dbf089b12defe5be9c11fd627c5d714f438020e3
[linux-2.6.git] / tools / perf / builtin-sched.c
1 #include "builtin.h"
2 #include "perf.h"
3
4 #include "util/util.h"
5 #include "util/cache.h"
6 #include "util/symbol.h"
7 #include "util/thread.h"
8 #include "util/header.h"
9
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 perf_header       *header;
28 static u64                      sample_type;
29
30 static char                     default_sort_order[] = "avg, max, switch, runtime";
31 static char                     *sort_order = default_sort_order;
32
33 static int                      profile_cpu = -1;
34
35 static char                     *cwd;
36 static int                      cwdlen;
37
38 #define PR_SET_NAME             15               /* Set process name */
39 #define MAX_CPUS                4096
40
41 static u64                      run_measurement_overhead;
42 static u64                      sleep_measurement_overhead;
43
44 #define COMM_LEN                20
45 #define SYM_LEN                 129
46
47 #define MAX_PID                 65536
48
49 static unsigned long            nr_tasks;
50
51 struct sched_atom;
52
53 struct task_desc {
54         unsigned long           nr;
55         unsigned long           pid;
56         char                    comm[COMM_LEN];
57
58         unsigned long           nr_events;
59         unsigned long           curr_event;
60         struct sched_atom       **atoms;
61
62         pthread_t               thread;
63         sem_t                   sleep_sem;
64
65         sem_t                   ready_for_work;
66         sem_t                   work_done_sem;
67
68         u64                     cpu_usage;
69 };
70
71 enum sched_event_type {
72         SCHED_EVENT_RUN,
73         SCHED_EVENT_SLEEP,
74         SCHED_EVENT_WAKEUP,
75         SCHED_EVENT_MIGRATION,
76 };
77
78 struct sched_atom {
79         enum sched_event_type   type;
80         u64                     timestamp;
81         u64                     duration;
82         unsigned long           nr;
83         int                     specific_wait;
84         sem_t                   *wait_sem;
85         struct task_desc        *wakee;
86 };
87
88 static struct task_desc         *pid_to_task[MAX_PID];
89
90 static struct task_desc         **tasks;
91
92 static pthread_mutex_t          start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
93 static u64                      start_time;
94
95 static pthread_mutex_t          work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
96
97 static unsigned long            nr_run_events;
98 static unsigned long            nr_sleep_events;
99 static unsigned long            nr_wakeup_events;
100
101 static unsigned long            nr_sleep_corrections;
102 static unsigned long            nr_run_events_optimized;
103
104 static unsigned long            targetless_wakeups;
105 static unsigned long            multitarget_wakeups;
106
107 static u64                      cpu_usage;
108 static u64                      runavg_cpu_usage;
109 static u64                      parent_cpu_usage;
110 static u64                      runavg_parent_cpu_usage;
111
112 static unsigned long            nr_runs;
113 static u64                      sum_runtime;
114 static u64                      sum_fluct;
115 static u64                      run_avg;
116
117 static unsigned long            replay_repeat = 10;
118 static unsigned long            nr_timestamps;
119 static unsigned long            nr_unordered_timestamps;
120 static unsigned long            nr_state_machine_bugs;
121 static unsigned long            nr_context_switch_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
162 static u64 get_nsecs(void)
163 {
164         struct timespec ts;
165
166         clock_gettime(CLOCK_MONOTONIC, &ts);
167
168         return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
169 }
170
171 static void burn_nsecs(u64 nsecs)
172 {
173         u64 T0 = get_nsecs(), T1;
174
175         do {
176                 T1 = get_nsecs();
177         } while (T1 + run_measurement_overhead < T0 + nsecs);
178 }
179
180 static void sleep_nsecs(u64 nsecs)
181 {
182         struct timespec ts;
183
184         ts.tv_nsec = nsecs % 999999999;
185         ts.tv_sec = nsecs / 999999999;
186
187         nanosleep(&ts, NULL);
188 }
189
190 static void calibrate_run_measurement_overhead(void)
191 {
192         u64 T0, T1, delta, min_delta = 1000000000ULL;
193         int i;
194
195         for (i = 0; i < 10; i++) {
196                 T0 = get_nsecs();
197                 burn_nsecs(0);
198                 T1 = get_nsecs();
199                 delta = T1-T0;
200                 min_delta = min(min_delta, delta);
201         }
202         run_measurement_overhead = min_delta;
203
204         printf("run measurement overhead: %Ld nsecs\n", min_delta);
205 }
206
207 static void calibrate_sleep_measurement_overhead(void)
208 {
209         u64 T0, T1, delta, min_delta = 1000000000ULL;
210         int i;
211
212         for (i = 0; i < 10; i++) {
213                 T0 = get_nsecs();
214                 sleep_nsecs(10000);
215                 T1 = get_nsecs();
216                 delta = T1-T0;
217                 min_delta = min(min_delta, delta);
218         }
219         min_delta -= 10000;
220         sleep_measurement_overhead = min_delta;
221
222         printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
223 }
224
225 static struct sched_atom *
226 get_new_event(struct task_desc *task, u64 timestamp)
227 {
228         struct sched_atom *event = calloc(1, sizeof(*event));
229         unsigned long idx = task->nr_events;
230         size_t size;
231
232         event->timestamp = timestamp;
233         event->nr = idx;
234
235         task->nr_events++;
236         size = sizeof(struct sched_atom *) * task->nr_events;
237         task->atoms = realloc(task->atoms, size);
238         BUG_ON(!task->atoms);
239
240         task->atoms[idx] = event;
241
242         return event;
243 }
244
245 static struct sched_atom *last_event(struct task_desc *task)
246 {
247         if (!task->nr_events)
248                 return NULL;
249
250         return task->atoms[task->nr_events - 1];
251 }
252
253 static void
254 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
255 {
256         struct sched_atom *event, *curr_event = last_event(task);
257
258         /*
259          * optimize an existing RUN event by merging this one
260          * to it:
261          */
262         if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
263                 nr_run_events_optimized++;
264                 curr_event->duration += duration;
265                 return;
266         }
267
268         event = get_new_event(task, timestamp);
269
270         event->type = SCHED_EVENT_RUN;
271         event->duration = duration;
272
273         nr_run_events++;
274 }
275
276 static void
277 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
278                        struct task_desc *wakee)
279 {
280         struct sched_atom *event, *wakee_event;
281
282         event = get_new_event(task, timestamp);
283         event->type = SCHED_EVENT_WAKEUP;
284         event->wakee = wakee;
285
286         wakee_event = last_event(wakee);
