Root/tools/perf/builtin-sched.c

1#include "builtin.h"
2#include "perf.h"
3
4#include "util/util.h"
5#include "util/cache.h"
6#include "util/symbol.h"
7#include "util/thread.h"
8#include "util/header.h"
9#include "util/session.h"
10
11#include "util/parse-options.h"
12#include "util/trace-event.h"
13
14#include "util/debug.h"
15
16#include <sys/prctl.h>
17
18#include <semaphore.h>
19#include <pthread.h>
20#include <math.h>
21
22static char const *input_name = "perf.data";
23
24static char default_sort_order[] = "avg, max, switch, runtime";
25static char *sort_order = default_sort_order;
26
27static int profile_cpu = -1;
28
29#define PR_SET_NAME 15 /* Set process name */
30#define MAX_CPUS 4096
31
32static u64 run_measurement_overhead;
33static u64 sleep_measurement_overhead;
34
35#define COMM_LEN 20
36#define SYM_LEN 129
37
38#define MAX_PID 65536
39
40static unsigned long nr_tasks;
41
42struct sched_atom;
43
44struct task_desc {
45    unsigned long nr;
46    unsigned long pid;
47    char comm[COMM_LEN];
48
49    unsigned long nr_events;
50    unsigned long curr_event;
51    struct sched_atom **atoms;
52
53    pthread_t thread;
54    sem_t sleep_sem;
55
56    sem_t ready_for_work;
57    sem_t work_done_sem;
58
59    u64 cpu_usage;
60};
61
62enum sched_event_type {
63    SCHED_EVENT_RUN,
64    SCHED_EVENT_SLEEP,
65    SCHED_EVENT_WAKEUP,
66    SCHED_EVENT_MIGRATION,
67};
68
69struct sched_atom {
70    enum sched_event_type type;
71    u64 timestamp;
72    u64 duration;
73    unsigned long nr;
74    int specific_wait;
75    sem_t *wait_sem;
76    struct task_desc *wakee;
77};
78
79static struct task_desc *pid_to_task[MAX_PID];
80
81static struct task_desc **tasks;
82
83static pthread_mutex_t start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
84static u64 start_time;
85
86static pthread_mutex_t work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
87
88static unsigned long nr_run_events;
89static unsigned long nr_sleep_events;
90static unsigned long nr_wakeup_events;
91
92static unsigned long nr_sleep_corrections;
93static unsigned long nr_run_events_optimized;
94
95static unsigned long targetless_wakeups;
96static unsigned long multitarget_wakeups;
97
98static u64 cpu_usage;
99static u64 runavg_cpu_usage;
100static u64 parent_cpu_usage;
101static u64 runavg_parent_cpu_usage;
102
103static unsigned long nr_runs;
104static u64 sum_runtime;
105static u64 sum_fluct;
106static u64 run_avg;
107
108static unsigned long replay_repeat = 10;
109static unsigned long nr_timestamps;
110static unsigned long nr_unordered_timestamps;
111static unsigned long nr_state_machine_bugs;
112static unsigned long nr_context_switch_bugs;
113static unsigned long nr_events;
114static unsigned long nr_lost_chunks;
115static unsigned long nr_lost_events;
116
117#define TASK_STATE_TO_CHAR_STR "RSDTtZX"
118
119enum thread_state {
120    THREAD_SLEEPING = 0,
121    THREAD_WAIT_CPU,
122    THREAD_SCHED_IN,
123    THREAD_IGNORE
124};
125
126struct work_atom {
127    struct list_head list;
128    enum thread_state state;
129    u64 sched_out_time;
130    u64 wake_up_time;
131    u64 sched_in_time;
132    u64 runtime;
133};
134
135struct work_atoms {
136    struct list_head work_list;
137    struct thread *thread;
138    struct rb_node node;
139    u64 max_lat;
140    u64 max_lat_at;
141    u64 total_lat;
142    u64 nb_atoms;
143    u64 total_runtime;
144};
145
146typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
147
148static struct rb_root atom_root, sorted_atom_root;
149
150static u64 all_runtime;
151static u64 all_count;
152
153
154static u64 get_nsecs(void)
155{
156    struct timespec ts;
157
158    clock_gettime(CLOCK_MONOTONIC, &ts);
159
160    return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
161}
162
163static void burn_nsecs(u64 nsecs)
164{
165    u64 T0 = get_nsecs(), T1;
166
167    do {
168        T1 = get_nsecs();
169    } while (T1 + run_measurement_overhead < T0 + nsecs);
170}
171
172static void sleep_nsecs(u64 nsecs)
173{
174    struct timespec ts;
