
1348                         break;
1349
1350                 rb_erase(node, &atom_root);
1351                 data = rb_entry(node, struct work_atoms, node);
1352                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1353         }
1354 }
1355
1356 static void __cmd_lat(void)
1357 {
1358         struct rb_node *next;
1359
1360         setup_pager();
1361         read_events();
1362         sort_lat();
1363
1364         printf("\n -----------------------------------------------------------------------------------------\n");
1365         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1366         printf(" -----------------------------------------------------------------------------------------\n");
1367
1368         next = rb_first(&sorted_atom_root);
1369
1370         while (next) {
1371                 struct work_atoms *work_list;
1372
1373                 work_list = rb_entry(next, struct work_atoms, node);
1374                 output_lat_thread(work_list);
1375                 next = rb_next(next);
1376         }
1377
1378         printf(" -----------------------------------------------------------------------------------------\n");
1379         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1380                 (double)all_runtime/1e6, all_count);
1381
1382         printf(" ---------------------------------------------------\n");
1383         if (nr_unordered_timestamps && nr_timestamps) {
1384                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1385                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1386                         nr_unordered_timestamps, nr_timestamps);
1387         } else {
1388         }
1389         if (nr_lost_events && nr_events) {
1390                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1391                         (double)nr_lost_events/(double)nr_events*100.0,
1392                         nr_lost_events, nr_events, nr_lost_chunks);
1393         }
1394         if (nr_state_machine_bugs && nr_timestamps) {
1395                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1396                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1397                         nr_state_machine_bugs, nr_timestamps);
1398                 if (nr_lost_events)
1399                         printf(" (due to lost events?)");
1400                 printf("\n");
1401         }
1402         printf("\n");
1403
1404 }
1405
1406 static struct trace_sched_handler *trace_handler;
1407
1408 static void
1409 process_sched_wakeup_event(struct raw_event_sample *raw,
1410                            struct event *event,
1411                            int cpu __used,
1412                            u64 timestamp __used,
1413                            struct thread *thread __used)
1414 {
1415         struct trace_wakeup_event wakeup_event;
1416
1417         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1418
1419         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1420         FILL_FIELD(wakeup_event, pid, event, raw->data);
1421         FILL_FIELD(wakeup_event, prio, event, raw->data);
1422         FILL_FIELD(wakeup_event, success, event, raw->data);
1423         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1424
1425         trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1426 }
1427
1428 static void
1429 process_sched_switch_event(struct raw_event_sample *raw,
1430                            struct event *event,
1431                            int cpu __used,
1432                            u64 timestamp __used,
1433                            struct thread *thread __used)
1434 {
1435         struct trace_switch_event switch_event;
1436
1437         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1438
1439         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1440         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1441         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1442         FILL_FIELD(switch_event, prev_state, event, raw->data);
1443         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1444         FILL_FIELD(switch_event, next_pid, event, raw->data);
1445         FILL_FIELD(switch_event, next_prio, event, raw->data);
1446
1447         trace_handler->switch_event(&switch_event, event, cpu, timestamp, thread);
1448 }
1449
1450 static void
1451 process_sched_runtime_event(struct raw_event_sample *raw,