287         if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
288                 targetless_wakeups++;
289                 return;
290         }
291         if (wakee_event->wait_sem) {
292                 multitarget_wakeups++;
293                 return;
294         }
295
296         wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
297         sem_init(wakee_event->wait_sem, 0, 0);
298         wakee_event->specific_wait = 1;
299         event->wait_sem = wakee_event->wait_sem;
300
301         nr_wakeup_events++;
302 }
303
304 static void
305 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
306                       u64 task_state __used)
307 {
308         struct sched_atom *event = get_new_event(task, timestamp);
309
310         event->type = SCHED_EVENT_SLEEP;
311
312         nr_sleep_events++;
313 }
314
315 static struct task_desc *register_pid(unsigned long pid, const char *comm)
316 {
317         struct task_desc *task;
318
319         BUG_ON(pid >= MAX_PID);
320
321         task = pid_to_task[pid];
322
323         if (task)
324                 return task;
325
326         task = calloc(1, sizeof(*task));
327         task->pid = pid;
328         task->nr = nr_tasks;
329         strcpy(task->comm, comm);
330         /*
331          * every task starts in sleeping state - this gets ignored
332          * if there's no wakeup pointing to this sleep state:
333          */
334         add_sched_event_sleep(task, 0, 0);
335
336         pid_to_task[pid] = task;
337         nr_tasks++;
338         tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
339         BUG_ON(!tasks);
340         tasks[task->nr] = task;
341
342         if (verbose)
343                 printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
344
345         return task;
346 }
347
348
349 static void print_task_traces(void)
350 {
351         struct task_desc *task;
352         unsigned long i;
353
354         for (i = 0; i < nr_tasks; i++) {
355                 task = tasks[i];
356                 printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
357                         task->nr, task->comm, task->pid, task->nr_events);
358         }
359 }
360
361 static void add_cross_task_wakeups(void)
362 {
363         struct task_desc *task1, *task2;
364         unsigned long i, j;
365
366         for (i = 0; i < nr_tasks; i++) {
367                 task1 = tasks[i];
368                 j = i + 1;
369                 if (j == nr_tasks)
370                         j = 0;
371                 task2 = tasks[j];
372                 add_sched_event_wakeup(task1, 0, task2);
373         }
374 }
375
376 static void
377 process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
378 {
379         int ret = 0;
380         u64 now;
381         long long delta;
382
383         now = get_nsecs();
384         delta = start_time + atom->timestamp - now;
385
386         switch (atom->type) {
387                 case SCHED_EVENT_RUN:
388                         burn_nsecs(atom->duration);
389                         break;
390                 case SCHED_EVENT_SLEEP:
391                         if (atom->wait_sem)
392                                 ret = sem_wait(atom->wait_sem);
393                         BUG_ON(ret);
394                         break;
395                 case SCHED_EVENT_WAKEUP:
396                         if (atom->wait_sem)
397                                 ret = sem_post(atom->wait_sem);
398                         BUG_ON(ret);
399                         break;
400                 case SCHED_EVENT_MIGRATION:
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 = threads__findnew(event->comm.tid);
640
641         dump_printf("%p [%p]: perf_event_comm: %s:%d\n",
642                 (void *)(offset + head),
643                 (void *)(long)(event->header.size),
644                 event->comm.comm, event->comm.pid);
645
646         if (thread == NULL ||
647             thread__set_comm(thread, event->comm.comm)) {
648                 dump_printf("problem processing perf_event_comm, skipping event.\n");
649                 return -1;
650         }
651         total_comm++;
652
653         return 0;
654 }
655
656
657 struct raw_event_sample {
658         u32 size;
659         char data[0];
660 };
661
662 #define FILL_FIELD(ptr, field, event, data)     \
663         ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
664
665 #define FILL_ARRAY(ptr, array, event, data)                     \
666 do {                                                            \
667         void *__array = raw_field_ptr(event, #array, data);     \
668         memcpy(ptr.array, __array, sizeof(ptr.array));  \
669 } while(0)
670
671 #define FILL_COMMON_FIELDS(ptr, event, data)                    \
672 do {                                                            \
673         FILL_FIELD(ptr, common_type, event, data);              \
674         FILL_FIELD(ptr, common_flags, event, data);             \
675         FILL_FIELD(ptr, common_preempt_count, event, data);     \
676         FILL_FIELD(ptr, common_pid, event, data);               \
677         FILL_FIELD(ptr, common_tgid, event, data);              \
678 } while (0)
679
680
681
682 struct trace_switch_event {
683         u32 size;
684
685         u16 common_type;
686         u8 common_flags;
687         u8 common_preempt_count;
688         u32 common_pid;
689         u32 common_tgid;
690
691         char prev_comm[16];
692         u32 prev_pid;
693         u32 prev_prio;
694         u64 prev_state;
695         char next_comm[16];
696         u32 next_pid;
697         u32 next_prio;
698 };
699
700 struct trace_runtime_event {
701         u32 size;
702
703         u16 common_type;
704         u8 common_flags;
705         u8 common_preempt_count;
706         u32 common_pid;
707         u32 common_tgid;
708
709         char comm[16];
710         u32 pid;
711         u64 runtime;
712         u64 vruntime;
713 };
714
715 struct trace_wakeup_event {
716         u32 size;
717
718         u16 common_type;
719         u8 common_flags;
720         u8 common_preempt_count;
721         u32 common_pid;
722         u32 common_tgid;
723
724         char comm[16];
725         u32 pid;
726
727         u32 prio;
728         u32 success;
729         u32 cpu;
730 };
731
732 struct trace_fork_event {
733         u32 size;
734
735         u16 common_type;
736         u8 common_flags;
737         u8 common_preempt_count;
738         u32 common_pid;
739         u32 common_tgid;
740
741         char parent_comm[16];
742         u32 parent_pid;
743         char child_comm[16];
744         u32 child_pid;
745 };
746
747 struct trace_migrate_task_event {
748         u32 size;
749
750         u16 common_type;
751         u8 common_flags;
752         u8 common_preempt_count;
753         u32 common_pid;
754         u32 common_tgid;
755
756         char comm[16];
757         u32 pid;
758
759         u32 prio;
760         u32 cpu;
761 };
762
763 struct trace_sched_handler {
764         void (*switch_event)(struct trace_switch_event *,
765                              struct event *,
766                              int cpu,
767                              u64 timestamp,
768                              struct thread *thread);
769
770         void (*runtime_event)(struct trace_runtime_event *,