175
176    ts.tv_nsec = nsecs % 999999999;
177    ts.tv_sec = nsecs / 999999999;
178
179    nanosleep(&ts, NULL);
180}
181
182static void calibrate_run_measurement_overhead(void)
183{
184    u64 T0, T1, delta, min_delta = 1000000000ULL;
185    int i;
186
187    for (i = 0; i < 10; i++) {
188        T0 = get_nsecs();
189        burn_nsecs(0);
190        T1 = get_nsecs();
191        delta = T1-T0;
192        min_delta = min(min_delta, delta);
193    }
194    run_measurement_overhead = min_delta;
195
196    printf("run measurement overhead: %Ld nsecs\n", min_delta);
197}
198
199static void calibrate_sleep_measurement_overhead(void)
200{
201    u64 T0, T1, delta, min_delta = 1000000000ULL;
202    int i;
203
204    for (i = 0; i < 10; i++) {
205        T0 = get_nsecs();
206        sleep_nsecs(10000);
207        T1 = get_nsecs();
208        delta = T1-T0;
209        min_delta = min(min_delta, delta);
210    }
211    min_delta -= 10000;
212    sleep_measurement_overhead = min_delta;
213
214    printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
215}
216
217static struct sched_atom *
218get_new_event(struct task_desc *task, u64 timestamp)
219{
220    struct sched_atom *event = zalloc(sizeof(*event));
221    unsigned long idx = task->nr_events;
222    size_t size;
223
224    event->timestamp = timestamp;
225    event->nr = idx;
226
227    task->nr_events++;
228    size = sizeof(struct sched_atom *) * task->nr_events;
229    task->atoms = realloc(task->atoms, size);
230    BUG_ON(!task->atoms);
231
232    task->atoms[idx] = event;
233
234    return event;
235}
236
237static struct sched_atom *last_event(struct task_desc *task)
238{
239    if (!task->nr_events)
240        return NULL;
241
242    return task->atoms[task->nr_events - 1];
243}
244
245static void
246add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
247{
248    struct sched_atom *event, *curr_event = last_event(task);
249
250    /*
251     * optimize an existing RUN event by merging this one
252     * to it:
253     */
254    if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
255        nr_run_events_optimized++;
256        curr_event->duration += duration;
257        return;
258    }
259
260    event = get_new_event(task, timestamp);
261
262    event->type = SCHED_EVENT_RUN;
263    event->duration = duration;
264
265    nr_run_events++;
266}
267
268static void
269add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
270               struct task_desc *wakee)
271{
272    struct sched_atom *event, *wakee_event;
273
274    event = get_new_event(task, timestamp);
275    event->type = SCHED_EVENT_WAKEUP;
276    event->wakee = wakee;
277
278    wakee_event = last_event(wakee);
279    if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
280        targetless_wakeups++;
281        return;
282    }
283    if (wakee_event->wait_sem) {
284        multitarget_wakeups++;
285        return;
286    }
287
288    wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
289    sem_init(wakee_event->wait_sem, 0, 0);
290    wakee_event->specific_wait = 1;
291    event->wait_sem = wakee_event->wait_sem;
292
293    nr_wakeup_events++;
294}
295
296static void
297add_sched_event_sleep(struct task_desc *task, u64 timestamp,
298              u64 task_state __used)
299{
300    struct sched_atom *event = get_new_event(task, timestamp);
301
302    event->type = SCHED_EVENT_SLEEP;
303
304    nr_sleep_events++;
305}
306
307static struct task_desc *register_pid(unsigned long pid, const char *comm)
308{
309    struct task_desc *task;
310
311    BUG_ON(pid >= MAX_PID);
312
313    task = pid_to_task[pid];
314
315    if (task)
316        return task;
317
318    task = zalloc(sizeof(*task));
319    task->pid = pid;
320    task->nr = nr_tasks;
321    strcpy(task->comm, comm);
322    /*
323     * every task starts in sleeping state - this gets ignored
324     * if there's no wakeup pointing to this sleep state:
325     */
326    add_sched_event_sleep(task, 0, 0);
327
328    pid_to_task[pid] = task;