1452                            struct event *event,
1453                            int cpu __used,
1454                            u64 timestamp __used,
1455                            struct thread *thread __used)
1456 {
1457         struct trace_runtime_event runtime_event;
1458
1459         FILL_ARRAY(runtime_event, comm, event, raw->data);
1460         FILL_FIELD(runtime_event, pid, event, raw->data);
1461         FILL_FIELD(runtime_event, runtime, event, raw->data);
1462         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1463
1464         trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1465 }
1466
1467 static void
1468 process_sched_fork_event(struct raw_event_sample *raw,
1469                          struct event *event,
1470                          int cpu __used,
1471                          u64 timestamp __used,
1472                          struct thread *thread __used)
1473 {
1474         struct trace_fork_event fork_event;
1475
1476         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1477
1478         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1479         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1480         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1481         FILL_FIELD(fork_event, child_pid, event, raw->data);
1482
1483         trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1484 }
1485
1486 static void
1487 process_sched_exit_event(struct event *event,
1488                          int cpu __used,
1489                          u64 timestamp __used,
1490                          struct thread *thread __used)
1491 {
1492         if (verbose)
1493                 printf("sched_exit event %p\n", event);
1494 }
1495
1496 static void
1497 process_raw_event(event_t *raw_event __used, void *more_data,
1498                   int cpu, u64 timestamp, struct thread *thread)
1499 {
1500         struct raw_event_sample *raw = more_data;
1501         struct event *event;
1502         int type;
1503
1504         type = trace_parse_common_type(raw->data);
1505         event = trace_find_event(type);
1506
1507         if (!strcmp(event->name, "sched_switch"))
1508                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1509         if (!strcmp(event->name, "sched_stat_runtime"))
1510                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1511         if (!strcmp(event->name, "sched_wakeup"))
1512                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1513         if (!strcmp(event->name, "sched_wakeup_new"))
1514                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1515         if (!strcmp(event->name, "sched_process_fork"))
1516                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1517         if (!strcmp(event->name, "sched_process_exit"))
1518                 process_sched_exit_event(event, cpu, timestamp, thread);
1519 }
1520
1521 static int
1522 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1523 {
1524         char level;
1525         int show = 0;
1526         struct dso *dso = NULL;
1527         struct thread *thread;
1528         u64 ip = event->ip.ip;
1529         u64 timestamp = -1;
1530         u32 cpu = -1;
1531         u64 period = 1;
1532         void *more_data = event->ip.__more_data;
1533         int cpumode;
1534
1535         thread = threads__findnew(event->ip.pid, &threads, &last_match);
1536
1537         if (sample_type & PERF_SAMPLE_TIME) {
1538                 timestamp = *(u64 *)more_data;
1539                 more_data += sizeof(u64);
1540         }
1541
1542         if (sample_type & PERF_SAMPLE_CPU) {
1543                 cpu = *(u32 *)more_data;
1544                 more_data += sizeof(u32);
1545                 more_data += sizeof(u32); /* reserved */
1546         }
1547
1548         if (sample_type & PERF_SAMPLE_PERIOD) {
1549                 period = *(u64 *)more_data;
1550                 more_data += sizeof(u64);
1551         }
1552
1553         dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1554                 (void *)(offset + head),
1555                 (void *)(long)(event->header.size),
1556                 event->header.misc,
1557                 event->ip.pid, event->ip.tid,
1558                 (void *)(long)ip,
1559                 (long long)period);
1560
1561         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1562
1563         if (thread == NULL) {
