perf tools: Improve thread comm resolution in perf sched
[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 struct thread *
1038 threads__findnew_from_ctx(u32 pid, struct trace_switch_event *switch_event)
1039 {
1040         struct thread *th;
1041
1042         th = threads__findnew_nocomm(pid, &threads, &last_match);
1043         if (th->comm)
1044                 return th;
1045
1046         if (pid == switch_event->prev_pid)
1047                 thread__set_comm(th, switch_event->prev_comm);
1048         else
1049                 thread__set_comm(th, switch_event->next_comm);
1050         return th;
1051 }
1052
1053 static struct thread *
1054 threads__findnew_from_wakeup(struct trace_wakeup_event *wakeup_event)
1055 {
1056         struct thread *th;
1057
1058         th =  threads__findnew_nocomm(wakeup_event->pid, &threads, &last_match);
1059         if (th->comm)
1060                 return th;
1061
1062         thread__set_comm(th, wakeup_event->comm);
1063
1064         return th;
1065 }
1066
1067 static void
1068 latency_switch_event(struct trace_switch_event *switch_event,
1069                      struct event *event __used,
1070                      int cpu,
1071                      u64 timestamp,
1072                      struct thread *thread __used)
1073 {
1074         struct work_atoms *out_events, *in_events;
1075         struct thread *sched_out, *sched_in;
1076         u64 timestamp0;
1077         s64 delta;
1078
1079         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1080
1081         timestamp0 = cpu_last_switched[cpu];
1082         cpu_last_switched[cpu] = timestamp;
1083         if (timestamp0)
1084                 delta = timestamp - timestamp0;
1085         else
1086                 delta = 0;
1087
1088         if (delta < 0)
1089                 die("hm, delta: %Ld < 0 ?\n", delta);
1090
1091
1092         sched_out = threads__findnew_from_ctx(switch_event->prev_pid,
1093                                               switch_event);
1094         sched_in = threads__findnew_from_ctx(switch_event->next_pid,
1095                                              switch_event);
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_from_wakeup(wakeup_event);
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 (!strcmp(work_list->thread->comm, "swapper"))
1207                 return;
1208
1209         all_runtime += work_list->total_runtime;
1210         all_count += work_list->nb_atoms;
1211
1212         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1213
1214         for (i = 0; i < 24 - ret; i++)
1215                 printf(" ");
1216
1217         avg = work_list->total_lat / work_list->nb_atoms;
1218
1219         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1220               (double)work_list->total_runtime / 1e6,
1221                  work_list->nb_atoms, (double)avg / 1e6,
1222                  (double)work_list->max_lat / 1e6);
1223 }
1224
1225 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1226 {
1227         if (l->thread->pid < r->thread->pid)
1228                 return -1;
1229         if (l->thread->pid > r->thread->pid)
1230                 return 1;
1231
1232         return 0;
1233 }
1234
1235 static struct sort_dimension pid_sort_dimension = {
1236         .name                   = "pid",
1237         .cmp                    = pid_cmp,
1238 };
1239
1240 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1241 {
1242         u64 avgl, avgr;
1243
1244         if (!l->nb_atoms)
1245                 return -1;
1246
1247         if (!r->nb_atoms)
1248                 return 1;
1249
1250         avgl = l->total_lat / l->nb_atoms;
1251         avgr = r->total_lat / r->nb_atoms;
1252
1253         if (avgl < avgr)
1254                 return -1;
1255         if (avgl > avgr)
1256                 return 1;
1257
1258         return 0;
1259 }
1260