329    nr_tasks++;
330    tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
331    BUG_ON(!tasks);
332    tasks[task->nr] = task;
333
334    if (verbose)
335        printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
336
337    return task;
338}
339
340
341static void print_task_traces(void)
342{
343    struct task_desc *task;
344    unsigned long i;
345
346    for (i = 0; i < nr_tasks; i++) {
347        task = tasks[i];
348        printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
349            task->nr, task->comm, task->pid, task->nr_events);
350    }
351}
352
353static void add_cross_task_wakeups(void)
354{
355    struct task_desc *task1, *task2;
356    unsigned long i, j;
357
358    for (i = 0; i < nr_tasks; i++) {
359        task1 = tasks[i];
360        j = i + 1;
361        if (j == nr_tasks)
362            j = 0;
363        task2 = tasks[j];
364        add_sched_event_wakeup(task1, 0, task2);
365    }
366}
367
368static void
369process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
370{
371    int ret = 0;
372    u64 now;
373    long long delta;
374
375    now = get_nsecs();
376    delta = start_time + atom->timestamp - now;
377
378    switch (atom->type) {
379        case SCHED_EVENT_RUN:
380            burn_nsecs(atom->duration);
381            break;
382        case SCHED_EVENT_SLEEP:
383            if (atom->wait_sem)
384                ret = sem_wait(atom->wait_sem);
385            BUG_ON(ret);
386            break;
387        case SCHED_EVENT_WAKEUP:
388            if (atom->wait_sem)
389                ret = sem_post(atom->wait_sem);
390            BUG_ON(ret);
391            break;
392        case SCHED_EVENT_MIGRATION:
393            break;
394        default:
395            BUG_ON(1);
396    }
397}
398
399static u64 get_cpu_usage_nsec_parent(void)
400{
401    struct rusage ru;
402    u64 sum;
403    int err;
404
405    err = getrusage(RUSAGE_SELF, &ru);
406    BUG_ON(err);
407
408    sum = ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
409    sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
410
411    return sum;
412}
413
414static int self_open_counters(void)
415{
416    struct perf_event_attr attr;
417    int fd;
418
419    memset(&attr, 0, sizeof(attr));
420
421    attr.type = PERF_TYPE_SOFTWARE;
422    attr.config = PERF_COUNT_SW_TASK_CLOCK;
423
424    fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
425
426    if (fd < 0)
427        die("Error: sys_perf_event_open() syscall returned"
428            "with %d (%s)\n", fd, strerror(errno));
429    return fd;
430}
431
432static u64 get_cpu_usage_nsec_self(int fd)
433{
434    u64 runtime;
435    int ret;
436
437    ret = read(fd, &runtime, sizeof(runtime));
438    BUG_ON(ret != sizeof(runtime));
439
440    return runtime;
441}
442
443static void *thread_func(void *ctx)
444{
445    struct task_desc *this_task = ctx;
446    u64 cpu_usage_0, cpu_usage_1;
447    unsigned long i, ret;
448    char comm2[22];
449    int fd;
450
451    sprintf(comm2, ":%s", this_task->comm);
452    prctl(PR_SET_NAME, comm2);
453    fd = self_open_counters();
454
455again:
456    ret = sem_post(&this_task->ready_for_work);
457    BUG_ON(ret);
458    ret = pthread_mutex_lock(&start_work_mutex);
459    BUG_ON(ret);
460    ret = pthread_mutex_unlock(&start_work_mutex);
461    BUG_ON(ret);
462
463    cpu_usage_0 = get_cpu_usage_nsec_self(fd);
464
465    for (i = 0; i < this_task->nr_events; i++) {
466        this_task->curr_event = i;
467        process_sched_event(this_task, this_task->atoms[i]);
468    }
469
470    cpu_usage_1 = get_cpu_usage_nsec_self(fd);
471    this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
472    ret = sem_post(&this_task->work_done_sem);
473    BUG_ON(ret);
474
475    ret = pthread_mutex_lock(&work_done_wait_mutex);
476    BUG_ON(ret);
477    ret = pthread_mutex_unlock(&work_done_wait_mutex);
478    BUG_ON(ret);
479
480    goto again;
481}
482
483static void create_tasks(void)
484{
485    struct task_desc *task;
486    pthread_attr_t attr;