771                               struct event *,
772                               int cpu,
773                               u64 timestamp,
774                               struct thread *thread);
775
776         void (*wakeup_event)(struct trace_wakeup_event *,
777                              struct event *,
778                              int cpu,
779                              u64 timestamp,
780                              struct thread *thread);
781
782         void (*fork_event)(struct trace_fork_event *,
783                            struct event *,
784                            int cpu,
785                            u64 timestamp,
786                            struct thread *thread);
787
788         void (*migrate_task_event)(struct trace_migrate_task_event *,
789                            struct event *,
790                            int cpu,
791                            u64 timestamp,
792                            struct thread *thread);
793 };
794
795
796 static void
797 replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
798                     struct event *event,
799                     int cpu __used,
800                     u64 timestamp __used,
801                     struct thread *thread __used)
802 {
803         struct task_desc *waker, *wakee;
804
805         if (verbose) {
806                 printf("sched_wakeup event %p\n", event);
807
808                 printf(" ... pid %d woke up %s/%d\n",
809                         wakeup_event->common_pid,
810                         wakeup_event->comm,
811                         wakeup_event->pid);
812         }
813
814         waker = register_pid(wakeup_event->common_pid, "<unknown>");
815         wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
816
817         add_sched_event_wakeup(waker, timestamp, wakee);
818 }
819
820 static u64 cpu_last_switched[MAX_CPUS];
821
822 static void
823 replay_switch_event(struct trace_switch_event *switch_event,
824                     struct event *event,
825                     int cpu,
826                     u64 timestamp,
827                     struct thread *thread __used)
828 {
829         struct task_desc *prev, *next;
830         u64 timestamp0;
831         s64 delta;
832
833         if (verbose)
834                 printf("sched_switch event %p\n", event);
835
836         if (cpu >= MAX_CPUS || cpu < 0)
837                 return;
838
839         timestamp0 = cpu_last_switched[cpu];
840         if (timestamp0)
841                 delta = timestamp - timestamp0;
842         else
843                 delta = 0;
844
845         if (delta < 0)
846                 die("hm, delta: %Ld < 0 ?\n", delta);
847
848         if (verbose) {
849                 printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
850                         switch_event->prev_comm, switch_event->prev_pid,
851                         switch_event->next_comm, switch_event->next_pid,
852                         delta);
853         }
854
855         prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
856         next = register_pid(switch_event->next_pid, switch_event->next_comm);
857
858         cpu_last_switched[cpu] = timestamp;
859
860         add_sched_event_run(prev, timestamp, delta);
861         add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
862 }
863
864
865 static void
866 replay_fork_event(struct trace_fork_event *fork_event,
867                   struct event *event,
868                   int cpu __used,
869                   u64 timestamp __used,
870                   struct thread *thread __used)
871 {
872         if (verbose) {
873                 printf("sched_fork event %p\n", event);
874                 printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
875                 printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
876         }
877         register_pid(fork_event->parent_pid, fork_event->parent_comm);
878         register_pid(fork_event->child_pid, fork_event->child_comm);
879 }
880
881 static struct trace_sched_handler replay_ops  = {
882         .wakeup_event           = replay_wakeup_event,
883         .switch_event           = replay_switch_event,
884         .fork_event             = replay_fork_event,
885 };
886
887 struct sort_dimension {
888         const char              *name;
889         sort_fn_t               cmp;
890         struct list_head        list;
891 };
892
893 static LIST_HEAD(cmp_pid);
894
895 static int
896 thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
897 {
898         struct sort_dimension *sort;
899         int ret = 0;
900
901         BUG_ON(list_empty(list));
902
903         list_for_each_entry(sort, list, list) {
904                 ret = sort->cmp(l, r);
905                 if (ret)
906                         return ret;
907         }
908
909         return ret;
910 }
911
912 static struct work_atoms *
913 thread_atoms_search(struct rb_root *root, struct thread *thread,
914                          struct list_head *sort_list)
915 {
916         struct rb_node *node = root->rb_node;
917         struct work_atoms key = { .thread = thread };
918
919         while (node) {
920                 struct work_atoms *atoms;
921                 int cmp;
922
923                 atoms = container_of(node, struct work_atoms, node);
924
925                 cmp = thread_lat_cmp(sort_list, &key, atoms);
926                 if (cmp > 0)
927                         node = node->rb_left;
928                 else if (cmp < 0)
929                         node = node->rb_right;
930                 else {
931                         BUG_ON(thread != atoms->thread);
932                         return atoms;
933                 }
934         }
935         return NULL;
936 }
937
938 static void
939 __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
940                          struct list_head *sort_list)
941 {
942         struct rb_node **new = &(root->rb_node), *parent = NULL;
943
944         while (*new) {
945                 struct work_atoms *this;
946                 int cmp;
947
948                 this = container_of(*new, struct work_atoms, node);
949                 parent = *new;
950
951                 cmp = thread_lat_cmp(sort_list, data, this);
952
953                 if (cmp > 0)
954                         new = &((*new)->rb_left);
955                 else
956                         new = &((*new)->rb_right);
957         }
958
959         rb_link_node(&data->node, parent, new);
960         rb_insert_color(&data->node, root);
961 }
962
963 static void thread_atoms_insert(struct thread *thread)
964 {
965         struct work_atoms *atoms;
966
967         atoms = calloc(sizeof(*atoms), 1);
968         if (!atoms)
969                 die("No memory");
970
971         atoms->thread = thread;
972         INIT_LIST_HEAD(&atoms->work_list);
973         __thread_latency_insert(&atom_root, atoms, &cmp_pid);
974 }
975
976 static void
977 latency_fork_event(struct trace_fork_event *fork_event __used,
978                    struct event *event __used,
979                    int cpu __used,
980                    u64 timestamp __used,
981                    struct thread *thread __used)
982 {
983         /* should insert the newcomer */