1261 static struct sort_dimension avg_sort_dimension = {
1262         .name                   = "avg",
1263         .cmp                    = avg_cmp,
1264 };
1265
1266 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1267 {
1268         if (l->max_lat < r->max_lat)
1269                 return -1;
1270         if (l->max_lat > r->max_lat)
1271                 return 1;
1272
1273         return 0;
1274 }
1275
1276 static struct sort_dimension max_sort_dimension = {
1277         .name                   = "max",
1278         .cmp                    = max_cmp,
1279 };
1280
1281 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1282 {
1283         if (l->nb_atoms < r->nb_atoms)
1284                 return -1;
1285         if (l->nb_atoms > r->nb_atoms)
1286                 return 1;
1287
1288         return 0;
1289 }
1290
1291 static struct sort_dimension switch_sort_dimension = {
1292         .name                   = "switch",
1293         .cmp                    = switch_cmp,
1294 };
1295
1296 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1297 {
1298         if (l->total_runtime < r->total_runtime)
1299                 return -1;
1300         if (l->total_runtime > r->total_runtime)
1301                 return 1;
1302
1303         return 0;
1304 }
1305
1306 static struct sort_dimension runtime_sort_dimension = {
1307         .name                   = "runtime",
1308         .cmp                    = runtime_cmp,
1309 };
1310
1311 static struct sort_dimension *available_sorts[] = {
1312         &pid_sort_dimension,
1313         &avg_sort_dimension,
1314         &max_sort_dimension,
1315         &switch_sort_dimension,
1316         &runtime_sort_dimension,
1317 };
1318
1319 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1320
1321 static LIST_HEAD(sort_list);
1322
1323 static int sort_dimension__add(char *tok, struct list_head *list)
1324 {
1325         int i;
1326
1327         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1328                 if (!strcmp(available_sorts[i]->name, tok)) {
1329                         list_add_tail(&available_sorts[i]->list, list);
1330
1331                         return 0;
1332                 }
1333         }
1334
1335         return -1;
1336 }
1337
1338 static void setup_sorting(void);
1339
1340 static void sort_lat(void)
1341 {
1342         struct rb_node *node;
1343
1344         for (;;) {
1345                 struct work_atoms *data;
1346                 node = rb_first(&atom_root);
1347                 if (!node)
1348                         break;
1349
1350                 rb_erase(node, &atom_root);
1351                 data = rb_entry(node, struct work_atoms, node);
1352                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1353         }
1354 }
1355
1356 static struct trace_sched_handler *trace_handler;
1357
1358 static void
1359 process_sched_wakeup_event(struct raw_event_sample *raw,
1360                            struct event *event,
1361                            int cpu __used,
1362                            u64 timestamp __used,
1363                            struct thread *thread __used)
1364 {
1365         struct trace_wakeup_event wakeup_event;
1366
1367         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1368
1369         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1370         FILL_FIELD(wakeup_event, pid, event, raw->data);
1371         FILL_FIELD(wakeup_event, prio, event, raw->data);
1372         FILL_FIELD(wakeup_event, success, event, raw->data);
1373         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1374
1375         if (trace_handler->wakeup_event)
1376                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1377 }
1378
1379 /*
1380  * Track the current task - that way we can know whether there's any
1381  * weird events, such as a task being switched away that is not current.