487    unsigned long i;
488    int err;
489
490    err = pthread_attr_init(&attr);
491    BUG_ON(err);
492    err = pthread_attr_setstacksize(&attr, (size_t)(16*1024));
493    BUG_ON(err);
494    err = pthread_mutex_lock(&start_work_mutex);
495    BUG_ON(err);
496    err = pthread_mutex_lock(&work_done_wait_mutex);
497    BUG_ON(err);
498    for (i = 0; i < nr_tasks; i++) {
499        task = tasks[i];
500        sem_init(&task->sleep_sem, 0, 0);
501        sem_init(&task->ready_for_work, 0, 0);
502        sem_init(&task->work_done_sem, 0, 0);
503        task->curr_event = 0;
504        err = pthread_create(&task->thread, &attr, thread_func, task);
505        BUG_ON(err);
506    }
507}
508
509static void wait_for_tasks(void)
510{
511    u64 cpu_usage_0, cpu_usage_1;
512    struct task_desc *task;
513    unsigned long i, ret;
514
515    start_time = get_nsecs();
516    cpu_usage = 0;
517    pthread_mutex_unlock(&work_done_wait_mutex);
518
519    for (i = 0; i < nr_tasks; i++) {
520        task = tasks[i];
521        ret = sem_wait(&task->ready_for_work);
522        BUG_ON(ret);
523        sem_init(&task->ready_for_work, 0, 0);
524    }
525    ret = pthread_mutex_lock(&work_done_wait_mutex);
526    BUG_ON(ret);
527
528    cpu_usage_0 = get_cpu_usage_nsec_parent();
529
530    pthread_mutex_unlock(&start_work_mutex);
531
532    for (i = 0; i < nr_tasks; i++) {
533        task = tasks[i];
534        ret = sem_wait(&task->work_done_sem);
535        BUG_ON(ret);
536        sem_init(&task->work_done_sem, 0, 0);
537        cpu_usage += task->cpu_usage;
538        task->cpu_usage = 0;
539    }
540
541    cpu_usage_1 = get_cpu_usage_nsec_parent();
542    if (!runavg_cpu_usage)
543        runavg_cpu_usage = cpu_usage;
544    runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
545
546    parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
547    if (!runavg_parent_cpu_usage)
548        runavg_parent_cpu_usage = parent_cpu_usage;
549    runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
550                   parent_cpu_usage)/10;
551
552    ret = pthread_mutex_lock(&start_work_mutex);
553    BUG_ON(ret);
554
555    for (i = 0; i < nr_tasks; i++) {
556        task = tasks[i];
557        sem_init(&task->sleep_sem, 0, 0);
558        task->curr_event = 0;
559    }
560}
561
562static void run_one_test(void)
563{
564    u64 T0, T1, delta, avg_delta, fluct, std_dev;
565
566    T0 = get_nsecs();
567    wait_for_tasks();
568    T1 = get_nsecs();
569
570    delta = T1 - T0;
571    sum_runtime += delta;
572    nr_runs++;
573
574    avg_delta = sum_runtime / nr_runs;
575    if (delta < avg_delta)
576        fluct = avg_delta - delta;
577    else
578        fluct = delta - avg_delta;
579    sum_fluct += fluct;
580    std_dev = sum_fluct / nr_runs / sqrt(nr_runs);
581    if (!run_avg)
582        run_avg = delta;
583    run_avg = (run_avg*9 + delta)/10;
584
585    printf("#%-3ld: %0.3f, ",
586        nr_runs, (double)delta/1000000.0);
587
588    printf("ravg: %0.2f, ",
589        (double)run_avg/1e6);
590
591    printf("cpu: %0.2f / %0.2f",
592        (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
593
594#if 0
595    /*
596     * rusage statistics done by the parent, these are less
597     * accurate than the sum_exec_runtime based statistics:
598     */
599    printf(" [%0.2f / %0.2f]",
600        (double)parent_cpu_usage/1e6,
601        (double)runavg_parent_cpu_usage/1e6);
602#endif
603
604    printf("\n");
605
606    if (nr_sleep_corrections)
607        printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
608    nr_sleep_corrections = 0;
609}
610
611static void test_calibrations(void)
612{
613    u64 T0, T1;
614
615    T0 = get_nsecs();
616    burn_nsecs(1e6);
617    T1 = get_nsecs();
618
619    printf("the run test took %Ld nsecs\n", T1-T0);
620
621    T0 = get_nsecs();
622    sleep_nsecs(1e6);
623    T1 = get_nsecs();
624
625    printf("the sleep test took %Ld nsecs\n", T1-T0);
626}
627
628#define FILL_FIELD(ptr, field, event, data) \