984 }
985
986 __used
987 static char sched_out_state(struct trace_switch_event *switch_event)
988 {
989         const char *str = TASK_STATE_TO_CHAR_STR;
990
991         return str[switch_event->prev_state];
992 }
993
994 static void
995 add_sched_out_event(struct work_atoms *atoms,
996                     char run_state,
997                     u64 timestamp)
998 {
999         struct work_atom *atom;
1000
1001         atom = calloc(sizeof(*atom), 1);
1002         if (!atom)
1003                 die("Non memory");
1004
1005         atom->sched_out_time = timestamp;
1006
1007         if (run_state == 'R') {
1008                 atom->state = THREAD_WAIT_CPU;
1009                 atom->wake_up_time = atom->sched_out_time;
1010         }
1011
1012         list_add_tail(&atom->list, &atoms->work_list);
1013 }
1014
1015 static void
1016 add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
1017 {
1018         struct work_atom *atom;
1019
1020         BUG_ON(list_empty(&atoms->work_list));
1021
1022         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1023
1024         atom->runtime += delta;
1025         atoms->total_runtime += delta;
1026 }
1027
1028 static void
1029 add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
1030 {
1031         struct work_atom *atom;
1032         u64 delta;
1033
1034         if (list_empty(&atoms->work_list))
1035                 return;
1036
1037         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1038
1039         if (atom->state != THREAD_WAIT_CPU)
1040                 return;
1041
1042         if (timestamp < atom->wake_up_time) {
1043                 atom->state = THREAD_IGNORE;
1044                 return;
1045         }
1046
1047         atom->state = THREAD_SCHED_IN;
1048         atom->sched_in_time = timestamp;
1049
1050         delta = atom->sched_in_time - atom->wake_up_time;
1051         atoms->total_lat += delta;
1052         if (delta > atoms->max_lat)
1053                 atoms->max_lat = delta;
1054         atoms->nb_atoms++;
1055 }
1056
1057 static void
1058 latency_switch_event(struct trace_switch_event *switch_event,
1059                      struct event *event __used,
1060                      int cpu,
1061                      u64 timestamp,
1062                      struct thread *thread __used)
1063 {
1064         struct work_atoms *out_events, *in_events;
1065         struct thread *sched_out, *sched_in;
1066         u64 timestamp0;
1067         s64 delta;
1068
1069         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1070
1071         timestamp0 = cpu_last_switched[cpu];
1072         cpu_last_switched[cpu] = timestamp;
1073         if (timestamp0)
1074                 delta = timestamp - timestamp0;
1075         else
1076                 delta = 0;
1077
1078         if (delta < 0)
1079                 die("hm, delta: %Ld < 0 ?\n", delta);
1080
1081
1082         sched_out = threads__findnew(switch_event->prev_pid);
1083         sched_in = threads__findnew(switch_event->next_pid);
1084
1085         out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1086         if (!out_events) {
1087                 thread_atoms_insert(sched_out);
1088                 out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1089                 if (!out_events)
1090                         die("out-event: Internal tree error");
1091         }
1092         add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1093
1094         in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1095         if (!in_events) {
1096                 thread_atoms_insert(sched_in);
1097                 in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1098                 if (!in_events)
1099                         die("in-event: Internal tree error");
1100                 /*
1101                  * Take came in we have not heard about yet,
1102                  * add in an initial atom in runnable state:
1103                  */
1104                 add_sched_out_event(in_events, 'R', timestamp);
1105         }
1106         add_sched_in_event(in_events, timestamp);
1107 }
1108
1109 static void
1110 latency_runtime_event(struct trace_runtime_event *runtime_event,
1111                      struct event *event __used,
1112                      int cpu,
1113                      u64 timestamp,
1114                      struct thread *this_thread __used)
1115 {
1116         struct thread *thread = threads__findnew(runtime_event->pid);
1117         struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1118
1119         BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1120         if (!atoms) {
1121                 thread_atoms_insert(thread);
1122                 atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1123                 if (!atoms)
1124                         die("in-event: Internal tree error");
1125                 add_sched_out_event(atoms, 'R', timestamp);
1126         }
1127
1128         add_runtime_event(atoms, runtime_event->runtime, timestamp);
1129 }
1130
1131 static void
1132 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1133                      struct event *__event __used,
1134                      int cpu __used,
1135                      u64 timestamp,
1136                      struct thread *thread __used)
1137 {
1138         struct work_atoms *atoms;
1139         struct work_atom *atom;
1140         struct thread *wakee;
1141
1142         /* Note for later, it may be interesting to observe the failing cases */
1143         if (!wakeup_event->success)
1144                 return;
1145
1146         wakee = threads__findnew(wakeup_event->pid);
1147         atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1148         if (!atoms) {
1149                 thread_atoms_insert(wakee);
1150                 atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1151                 if (!atoms)
1152                         die("wakeup-event: Internal tree error");
1153                 add_sched_out_event(atoms, 'S', timestamp);
1154         }
1155
1156         BUG_ON(list_empty(&atoms->work_list));
1157
1158         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1159
1160         /*
1161          * You WILL be missing events if you've recorded only
1162          * one CPU, or are only looking at only one, so don't
1163          * make useless noise.
1164          */
1165         if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1166                 nr_state_machine_bugs++;
1167
1168         nr_timestamps++;
1169         if (atom->sched_out_time > timestamp) {
1170                 nr_unordered_timestamps++;
1171                 return;
1172         }
1173
1174         atom->state = THREAD_WAIT_CPU;
1175         atom->wake_up_time = timestamp;
1176 }
1177
1178 static void
1179 latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1180                      struct event *__event __used,
1181                      int cpu __used,
1182                      u64 timestamp,
1183                      struct thread *thread __used)
1184 {
1185         struct work_atoms *atoms;
1186         struct work_atom *atom;
1187         struct thread *migrant;
1188
1189         /*
1190          * Only need to worry about migration when profiling one CPU.