1382  */
1383 static int max_cpu;
1384
1385 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1386
1387 static struct thread *curr_thread[MAX_CPUS];
1388
1389 static char next_shortname1 = 'A';
1390 static char next_shortname2 = '0';
1391
1392 static void
1393 map_switch_event(struct trace_switch_event *switch_event,
1394                  struct event *event __used,
1395                  int this_cpu,
1396                  u64 timestamp,
1397                  struct thread *thread __used)
1398 {
1399         struct thread *sched_out, *sched_in;
1400         int new_shortname;
1401         u64 timestamp0;
1402         s64 delta;
1403         int cpu;
1404
1405         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1406
1407         if (this_cpu > max_cpu)
1408                 max_cpu = this_cpu;
1409
1410         timestamp0 = cpu_last_switched[this_cpu];
1411         cpu_last_switched[this_cpu] = timestamp;
1412         if (timestamp0)
1413                 delta = timestamp - timestamp0;
1414         else
1415                 delta = 0;
1416
1417         if (delta < 0)
1418                 die("hm, delta: %Ld < 0 ?\n", delta);
1419
1420
1421         sched_out = threads__findnew_from_ctx(switch_event->prev_pid,
1422                                               switch_event);
1423         sched_in = threads__findnew_from_ctx(switch_event->next_pid,
1424                                              switch_event);
1425
1426         curr_thread[this_cpu] = sched_in;
1427
1428         printf("  ");
1429
1430         new_shortname = 0;
1431         if (!sched_in->shortname[0]) {
1432                 sched_in->shortname[0] = next_shortname1;
1433                 sched_in->shortname[1] = next_shortname2;
1434
1435                 if (next_shortname1 < 'Z') {
1436                         next_shortname1++;
1437                 } else {
1438                         next_shortname1='A';
1439                         if (next_shortname2 < '9') {
1440                                 next_shortname2++;
1441                         } else {
1442                                 next_shortname2='0';
1443                         }
1444                 }
1445                 new_shortname = 1;
1446         }
1447
1448         for (cpu = 0; cpu <= max_cpu; cpu++) {
1449                 if (cpu != this_cpu)
1450                         printf(" ");
1451                 else
1452                         printf("*");
1453
1454                 if (curr_thread[cpu]) {
1455                         if (curr_thread[cpu]->pid)
1456                                 printf("%2s ", curr_thread[cpu]->shortname);
1457                         else
1458                                 printf(".  ");
1459                 } else
1460                         printf("   ");
1461         }
1462
1463         printf("  %12.6f secs ", (double)timestamp/1e9);
1464         if (new_shortname) {
1465                 printf("%s => %s:%d\n",
1466                         sched_in->shortname, sched_in->comm, sched_in->pid);
1467         } else {
1468                 printf("\n");
1469         }
1470 }
1471
1472
1473 static void
1474 process_sched_switch_event(struct raw_event_sample *raw,
1475                            struct event *event,
1476                            int this_cpu,
1477                            u64 timestamp __used,
1478                            struct thread *thread __used)
1479 {
1480         struct trace_switch_event switch_event;
1481
1482         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1483
1484         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1485         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1486         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1487         FILL_FIELD(switch_event, prev_state, event, raw->data);
1488         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1489         FILL_FIELD(switch_event, next_pid, event, raw->data);
1490         FILL_FIELD(switch_event, next_prio, event, raw->data);
1491
1492         if (curr_pid[this_cpu] != (u32)-1) {
1493                 /*
1494                  * Are we trying to switch away a PID that is
1495                  * not current?
1496                  */
1497                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1498                         nr_context_switch_bugs++;
1499         }
1500         if (trace_handler->switch_event)
1501                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1502
1503         curr_pid[this_cpu] = switch_event.next_pid;
1504 }
1505
1506 static void
1507 process_sched_runtime_event(struct raw_event_sample *raw,
1508                            struct event *event,
1509                            int cpu __used,
1510                            u64 timestamp __used,
1511                            struct thread *thread __used)
1512 {
1513         struct trace_runtime_event runtime_event;
1514
1515         FILL_ARRAY(runtime_event, comm, event, raw->data);
1516         FILL_FIELD(runtime_event, pid, event, raw->data);
1517         FILL_FIELD(runtime_event, runtime, event, raw->data);