629    ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
630
631#define FILL_ARRAY(ptr, array, event, data) \
632do { \
633    void *__array = raw_field_ptr(event, #array, data); \
634    memcpy(ptr.array, __array, sizeof(ptr.array)); \
635} while(0)
636
637#define FILL_COMMON_FIELDS(ptr, event, data) \
638do { \
639    FILL_FIELD(ptr, common_type, event, data); \
640    FILL_FIELD(ptr, common_flags, event, data); \
641    FILL_FIELD(ptr, common_preempt_count, event, data); \
642    FILL_FIELD(ptr, common_pid, event, data); \
643    FILL_FIELD(ptr, common_tgid, event, data); \
644} while (0)
645
646
647
648struct trace_switch_event {
649    u32 size;
650
651    u16 common_type;
652    u8 common_flags;
653    u8 common_preempt_count;
654    u32 common_pid;
655    u32 common_tgid;
656
657    char prev_comm[16];
658    u32 prev_pid;
659    u32 prev_prio;
660    u64 prev_state;
661    char next_comm[16];
662    u32 next_pid;
663    u32 next_prio;
664};
665
666struct trace_runtime_event {
667    u32 size;
668
669    u16 common_type;
670    u8 common_flags;
671    u8 common_preempt_count;
672    u32 common_pid;
673    u32 common_tgid;
674
675    char comm[16];
676    u32 pid;
677    u64 runtime;
678    u64 vruntime;
679};
680
681struct trace_wakeup_event {
682    u32 size;
683
684    u16 common_type;
685    u8 common_flags;
686    u8 common_preempt_count;
687    u32 common_pid;
688    u32 common_tgid;
689
690    char comm[16];
691    u32 pid;
692
693    u32 prio;
694    u32 success;
695    u32 cpu;
696};
697
698struct trace_fork_event {
699    u32 size;
700
701    u16 common_type;
702    u8 common_flags;
703    u8 common_preempt_count;
704    u32 common_pid;
705    u32 common_tgid;
706
707    char parent_comm[16];
708    u32 parent_pid;
709    char child_comm[16];
710    u32 child_pid;
711};
712
713struct trace_migrate_task_event {
714    u32 size;
715
716    u16 common_type;
717    u8 common_flags;
718    u8 common_preempt_count;
719    u32 common_pid;
720    u32 common_tgid;
721
722    char comm[16];
723    u32 pid;
724
725    u32 prio;
726    u32 cpu;
727};
728
729struct trace_sched_handler {
730    void (*switch_event)(struct trace_switch_event *,
731                 struct perf_session *,
732                 struct event *,
733                 int cpu,
734                 u64 timestamp,
735                 struct thread *thread);
736
737    void (*runtime_event)(struct trace_runtime_event *,
738                  struct perf_session *,
739                  struct event *,
740                  int cpu,
741                  u64 timestamp,
742                  struct thread *thread);
743
744    void (*wakeup_event)(struct trace_wakeup_event *,
745                 struct perf_session *,
746                 struct event *,
747                 int cpu,
748                 u64 timestamp,
749                 struct thread *thread);
750
751    void (*fork_event)(struct trace_fork_event *,
752               struct event *,
753               int cpu,
754               u64 timestamp,
755               struct thread *thread);
756
757    void (*migrate_task_event)(struct trace_migrate_task_event *,
758               struct perf_session *session,
759               struct event *,
760               int cpu,
761               u64 timestamp,
762               struct thread *thread);
763};
764
765
766static void
767replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
768            struct perf_session *session __used,
769            struct event *event,
770            int cpu __used,
771            u64 timestamp __used,
772            struct thread *thread __used)
773{
774    struct task_desc *waker, *wakee;
775
776    if (verbose) {
777        printf("sched_wakeup event %p\n", event);
778
779        printf(" ... pid %d woke up %s/%d\n",
780            wakeup_event->common_pid,
781            wakeup_event->comm,
782            wakeup_event->pid);
783    }
784
785    waker = register_pid(wakeup_event->common_pid, "<unknown>");