1564                 eprintf("problem processing %d event, skipping it.\n",
1565                         event->header.type);
1566                 return -1;
1567         }
1568
1569         cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
1570
1571         if (cpumode == PERF_EVENT_MISC_KERNEL) {
1572                 show = SHOW_KERNEL;
1573                 level = 'k';
1574
1575                 dso = kernel_dso;
1576
1577                 dump_printf(" ...... dso: %s\n", dso->name);
1578
1579         } else if (cpumode == PERF_EVENT_MISC_USER) {
1580
1581                 show = SHOW_USER;
1582                 level = '.';
1583
1584         } else {
1585                 show = SHOW_HV;
1586                 level = 'H';
1587
1588                 dso = hypervisor_dso;
1589
1590                 dump_printf(" ...... dso: [hypervisor]\n");
1591         }
1592
1593         if (sample_type & PERF_SAMPLE_RAW)
1594                 process_raw_event(event, more_data, cpu, timestamp, thread);
1595
1596         return 0;
1597 }
1598
1599 static int
1600 process_event(event_t *event, unsigned long offset, unsigned long head)
1601 {
1602         trace_event(event);
1603
1604         nr_events++;
1605         switch (event->header.type) {
1606         case PERF_EVENT_MMAP:
1607                 return 0;
1608         case PERF_EVENT_LOST:
1609                 nr_lost_chunks++;
1610                 nr_lost_events += event->lost.lost;
1611                 return 0;
1612
1613         case PERF_EVENT_COMM:
1614                 return process_comm_event(event, offset, head);
1615
1616         case PERF_EVENT_EXIT ... PERF_EVENT_READ:
1617                 return 0;
1618
1619         case PERF_EVENT_SAMPLE:
1620                 return process_sample_event(event, offset, head);
1621
1622         case PERF_EVENT_MAX:
1623         default:
1624                 return -1;
1625         }
1626
1627         return 0;
1628 }
1629
1630 static int read_events(void)
1631 {
1632         int ret, rc = EXIT_FAILURE;
1633         unsigned long offset = 0;
1634         unsigned long head = 0;
1635         struct stat perf_stat;
1636         event_t *event;
1637         uint32_t size;
1638         char *buf;
1639
1640         trace_report();
1641         register_idle_thread(&threads, &last_match);
1642
1643         input = open(input_name, O_RDONLY);
1644         if (input < 0) {
1645                 perror("failed to open file");
1646                 exit(-1);
1647         }
1648
1649         ret = fstat(input, &perf_stat);
1650         if (ret < 0) {
1651                 perror("failed to stat file");
1652                 exit(-1);
1653         }
1654
1655         if (!perf_stat.st_size) {
1656                 fprintf(stderr, "zero-sized file, nothing to do!\n");
1657                 exit(0);
1658         }
1659         header = perf_header__read(input);
1660         head = header->data_offset;
1661         sample_type = perf_header__sample_type(header);
1662
1663         if (!(sample_type & PERF_SAMPLE_RAW))
1664                 die("No trace sample to read. Did you call perf record "
1665                     "without -R?");
1666
1667         if (load_kernel() < 0) {
1668                 perror("failed to load kernel symbols");
1669                 return EXIT_FAILURE;
1670         }
1671
1672 remap:
1673         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
1674                            MAP_SHARED, input, offset);
1675         if (buf == MAP_FAILED) {
1676                 perror("failed to mmap file");
1677                 exit(-1);
1678         }
1679
1680 more:
1681         event = (event_t *)(buf + head);
1682
1683         size = event->header.size;
1684         if (!size)
1685                 size = 8;
1686
1687         if (head + event->header.size >= page_size * mmap_window) {
1688                 unsigned long shift = page_size * (head / page_size);
1689                 int res;
1690
1691                 res = munmap(buf, page_size * mmap_window);
1692                 assert(res == 0);
1693
1694                 offset += shift;
1695                 head -= shift;
1696                 goto remap;
1697         }
1698
1699         size = event->header.size;
1700
1701
1702         if (!size || process_event(event, offset, head) < 0) {
1703
1704                 /*
1705                  * assume we lost track of the stream, check alignment, and
1706                  * increment a single u64 in the hope to catch on again 'soon'.