1518         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1519
1520         if (trace_handler->runtime_event)
1521                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1522 }
1523
1524 static void
1525 process_sched_fork_event(struct raw_event_sample *raw,
1526                          struct event *event,
1527                          int cpu __used,
1528                          u64 timestamp __used,
1529                          struct thread *thread __used)
1530 {
1531         struct trace_fork_event fork_event;
1532
1533         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1534
1535         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1536         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1537         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1538         FILL_FIELD(fork_event, child_pid, event, raw->data);
1539
1540         if (trace_handler->fork_event)
1541                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1542 }
1543
1544 static void
1545 process_sched_exit_event(struct event *event,
1546                          int cpu __used,
1547                          u64 timestamp __used,
1548                          struct thread *thread __used)
1549 {
1550         if (verbose)
1551                 printf("sched_exit event %p\n", event);
1552 }
1553
1554 static void
1555 process_raw_event(event_t *raw_event __used, void *more_data,
1556                   int cpu, u64 timestamp, struct thread *thread)
1557 {
1558         struct raw_event_sample *raw = more_data;
1559         struct event *event;
1560         int type;
1561
1562         type = trace_parse_common_type(raw->data);
1563         event = trace_find_event(type);
1564
1565         if (!strcmp(event->name, "sched_switch"))
1566                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1567         if (!strcmp(event->name, "sched_stat_runtime"))
1568                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1569         if (!strcmp(event->name, "sched_wakeup"))
1570                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1571         if (!strcmp(event->name, "sched_wakeup_new"))
1572                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1573         if (!strcmp(event->name, "sched_process_fork"))
1574                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1575         if (!strcmp(event->name, "sched_process_exit"))
1576                 process_sched_exit_event(event, cpu, timestamp, thread);
1577 }
1578
1579 static int
1580 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1581 {
1582         struct thread *thread;
1583         u64 ip = event->ip.ip;
1584         u64 timestamp = -1;
1585         u32 cpu = -1;
1586         u64 period = 1;
1587         void *more_data = event->ip.__more_data;
1588
1589         if (!(sample_type & PERF_SAMPLE_RAW))
1590                 return 0;
1591
1592         thread = threads__findnew(event->ip.pid, &threads, &last_match);
1593
1594         if (sample_type & PERF_SAMPLE_TIME) {
1595                 timestamp = *(u64 *)more_data;
1596                 more_data += sizeof(u64);
1597         }
1598
1599         if (sample_type & PERF_SAMPLE_CPU) {
1600                 cpu = *(u32 *)more_data;
1601                 more_data += sizeof(u32);
1602                 more_data += sizeof(u32); /* reserved */
1603         }
1604
1605         if (sample_type & PERF_SAMPLE_PERIOD) {
1606                 period = *(u64 *)more_data;
1607                 more_data += sizeof(u64);
1608         }
1609
1610         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1611                 (void *)(offset + head),
1612                 (void *)(long)(event->header.size),
1613                 event->header.misc,
1614                 event->ip.pid, event->ip.tid,
1615                 (void *)(long)ip,
1616                 (long long)period);
1617
1618         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1619
1620         if (thread == NULL) {
1621                 eprintf("problem processing %d event, skipping it.\n",
1622                         event->header.type);
1623                 return -1;
1624         }
1625
1626         process_raw_event(event, more_data, cpu, timestamp, thread);
1627
1628         return 0;
1629 }
1630
1631 static int
1632 process_lost_event(event_t *event __used,
1633                    unsigned long offset __used,
1634                    unsigned long head __used)
1635 {
1636         nr_lost_chunks++;
1637         nr_lost_events += event->lost.lost;
1638
1639         return 0;
1640 }
1641
1642 static int sample_type_check(u64 type)
1643 {
1644         sample_type = type;
1645
1646         if (!(sample_type & PERF_SAMPLE_RAW)) {
1647                 fprintf(stderr,
1648                         "No trace sample to read. Did you call perf record "
1649                         "without -R?");
1650                 return -1;
1651         }
1652
1653         return 0;
1654 }
1655
1656 static struct perf_file_handler file_handler = {