786    wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
787
788    add_sched_event_wakeup(waker, timestamp, wakee);
789}
790
791static u64 cpu_last_switched[MAX_CPUS];
792
793static void
794replay_switch_event(struct trace_switch_event *switch_event,
795            struct perf_session *session __used,
796            struct event *event,
797            int cpu,
798            u64 timestamp,
799            struct thread *thread __used)
800{
801    struct task_desc *prev, *next;
802    u64 timestamp0;
803    s64 delta;
804
805    if (verbose)
806        printf("sched_switch event %p\n", event);
807
808    if (cpu >= MAX_CPUS || cpu < 0)
809        return;
810
811    timestamp0 = cpu_last_switched[cpu];
812    if (timestamp0)
813        delta = timestamp - timestamp0;
814    else
815        delta = 0;
816
817    if (delta < 0)
818        die("hm, delta: %Ld < 0 ?\n", delta);
819
820    if (verbose) {
821        printf(" ... switch from %s/%d to %s/%d [ran %Ld nsecs]\n",
822            switch_event->prev_comm, switch_event->prev_pid,
823            switch_event->next_comm, switch_event->next_pid,
824            delta);
825    }
826
827    prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
828    next = register_pid(switch_event->next_pid, switch_event->next_comm);
829
830    cpu_last_switched[cpu] = timestamp;
831
832    add_sched_event_run(prev, timestamp, delta);
833    add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
834}
835
836
837static void
838replay_fork_event(struct trace_fork_event *fork_event,
839          struct event *event,
840          int cpu __used,
841          u64 timestamp __used,
842          struct thread *thread __used)
843{
844    if (verbose) {
845        printf("sched_fork event %p\n", event);
846        printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
847        printf("... child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
848    }
849    register_pid(fork_event->parent_pid, fork_event->parent_comm);
850    register_pid(fork_event->child_pid, fork_event->child_comm);
851}
852
853static struct trace_sched_handler replay_ops = {
854    .wakeup_event = replay_wakeup_event,
855    .switch_event = replay_switch_event,
856    .fork_event = replay_fork_event,
857};
858
859struct sort_dimension {
860    const char *name;
861    sort_fn_t cmp;
862    struct list_head list;
863};
864
865static LIST_HEAD(cmp_pid);
866
867static int
868thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
869{
870    struct sort_dimension *sort;
871    int ret = 0;
872
873    BUG_ON(list_empty(list));
874
875    list_for_each_entry(sort, list, list) {
876        ret = sort->cmp(l, r);
877        if (ret)
878            return ret;
879    }
880
881    return ret;
882}
883
884static struct work_atoms *
885thread_atoms_search(struct rb_root *root, struct thread *thread,
886             struct list_head *sort_list)
887{
888    struct rb_node *node = root->rb_node;
889    struct work_atoms key = { .thread = thread };
890
891    while (node) {
892        struct work_atoms *atoms;
893        int cmp;
894
895        atoms = container_of(node, struct work_atoms, node);
896
897        cmp = thread_lat_cmp(sort_list, &key, atoms);
898        if (cmp > 0)
899            node = node->rb_left;
900        else if (cmp < 0)
901            node = node->rb_right;
902        else {
903            BUG_ON(thread != atoms->thread);
904            return atoms;
905        }
906    }
907    return NULL;
908}
909
910static void
911__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
912             struct list_head *sort_list)
913{
914    struct rb_node **new = &(root->rb_node), *parent = NULL;
915
916    while (*new) {
917        struct work_atoms *this;
918        int cmp;
919
920        this = container_of(*new, struct work_atoms, node);
921        parent = *new;
922
923        cmp = thread_lat_cmp(sort_list, data, this);
924
925        if (cmp > 0)
926            new = &((*new)->rb_left);
927        else
928            new = &((*new)->rb_right);
929    }