1191          */
1192         if (profile_cpu == -1)
1193                 return;
1194
1195         migrant = threads__findnew(migrate_task_event->pid);
1196         atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1197         if (!atoms) {
1198                 thread_atoms_insert(migrant);
1199                 register_pid(migrant->pid, migrant->comm);
1200                 atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1201                 if (!atoms)
1202                         die("migration-event: Internal tree error");
1203                 add_sched_out_event(atoms, 'R', timestamp);
1204         }
1205
1206         BUG_ON(list_empty(&atoms->work_list));
1207
1208         atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1209         atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1210
1211         nr_timestamps++;
1212
1213         if (atom->sched_out_time > timestamp)
1214                 nr_unordered_timestamps++;
1215 }
1216
1217 static struct trace_sched_handler lat_ops  = {
1218         .wakeup_event           = latency_wakeup_event,
1219         .switch_event           = latency_switch_event,
1220         .runtime_event          = latency_runtime_event,
1221         .fork_event             = latency_fork_event,
1222         .migrate_task_event     = latency_migrate_task_event,
1223 };
1224
1225 static void output_lat_thread(struct work_atoms *work_list)
1226 {
1227         int i;
1228         int ret;
1229         u64 avg;
1230
1231         if (!work_list->nb_atoms)
1232                 return;
1233         /*
1234          * Ignore idle threads:
1235          */
1236         if (!strcmp(work_list->thread->comm, "swapper"))
1237                 return;
1238
1239         all_runtime += work_list->total_runtime;
1240         all_count += work_list->nb_atoms;
1241
1242         ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
1243
1244         for (i = 0; i < 24 - ret; i++)
1245                 printf(" ");
1246
1247         avg = work_list->total_lat / work_list->nb_atoms;
1248
1249         printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
1250               (double)work_list->total_runtime / 1e6,
1251                  work_list->nb_atoms, (double)avg / 1e6,
1252                  (double)work_list->max_lat / 1e6);
1253 }
1254
1255 static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1256 {
1257         if (l->thread->pid < r->thread->pid)
1258                 return -1;
1259         if (l->thread->pid > r->thread->pid)
1260                 return 1;
1261
1262         return 0;
1263 }
1264
1265 static struct sort_dimension pid_sort_dimension = {
1266         .name                   = "pid",
1267         .cmp                    = pid_cmp,
1268 };
1269
1270 static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1271 {
1272         u64 avgl, avgr;
1273
1274         if (!l->nb_atoms)
1275                 return -1;
1276
1277         if (!r->nb_atoms)
1278                 return 1;
1279
1280         avgl = l->total_lat / l->nb_atoms;
1281         avgr = r->total_lat / r->nb_atoms;
1282
1283         if (avgl < avgr)
1284                 return -1;
1285         if (avgl > avgr)
1286                 return 1;
1287
1288         return 0;
1289 }
1290
1291 static struct sort_dimension avg_sort_dimension = {
1292         .name                   = "avg",
1293         .cmp                    = avg_cmp,
1294 };
1295
1296 static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1297 {
1298         if (l->max_lat < r->max_lat)
1299                 return -1;
1300         if (l->max_lat > r->max_lat)
1301                 return 1;
1302
1303         return 0;
1304 }
1305
1306 static struct sort_dimension max_sort_dimension = {
1307         .name                   = "max",
1308         .cmp                    = max_cmp,
1309 };
1310
1311 static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1312 {
1313         if (l->nb_atoms < r->nb_atoms)
1314                 return -1;
1315         if (l->nb_atoms > r->nb_atoms)
1316                 return 1;
1317
1318         return 0;
1319 }
1320
1321 static struct sort_dimension switch_sort_dimension = {
1322         .name                   = "switch",
1323         .cmp                    = switch_cmp,
1324 };
1325
1326 static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1327 {
1328         if (l->total_runtime < r->total_runtime)
1329                 return -1;
1330         if (l->total_runtime > r->total_runtime)
1331                 return 1;
1332
1333         return 0;
1334 }
1335
1336 static struct sort_dimension runtime_sort_dimension = {
1337         .name                   = "runtime",
1338         .cmp                    = runtime_cmp,
1339 };
1340
1341 static struct sort_dimension *available_sorts[] = {
1342         &pid_sort_dimension,
1343         &avg_sort_dimension,
1344         &max_sort_dimension,
1345         &switch_sort_dimension,
1346         &runtime_sort_dimension,
1347 };
1348
1349 #define NB_AVAILABLE_SORTS      (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1350
1351 static LIST_HEAD(sort_list);
1352
1353 static int sort_dimension__add(const char *tok, struct list_head *list)
1354 {
1355         int i;
1356
1357         for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1358                 if (!strcmp(available_sorts[i]->name, tok)) {
1359                         list_add_tail(&available_sorts[i]->list, list);
1360
1361                         return 0;
1362                 }
1363         }
1364
1365         return -1;
1366 }
1367
1368 static void setup_sorting(void);
1369
1370 static void sort_lat(void)
1371 {
1372         struct rb_node *node;
1373
1374         for (;;) {
1375                 struct work_atoms *data;
1376                 node = rb_first(&atom_root);
1377                 if (!node)
1378                         break;
1379
1380                 rb_erase(node, &atom_root);
1381                 data = rb_entry(node, struct work_atoms, node);
1382                 __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1383         }
1384 }
1385
1386 static struct trace_sched_handler *trace_handler;
1387
1388 static void
1389 process_sched_wakeup_event(struct raw_event_sample *raw,
1390                            struct event *event,
1391                            int cpu __used,
1392                            u64 timestamp __used,
1393                            struct thread *thread __used)
1394 {
1395         struct trace_wakeup_event wakeup_event;
1396
1397         FILL_COMMON_FIELDS(wakeup_event, event, raw->data);
1398
1399         FILL_ARRAY(wakeup_event, comm, event, raw->data);
1400         FILL_FIELD(wakeup_event, pid, event, raw->data);
1401         FILL_FIELD(wakeup_event, prio, event, raw->data);
1402         FILL_FIELD(wakeup_event, success, event, raw->data);
1403         FILL_FIELD(wakeup_event, cpu, event, raw->data);
1404
1405         if (trace_handler->wakeup_event)
1406                 trace_handler->wakeup_event(&wakeup_event, event, cpu, timestamp, thread);
1407 }
1408
1409 /*
1410  * Track the current task - that way we can know whether there's any
1411  * weird events, such as a task being switched away that is not current.