1707                  */
1708
1709                 if (unlikely(head & 7))
1710                         head &= ~7ULL;
1711
1712                 size = 8;
1713         }
1714
1715         head += size;
1716
1717         if (offset + head < (unsigned long)perf_stat.st_size)
1718                 goto more;
1719
1720         rc = EXIT_SUCCESS;
1721         close(input);
1722
1723         return rc;
1724 }
1725
1726 static const char * const sched_usage[] = {
1727         "perf sched [<options>] {record|latency|replay|trace}",
1728         NULL
1729 };
1730
1731 static const struct option sched_options[] = {
1732         OPT_BOOLEAN('v', "verbose", &verbose,
1733                     "be more verbose (show symbol address, etc)"),
1734         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1735                     "dump raw trace in ASCII"),
1736         OPT_END()
1737 };
1738
1739 static const char * const latency_usage[] = {
1740         "perf sched latency [<options>]",
1741         NULL
1742 };
1743
1744 static const struct option latency_options[] = {
1745         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1746                    "sort by key(s): runtime, switch, avg, max"),
1747         OPT_BOOLEAN('v', "verbose", &verbose,
1748                     "be more verbose (show symbol address, etc)"),
1749         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1750                     "dump raw trace in ASCII"),
1751         OPT_END()
1752 };
1753
1754 static const char * const replay_usage[] = {
1755         "perf sched replay [<options>]",
1756         NULL
1757 };
1758
1759 static const struct option replay_options[] = {
1760         OPT_INTEGER('r', "repeat", &replay_repeat,
1761                     "repeat the workload replay N times (-1: infinite)"),
1762         OPT_BOOLEAN('v', "verbose", &verbose,
1763                     "be more verbose (show symbol address, etc)"),
1764         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1765                     "dump raw trace in ASCII"),
1766         OPT_END()
1767 };
1768
1769 static void setup_sorting(void)
1770 {
1771         char *tmp, *tok, *str = strdup(sort_order);
1772
1773         for (tok = strtok_r(str, ", ", &tmp);
1774                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1775                 if (sort_dimension__add(tok, &sort_list) < 0) {
1776                         error("Unknown --sort key: `%s'", tok);
1777                         usage_with_options(latency_usage, latency_options);
1778                 }
1779         }
1780
1781         free(str);
1782
1783         sort_dimension__add((char *)"pid", &cmp_pid);
1784 }
1785
1786 static const char *record_args[] = {
1787         "record",
1788         "-a",
1789         "-R",
1790         "-M",
1791         "-f",
1792         "-m", "1024",
1793         "-c", "1",
1794         "-e", "sched:sched_switch:r",
1795         "-e", "sched:sched_stat_wait:r",
1796         "-e", "sched:sched_stat_sleep:r",
1797         "-e", "sched:sched_stat_iowait:r",
1798         "-e", "sched:sched_stat_runtime:r",
1799         "-e", "sched:sched_process_exit:r",
1800         "-e", "sched:sched_process_fork:r",
1801         "-e", "sched:sched_wakeup:r",
1802         "-e", "sched:sched_migrate_task:r",
1803 };
1804
1805 static int __cmd_record(int argc, const char **argv)
1806 {
1807         unsigned int rec_argc, i, j;
1808         const char **rec_argv;
1809
1810         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1811         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1812
1813         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1814                 rec_argv[i] = strdup(record_args[i]);
1815
1816         for (j = 1; j < (unsigned int)argc; j++, i++)
1817                 rec_argv[i] = argv[j];
1818
1819         BUG_ON(i != rec_argc);
1820
1821         return cmd_record(i, rec_argv, NULL);
1822 }
1823
1824 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1825 {
1826         symbol__init();
1827         page_size = getpagesize();
1828
1829         argc = parse_options(argc, argv, sched_options, sched_usage,
1830                              PARSE_OPT_STOP_AT_NON_OPTION);
1831         if (!argc)
1832                 usage_with_options(sched_usage, sched_options);
1833
1834         if (!strncmp(argv[0], "rec", 3)) {
1835                 return __cmd_record(argc, argv);
1836         } else if (!strncmp(argv[0], "lat", 3)) {
1837                 trace_handler = &lat_ops;
1838                 if (argc > 1) {
1839                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1840                         if (argc)
1841                                 usage_with_options(latency_usage, latency_options);
1842                 }
1843                 setup_sorting();
1844                 __cmd_lat();
1845         } else if (!strncmp(argv[0], "rep", 3)) {
1846                 trace_handler = &replay_ops;
1847                 if (argc) {
1848                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1849                         if (argc)
1850                                 usage_with_options(replay_usage, replay_options);
1851                 }
1852                 __cmd_replay();
1853         } else if (!strcmp(argv[0], "trace")) {
1854                 /*
1855                  * Aliased to 'perf trace' for now:
1856                  */
1857                 return cmd_trace(argc, argv, prefix);
1858         } else {
1859                 usage_with_options(sched_usage, sched_options);
1860         }
1861
1862         return 0;
1863 }