930
931    rb_link_node(&data->node, parent, new);
932    rb_insert_color(&data->node, root);
933}
934
935static void thread_atoms_insert(struct thread *thread)
936{
937    struct work_atoms *atoms = zalloc(sizeof(*atoms));
938    if (!atoms)
939        die("No memory");
940
941    atoms->thread = thread;
942    INIT_LIST_HEAD(&atoms->work_list);
943    __thread_latency_insert(&atom_root, atoms, &cmp_pid);
944}
945
946static void
947latency_fork_event(struct trace_fork_event *fork_event __used,
948           struct event *event __used,
949           int cpu __used,
950           u64 timestamp __used,
951           struct thread *thread __used)
952{
953    /* should insert the newcomer */
954}
955
956__used
957static char sched_out_state(struct trace_switch_event *switch_event)
958{
959    const char *str = TASK_STATE_TO_CHAR_STR;
960
961    return str[switch_event->prev_state];
962}
963
964static void
965add_sched_out_event(struct work_atoms *atoms,
966            char run_state,
967            u64 timestamp)
968{
969    struct work_atom *atom = zalloc(sizeof(*atom));
970    if (!atom)
971        die("Non memory");
972
973    atom->sched_out_time = timestamp;
974
975    if (run_state == 'R') {
976        atom->state = THREAD_WAIT_CPU;
977        atom->wake_up_time = atom->sched_out_time;
978    }
979
980    list_add_tail(&atom->list, &atoms->work_list);
981}
982
983static void
984add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
985{
986    struct work_atom *atom;
987
988    BUG_ON(list_empty(&atoms->work_list));
989
990    atom = list_entry(atoms->work_list.prev, struct work_atom, list);
991
992    atom->runtime += delta;
993    atoms->total_runtime += delta;
994}
995
996static void
997add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
998{
999    struct work_atom *atom;
1000    u64 delta;
1001
1002    if (list_empty(&atoms->work_list))
1003        return;
1004
1005    atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1006
1007    if (atom->state != THREAD_WAIT_CPU)
1008        return;
1009
1010    if (timestamp < atom->wake_up_time) {
1011        atom->state = THREAD_IGNORE;
1012        return;
1013    }
1014
1015    atom->state = THREAD_SCHED_IN;
1016    atom->sched_in_time = timestamp;
1017
1018    delta = atom->sched_in_time - atom->wake_up_time;
1019    atoms->total_lat += delta;
1020    if (delta > atoms->max_lat) {
1021        atoms->max_lat = delta;
1022        atoms->max_lat_at = timestamp;
1023    }
1024    atoms->nb_atoms++;
1025}
1026
1027static void
1028latency_switch_event(struct trace_switch_event *switch_event,
1029             struct perf_session *session,
1030             struct event *event __used,
1031             int cpu,
1032             u64 timestamp,
1033             struct thread *thread __used)
1034{
1035    struct work_atoms *out_events, *in_events;
1036    struct thread *sched_out, *sched_in;
1037    u64 timestamp0;
1038    s64 delta;
1039
1040    BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1041
1042    timestamp0 = cpu_last_switched[cpu];
1043    cpu_last_switched[cpu] = timestamp;
1044    if (timestamp0)
1045        delta = timestamp - timestamp0;
1046    else
1047        delta = 0;
1048
1049    if (delta < 0)
1050        die("hm, delta: %Ld < 0 ?\n", delta);
1051
1052
1053    sched_out = perf_session__findnew(session, switch_event->prev_pid);
1054    sched_in = perf_session__findnew(session, switch_event->next_pid);
1055
1056    out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1057    if (!out_events) {
1058        thread_atoms_insert(sched_out);
1059        out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
1060        if (!out_events)
1061            die("out-event: Internal tree error");
1062    }
1063    add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
1064
1065    in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1066    if (!in_events) {
1067        thread_atoms_insert(sched_in);
1068        in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