1412  */
1413 static int max_cpu;
1414
1415 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1416
1417 static struct thread *curr_thread[MAX_CPUS];
1418
1419 static char next_shortname1 = 'A';
1420 static char next_shortname2 = '0';
1421
1422 static void
1423 map_switch_event(struct trace_switch_event *switch_event,
1424                  struct event *event __used,
1425                  int this_cpu,
1426                  u64 timestamp,
1427                  struct thread *thread __used)
1428 {
1429         struct thread *sched_out, *sched_in;
1430         int new_shortname;
1431         u64 timestamp0;
1432         s64 delta;
1433         int cpu;
1434
1435         BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1436
1437         if (this_cpu > max_cpu)
1438                 max_cpu = this_cpu;
1439
1440         timestamp0 = cpu_last_switched[this_cpu];
1441         cpu_last_switched[this_cpu] = timestamp;
1442         if (timestamp0)
1443                 delta = timestamp - timestamp0;
1444         else
1445                 delta = 0;
1446
1447         if (delta < 0)
1448                 die("hm, delta: %Ld < 0 ?\n", delta);
1449
1450
1451         sched_out = threads__findnew(switch_event->prev_pid);
1452         sched_in = threads__findnew(switch_event->next_pid);
1453
1454         curr_thread[this_cpu] = sched_in;
1455
1456         printf("  ");
1457
1458         new_shortname = 0;
1459         if (!sched_in->shortname[0]) {
1460                 sched_in->shortname[0] = next_shortname1;
1461                 sched_in->shortname[1] = next_shortname2;
1462
1463                 if (next_shortname1 < 'Z') {
1464                         next_shortname1++;
1465                 } else {
1466                         next_shortname1='A';
1467                         if (next_shortname2 < '9') {
1468                                 next_shortname2++;
1469                         } else {
1470                                 next_shortname2='0';
1471                         }
1472                 }
1473                 new_shortname = 1;
1474         }
1475
1476         for (cpu = 0; cpu <= max_cpu; cpu++) {
1477                 if (cpu != this_cpu)
1478                         printf(" ");
1479                 else
1480                         printf("*");
1481
1482                 if (curr_thread[cpu]) {
1483                         if (curr_thread[cpu]->pid)
1484                                 printf("%2s ", curr_thread[cpu]->shortname);
1485                         else
1486                                 printf(".  ");
1487                 } else
1488                         printf("   ");
1489         }
1490
1491         printf("  %12.6f secs ", (double)timestamp/1e9);
1492         if (new_shortname) {
1493                 printf("%s => %s:%d\n",
1494                         sched_in->shortname, sched_in->comm, sched_in->pid);
1495         } else {
1496                 printf("\n");
1497         }
1498 }
1499
1500
1501 static void
1502 process_sched_switch_event(struct raw_event_sample *raw,
1503                            struct event *event,
1504                            int this_cpu,
1505                            u64 timestamp __used,
1506                            struct thread *thread __used)
1507 {
1508         struct trace_switch_event switch_event;
1509
1510         FILL_COMMON_FIELDS(switch_event, event, raw->data);
1511
1512         FILL_ARRAY(switch_event, prev_comm, event, raw->data);
1513         FILL_FIELD(switch_event, prev_pid, event, raw->data);
1514         FILL_FIELD(switch_event, prev_prio, event, raw->data);
1515         FILL_FIELD(switch_event, prev_state, event, raw->data);
1516         FILL_ARRAY(switch_event, next_comm, event, raw->data);
1517         FILL_FIELD(switch_event, next_pid, event, raw->data);
1518         FILL_FIELD(switch_event, next_prio, event, raw->data);
1519
1520         if (curr_pid[this_cpu] != (u32)-1) {
1521                 /*
1522                  * Are we trying to switch away a PID that is
1523                  * not current?
1524                  */
1525                 if (curr_pid[this_cpu] != switch_event.prev_pid)
1526                         nr_context_switch_bugs++;
1527         }
1528         if (trace_handler->switch_event)
1529                 trace_handler->switch_event(&switch_event, event, this_cpu, timestamp, thread);
1530
1531         curr_pid[this_cpu] = switch_event.next_pid;
1532 }
1533
1534 static void
1535 process_sched_runtime_event(struct raw_event_sample *raw,
1536                            struct event *event,
1537                            int cpu __used,
1538                            u64 timestamp __used,
1539                            struct thread *thread __used)
1540 {
1541         struct trace_runtime_event runtime_event;
1542
1543         FILL_ARRAY(runtime_event, comm, event, raw->data);
1544         FILL_FIELD(runtime_event, pid, event, raw->data);
1545         FILL_FIELD(runtime_event, runtime, event, raw->data);
1546         FILL_FIELD(runtime_event, vruntime, event, raw->data);
1547
1548         if (trace_handler->runtime_event)
1549                 trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
1550 }
1551
1552 static void
1553 process_sched_fork_event(struct raw_event_sample *raw,
1554                          struct event *event,
1555                          int cpu __used,
1556                          u64 timestamp __used,
1557                          struct thread *thread __used)
1558 {
1559         struct trace_fork_event fork_event;
1560
1561         FILL_COMMON_FIELDS(fork_event, event, raw->data);
1562
1563         FILL_ARRAY(fork_event, parent_comm, event, raw->data);
1564         FILL_FIELD(fork_event, parent_pid, event, raw->data);
1565         FILL_ARRAY(fork_event, child_comm, event, raw->data);
1566         FILL_FIELD(fork_event, child_pid, event, raw->data);
1567
1568         if (trace_handler->fork_event)
1569                 trace_handler->fork_event(&fork_event, event, cpu, timestamp, thread);
1570 }
1571
1572 static void
1573 process_sched_exit_event(struct event *event,
1574                          int cpu __used,
1575                          u64 timestamp __used,
1576                          struct thread *thread __used)
1577 {
1578         if (verbose)
1579                 printf("sched_exit event %p\n", event);
1580 }
1581
1582 static void
1583 process_sched_migrate_task_event(struct raw_event_sample *raw,
1584                            struct event *event,
1585                            int cpu __used,
1586                            u64 timestamp __used,
1587                            struct thread *thread __used)
1588 {
1589         struct trace_migrate_task_event migrate_task_event;
1590
1591         FILL_COMMON_FIELDS(migrate_task_event, event, raw->data);
1592
1593         FILL_ARRAY(migrate_task_event, comm, event, raw->data);
1594         FILL_FIELD(migrate_task_event, pid, event, raw->data);
1595         FILL_FIELD(migrate_task_event, prio, event, raw->data);
1596         FILL_FIELD(migrate_task_event, cpu, event, raw->data);
1597
1598         if (trace_handler->migrate_task_event)
1599                 trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