1657         .process_sample_event   = process_sample_event,
1658         .process_comm_event     = process_comm_event,
1659         .process_lost_event     = process_lost_event,
1660         .sample_type_check      = sample_type_check,
1661 };
1662
1663 static int read_events(void)
1664 {
1665         register_idle_thread(&threads, &last_match);
1666         register_perf_file_handler(&file_handler);
1667
1668         return mmap_dispatch_perf_file(&header, input_name, 0, 0, &cwdlen, &cwd);
1669 }
1670
1671 static void print_bad_events(void)
1672 {
1673         if (nr_unordered_timestamps && nr_timestamps) {
1674                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1675                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1676                         nr_unordered_timestamps, nr_timestamps);
1677         }
1678         if (nr_lost_events && nr_events) {
1679                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1680                         (double)nr_lost_events/(double)nr_events*100.0,
1681                         nr_lost_events, nr_events, nr_lost_chunks);
1682         }
1683         if (nr_state_machine_bugs && nr_timestamps) {
1684                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1685                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1686                         nr_state_machine_bugs, nr_timestamps);
1687                 if (nr_lost_events)
1688                         printf(" (due to lost events?)");
1689                 printf("\n");
1690         }
1691         if (nr_context_switch_bugs && nr_timestamps) {
1692                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1693                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1694                         nr_context_switch_bugs, nr_timestamps);
1695                 if (nr_lost_events)
1696                         printf(" (due to lost events?)");
1697                 printf("\n");
1698         }
1699 }
1700
1701 static void __cmd_lat(void)
1702 {
1703         struct rb_node *next;
1704
1705         setup_pager();
1706         read_events();
1707         sort_lat();
1708
1709         printf("\n -----------------------------------------------------------------------------------------\n");
1710         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1711         printf(" -----------------------------------------------------------------------------------------\n");
1712
1713         next = rb_first(&sorted_atom_root);
1714
1715         while (next) {
1716                 struct work_atoms *work_list;
1717
1718                 work_list = rb_entry(next, struct work_atoms, node);
1719                 output_lat_thread(work_list);
1720                 next = rb_next(next);
1721         }
1722
1723         printf(" -----------------------------------------------------------------------------------------\n");
1724         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1725                 (double)all_runtime/1e6, all_count);
1726
1727         printf(" ---------------------------------------------------\n");
1728
1729         print_bad_events();
1730         printf("\n");
1731
1732 }
1733
1734 static struct trace_sched_handler map_ops  = {
1735         .wakeup_event           = NULL,
1736         .switch_event           = map_switch_event,
1737         .runtime_event          = NULL,
1738         .fork_event             = NULL,
1739 };
1740
1741 static void __cmd_map(void)
1742 {
1743         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1744
1745         setup_pager();
1746         read_events();
1747         print_bad_events();
1748 }
1749
1750 static void __cmd_replay(void)
1751 {
1752         unsigned long i;
1753
1754         calibrate_run_measurement_overhead();
1755         calibrate_sleep_measurement_overhead();
1756
1757         test_calibrations();
1758
1759         read_events();
1760
1761         printf("nr_run_events:        %ld\n", nr_run_events);
1762         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1763         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1764
1765         if (targetless_wakeups)
1766                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1767         if (multitarget_wakeups)
1768                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1769         if (nr_run_events_optimized)
1770                 printf("run atoms optimized: %ld\n",
1771                         nr_run_events_optimized);
1772
1773         print_task_traces();
1774         add_cross_task_wakeups();
1775
1776         create_tasks();
1777         printf("------------------------------------------------------------\n");
1778         for (i = 0; i < replay_repeat; i++)
1779                 run_one_test();
1780 }
1781
1782
1783 static const char * const sched_usage[] = {
1784         "perf sched [<options>] {record|latency|map|replay|trace}",
1785         NULL
1786 };
1787
1788 static const struct option sched_options[] = {
1789         OPT_STRING('i', "input", &input_name, "file",
1790                     "input file name"),