1600 }
1601
1602 static void
1603 process_raw_event(event_t *raw_event __used, void *more_data,
1604                   int cpu, u64 timestamp, struct thread *thread)
1605 {
1606         struct raw_event_sample *raw = more_data;
1607         struct event *event;
1608         int type;
1609
1610         type = trace_parse_common_type(raw->data);
1611         event = trace_find_event(type);
1612
1613         if (!strcmp(event->name, "sched_switch"))
1614                 process_sched_switch_event(raw, event, cpu, timestamp, thread);
1615         if (!strcmp(event->name, "sched_stat_runtime"))
1616                 process_sched_runtime_event(raw, event, cpu, timestamp, thread);
1617         if (!strcmp(event->name, "sched_wakeup"))
1618                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1619         if (!strcmp(event->name, "sched_wakeup_new"))
1620                 process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
1621         if (!strcmp(event->name, "sched_process_fork"))
1622                 process_sched_fork_event(raw, event, cpu, timestamp, thread);
1623         if (!strcmp(event->name, "sched_process_exit"))
1624                 process_sched_exit_event(event, cpu, timestamp, thread);
1625         if (!strcmp(event->name, "sched_migrate_task"))
1626                 process_sched_migrate_task_event(raw, event, cpu, timestamp, thread);
1627 }
1628
1629 static int
1630 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
1631 {
1632         struct thread *thread;
1633         u64 ip = event->ip.ip;
1634         u64 timestamp = -1;
1635         u32 cpu = -1;
1636         u64 period = 1;
1637         void *more_data = event->ip.__more_data;
1638
1639         if (!(sample_type & PERF_SAMPLE_RAW))
1640                 return 0;
1641
1642         thread = threads__findnew(event->ip.pid);
1643
1644         if (sample_type & PERF_SAMPLE_TIME) {
1645                 timestamp = *(u64 *)more_data;
1646                 more_data += sizeof(u64);
1647         }
1648
1649         if (sample_type & PERF_SAMPLE_CPU) {
1650                 cpu = *(u32 *)more_data;
1651                 more_data += sizeof(u32);
1652                 more_data += sizeof(u32); /* reserved */
1653         }
1654
1655         if (sample_type & PERF_SAMPLE_PERIOD) {
1656                 period = *(u64 *)more_data;
1657                 more_data += sizeof(u64);
1658         }
1659
1660         dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
1661                 (void *)(offset + head),
1662                 (void *)(long)(event->header.size),
1663                 event->header.misc,
1664                 event->ip.pid, event->ip.tid,
1665                 (void *)(long)ip,
1666                 (long long)period);
1667
1668         if (thread == NULL) {
1669                 pr_debug("problem processing %d event, skipping it.\n",
1670                          event->header.type);
1671                 return -1;
1672         }
1673
1674         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1675
1676         if (profile_cpu != -1 && profile_cpu != (int) cpu)
1677                 return 0;
1678
1679         process_raw_event(event, more_data, cpu, timestamp, thread);
1680
1681         return 0;
1682 }
1683
1684 static int
1685 process_lost_event(event_t *event __used,
1686                    unsigned long offset __used,
1687                    unsigned long head __used)
1688 {
1689         nr_lost_chunks++;
1690         nr_lost_events += event->lost.lost;
1691
1692         return 0;
1693 }
1694
1695 static int sample_type_check(u64 type)
1696 {
1697         sample_type = type;
1698
1699         if (!(sample_type & PERF_SAMPLE_RAW)) {
1700                 fprintf(stderr,
1701                         "No trace sample to read. Did you call perf record "
1702                         "without -R?");
1703                 return -1;
1704         }
1705
1706         return 0;
1707 }
1708
1709 static struct perf_file_handler file_handler = {
1710         .process_sample_event   = process_sample_event,
1711         .process_comm_event     = process_comm_event,
1712         .process_lost_event     = process_lost_event,
1713         .sample_type_check      = sample_type_check,
1714 };
1715
1716 static int read_events(void)
1717 {
1718         register_idle_thread();
1719         register_perf_file_handler(&file_handler);
1720
1721         return mmap_dispatch_perf_file(&header, input_name, 0, 0,
1722                                        &cwdlen, &cwd);
1723 }
1724
1725 static void print_bad_events(void)
1726 {
1727         if (nr_unordered_timestamps && nr_timestamps) {
1728                 printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1729                         (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1730                         nr_unordered_timestamps, nr_timestamps);
1731         }
1732         if (nr_lost_events && nr_events) {
1733                 printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1734                         (double)nr_lost_events/(double)nr_events*100.0,
1735                         nr_lost_events, nr_events, nr_lost_chunks);
1736         }
1737         if (nr_state_machine_bugs && nr_timestamps) {
1738                 printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
1739                         (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1740                         nr_state_machine_bugs, nr_timestamps);
1741                 if (nr_lost_events)
1742                         printf(" (due to lost events?)");
1743                 printf("\n");
1744         }
1745         if (nr_context_switch_bugs && nr_timestamps) {
1746                 printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
1747                         (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1748                         nr_context_switch_bugs, nr_timestamps);
1749                 if (nr_lost_events)
1750                         printf(" (due to lost events?)");
1751                 printf("\n");
1752         }
1753 }
1754
1755 static void __cmd_lat(void)
1756 {
1757         struct rb_node *next;
1758
1759         setup_pager();
1760         read_events();
1761         sort_lat();
1762
1763         printf("\n -----------------------------------------------------------------------------------------\n");
1764         printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms |\n");
1765         printf(" -----------------------------------------------------------------------------------------\n");
1766
1767         next = rb_first(&sorted_atom_root);
1768
1769         while (next) {
1770                 struct work_atoms *work_list;
1771
1772                 work_list = rb_entry(next, struct work_atoms, node);
1773                 output_lat_thread(work_list);
1774                 next = rb_next(next);
1775         }
1776
1777         printf(" -----------------------------------------------------------------------------------------\n");
1778         printf("  TOTAL:                |%11.3f ms |%9Ld |\n",
1779                 (double)all_runtime/1e6, all_count);
1780
1781         printf(" ---------------------------------------------------\n");
1782
1783         print_bad_events();
1784         printf("\n");
1785
1786 }
1787
1788 static struct trace_sched_handler map_ops  = {
1789         .wakeup_event           = NULL,
1790         .switch_event           = map_switch_event,
1791         .runtime_event          = NULL,
1792         .fork_event             = NULL,
1793 };
1794
1795 static void __cmd_map(void)
1796 {
1797         max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1798
1799         setup_pager();
1800         read_events();
1801         print_bad_events();
1802 }
1803
1804 static void __cmd_replay(void)
1805 {
1806         unsigned long i;
1807
1808         calibrate_run_measurement_overhead();
1809         calibrate_sleep_measurement_overhead();
1810
1811         test_calibrations();
1812
1813         read_events();
1814
1815         printf("nr_run_events:        %ld\n", nr_run_events);
1816         printf("nr_sleep_events:      %ld\n", nr_sleep_events);
1817         printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
1818
1819         if (targetless_wakeups)
1820                 printf("target-less wakeups:  %ld\n", targetless_wakeups);
1821         if (multitarget_wakeups)
1822                 printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1823         if (nr_run_events_optimized)
1824                 printf("run atoms optimized: %ld\n",
1825                         nr_run_events_optimized);
1826
1827         print_task_traces();
1828         add_cross_task_wakeups();
1829
1830         create_tasks();
1831         printf("------------------------------------------------------------\n");
1832         for (i = 0; i < replay_repeat; i++)
1833                 run_one_test();
1834 }
1835
1836
1837 static const char * const sched_usage[] = {
1838         "perf sched [<options>] {record|latency|map|replay|trace}",
1839         NULL
1840 };
1841
1842 static const struct option sched_options[] = {
1843         OPT_STRING('i', "input", &input_name, "file",
1844                     "input file name"),
1845         OPT_BOOLEAN('v', "verbose", &verbose,
1846                     "be more verbose (show symbol address, etc)"),
1847         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1848                     "dump raw trace in ASCII"),
1849         OPT_END()
1850 };
1851
1852 static const char * const latency_usage[] = {
1853         "perf sched latency [<options>]",
1854         NULL
1855 };
1856
1857 static const struct option latency_options[] = {
1858         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1859                    "sort by key(s): runtime, switch, avg, max"),
1860         OPT_BOOLEAN('v', "verbose", &verbose,
1861                     "be more verbose (show symbol address, etc)"),
1862         OPT_INTEGER('C', "CPU", &profile_cpu,
1863                     "CPU to profile on"),
1864         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1865                     "dump raw trace in ASCII"),
1866         OPT_END()
1867 };
1868
1869 static const char * const replay_usage[] = {
1870         "perf sched replay [<options>]",
1871         NULL
1872 };
1873
1874 static const struct option replay_options[] = {
1875         OPT_INTEGER('r', "repeat", &replay_repeat,
1876                     "repeat the workload replay N times (-1: infinite)"),
1877         OPT_BOOLEAN('v', "verbose", &verbose,
1878                     "be more verbose (show symbol address, etc)"),
1879         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1880                     "dump raw trace in ASCII"),
1881         OPT_END()
1882 };
1883
1884 static void setup_sorting(void)
1885 {
1886         char *tmp, *tok, *str = strdup(sort_order);
1887
1888         for (tok = strtok_r(str, ", ", &tmp);
1889                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
1890                 if (sort_dimension__add(tok, &sort_list) < 0) {
1891                         error("Unknown --sort key: `%s'", tok);
1892                         usage_with_options(latency_usage, latency_options);
1893                 }
1894         }
1895
1896         free(str);
1897
1898         sort_dimension__add("pid", &cmp_pid);
1899 }
1900
1901 static const char *record_args[] = {
1902         "record",
1903         "-a",
1904         "-R",
1905         "-M",
1906         "-f",
1907         "-m", "1024",
1908         "-c", "1",
1909         "-e", "sched:sched_switch:r",
1910         "-e", "sched:sched_stat_wait:r",
1911         "-e", "sched:sched_stat_sleep:r",
1912         "-e", "sched:sched_stat_iowait:r",
1913         "-e", "sched:sched_stat_runtime:r",
1914         "-e", "sched:sched_process_exit:r",
1915         "-e", "sched:sched_process_fork:r",
1916         "-e", "sched:sched_wakeup:r",
1917         "-e", "sched:sched_migrate_task:r",
1918 };
1919
1920 static int __cmd_record(int argc, const char **argv)
1921 {
1922         unsigned int rec_argc, i, j;
1923         const char **rec_argv;
1924
1925         rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1926         rec_argv = calloc(rec_argc + 1, sizeof(char *));
1927
1928         for (i = 0; i < ARRAY_SIZE(record_args); i++)
1929                 rec_argv[i] = strdup(record_args[i]);
1930
1931         for (j = 1; j < (unsigned int)argc; j++, i++)
1932                 rec_argv[i] = argv[j];
1933
1934         BUG_ON(i != rec_argc);
1935
1936         return cmd_record(i, rec_argv, NULL);
1937 }
1938
1939 int cmd_sched(int argc, const char **argv, const char *prefix __used)
1940 {
1941         symbol__init(0);
1942
1943         argc = parse_options(argc, argv, sched_options, sched_usage,
1944                              PARSE_OPT_STOP_AT_NON_OPTION);
1945         if (!argc)
1946                 usage_with_options(sched_usage, sched_options);
1947
1948         if (!strncmp(argv[0], "rec", 3)) {
1949                 return __cmd_record(argc, argv);
1950         } else if (!strncmp(argv[0], "lat", 3)) {
1951                 trace_handler = &lat_ops;
1952                 if (argc > 1) {
1953                         argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1954                         if (argc)
1955                                 usage_with_options(latency_usage, latency_options);
1956                 }
1957                 setup_sorting();
1958                 __cmd_lat();
1959         } else if (!strcmp(argv[0], "map")) {
1960                 trace_handler = &map_ops;
1961                 setup_sorting();
1962                 __cmd_map();
1963         } else if (!strncmp(argv[0], "rep", 3)) {
1964                 trace_handler = &replay_ops;
1965                 if (argc) {
1966                         argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1967                         if (argc)
1968                                 usage_with_options(replay_usage, replay_options);
1969                 }
1970                 __cmd_replay();
1971         } else if (!strcmp(argv[0], "trace")) {
1972                 /*
1973                  * Aliased to 'perf trace' for now:
1974                  */
1975                 return cmd_trace(argc, argv, prefix);
1976         } else {
1977                 usage_with_options(sched_usage, sched_options);
1978         }
1979
1980         return 0;
1981 }