1791         OPT_BOOLEAN('v', "verbose", &verbose,
1792                     "be more verbose (show symbol address, etc)"),
1793         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1794                     "dump raw trace in ASCII"),
1795         OPT_END()
1796 };
1797
1798 static const char * const latency_usage[] = {
1799         "perf sched latency [<options>]",
1800         NULL
1801 };
1802
1803 static const struct option latency_options[] = {
1804         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1805                    "sort by key(s): runtime, switch, avg, max"),
1806         OPT_BOOLEAN('v', "verbose", &verbose,
1807                     "be more verbose (show symbol address, etc)"),
1808         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1809                     "dump raw trace in ASCII"),
1810         OPT_END()
1811 };
1812
1813 static const char * const replay_usage[] = {
1814         "perf sched replay [<options>]",
1815         NULL
1816 };
1817
1818 static const struct option replay_options[] = {
1819         OPT_INTEGER('r', "repeat", &replay_repeat,
1820                     "repeat the workload replay N times (-1: infinite)"),
1821         OPT_BOOLEAN('v', "verbose", &verbose,
1822                     "be more verbose (show symbol address, etc)"),
1823         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1824                     "dump raw trace in ASCII"),
1825         OPT_END()
1826 };
1827
1828 static void setup_sorting(void)
1829 {
1830         char *tmp, *tok, *str = strdup(sort_order);
1831
1832         for (tok = strtok_r(str, ", ", &tmp);
1833                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1834                 if (sort_dimension__add(tok, &sort_list) < 0) {
1835                         error("Unknown --sort key: `%s'", tok);
1836                         usage_with_options(latency_usage, latency_options);
1837                 }
1838         }
1839
1840         free(str);
1841
1842         sort_dimension__add((char *)"pid", &cmp_pid);
1843 }
1844
1845 static const char *record_args[] = {
1846         "record",
1847         "-a",
1848         "-R",
1849         "-M",
1850         "-f",
1851         "-m", "1024",
1852         "-c", "1",
1853         "-e", "sched:sched_switch:r",
1854         "-e", "sched:sched_stat_wait:r",
1855         "-e", "sched:sched_stat_sleep:r",
1856         "-e", "sched:sched_stat_iowait:r",
1857         "-e", "sched:sched_stat_runtime:r",
1858         "-e", "sched:sched_process_exit:r",
1859         "-e", "sched:sched_process_fork:r",
1860         "-e", "sched:sched_wakeup:r",
1861         "-e", "sched:sched_migrate_task:r",
1862 };
1863
1864 static int __cmd_record(int argc, const char **argv)
1865 {
1866         unsigned int rec_argc, i, j;
1867         const char **rec_argv;
1868
1869         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1870         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1871
1872         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1873                 rec_argv[i] = strdup(record_args[i]);
1874
1875         for (j = 1; j < (unsigned int)argc; j++, i++)
1876                 rec_argv[i] = argv[j];
1877
1878         BUG_ON(i != rec_argc);
1879
1880         return cmd_record(i, rec_argv, NULL);
1881 }
1882
1883 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1884 {
1885         symbol__init();
1886
1887         argc = parse_options(argc, argv, sched_options, sched_usage,
1888                              PARSE_OPT_STOP_AT_NON_OPTION);
1889         if (!argc)
1890                 usage_with_options(sched_usage, sched_options);
1891
1892         if (!strncmp(argv[0], "rec", 3)) {
1893                 return __cmd_record(argc, argv);
1894         } else if (!strncmp(argv[0], "lat", 3)) {
1895                 trace_handler = &lat_ops;
1896                 if (argc > 1) {
1897                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1898                         if (argc)
1899                                 usage_with_options(latency_usage, latency_options);
1900                 }
1901                 setup_sorting();
1902                 __cmd_lat();
1903         } else if (!strcmp(argv[0], "map")) {
1904                 trace_handler = &map_ops;
1905                 setup_sorting();
1906                 __cmd_map();
1907         } else if (!strncmp(argv[0], "rep", 3)) {
1908                 trace_handler = &replay_ops;
1909                 if (argc) {
1910                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1911                         if (argc)
1912                                 usage_with_options(replay_usage, replay_options);
1913                 }
1914                 __cmd_replay();
1915         } else if (!strcmp(argv[0], "trace")) {
1916                 /*
1917                  * Aliased to 'perf trace' for now:
1918                  */
1919                 return cmd_trace(argc, argv, prefix);
1920         } else {
1921                 usage_with_options(sched_usage, sched_options);
1922         }
1923
1924         return 0;
1925 }