1069        if (!in_events)
1070            die("in-event: Internal tree error");
1071        /*
1072         * Take came in we have not heard about yet,
1073         * add in an initial atom in runnable state:
1074         */
1075        add_sched_out_event(in_events, 'R', timestamp);
1076    }
1077    add_sched_in_event(in_events, timestamp);
1078}
1079
1080static void
1081latency_runtime_event(struct trace_runtime_event *runtime_event,
1082             struct perf_session *session,
1083             struct event *event __used,
1084             int cpu,
1085             u64 timestamp,
1086             struct thread *this_thread __used)
1087{
1088    struct thread *thread = perf_session__findnew(session, runtime_event->pid);
1089    struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1090
1091    BUG_ON(cpu >= MAX_CPUS || cpu < 0);
1092    if (!atoms) {
1093        thread_atoms_insert(thread);
1094        atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
1095        if (!atoms)
1096            die("in-event: Internal tree error");
1097        add_sched_out_event(atoms, 'R', timestamp);
1098    }
1099
1100    add_runtime_event(atoms, runtime_event->runtime, timestamp);
1101}
1102
1103static void
1104latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
1105             struct perf_session *session,
1106             struct event *__event __used,
1107             int cpu __used,
1108             u64 timestamp,
1109             struct thread *thread __used)
1110{
1111    struct work_atoms *atoms;
1112    struct work_atom *atom;
1113    struct thread *wakee;
1114
1115    /* Note for later, it may be interesting to observe the failing cases */
1116    if (!wakeup_event->success)
1117        return;
1118
1119    wakee = perf_session__findnew(session, wakeup_event->pid);
1120    atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1121    if (!atoms) {
1122        thread_atoms_insert(wakee);
1123        atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
1124        if (!atoms)
1125            die("wakeup-event: Internal tree error");
1126        add_sched_out_event(atoms, 'S', timestamp);
1127    }
1128
1129    BUG_ON(list_empty(&atoms->work_list));
1130
1131    atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1132
1133    /*
1134     * You WILL be missing events if you've recorded only
1135     * one CPU, or are only looking at only one, so don't
1136     * make useless noise.
1137     */
1138    if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
1139        nr_state_machine_bugs++;
1140
1141    nr_timestamps++;
1142    if (atom->sched_out_time > timestamp) {
1143        nr_unordered_timestamps++;
1144        return;
1145    }
1146
1147    atom->state = THREAD_WAIT_CPU;
1148    atom->wake_up_time = timestamp;
1149}
1150
1151static void
1152latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
1153             struct perf_session *session,
1154             struct event *__event __used,
1155             int cpu __used,
1156             u64 timestamp,
1157             struct thread *thread __used)
1158{
1159    struct work_atoms *atoms;
1160    struct work_atom *atom;
1161    struct thread *migrant;
1162
1163    /*
1164     * Only need to worry about migration when profiling one CPU.
1165     */
1166    if (profile_cpu == -1)
1167        return;
1168
1169    migrant = perf_session__findnew(session, migrate_task_event->pid);
1170    atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1171    if (!atoms) {
1172        thread_atoms_insert(migrant);
1173        register_pid(migrant->pid, migrant->comm);
1174        atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
1175        if (!atoms)
1176            die("migration-event: Internal tree error");
1177        add_sched_out_event(atoms, 'R', timestamp);
1178    }
1179
1180    BUG_ON(list_empty(&atoms->work_list));
1181
1182    atom = list_entry(atoms->work_list.prev, struct work_atom, list);
1183    atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
1184
1185    nr_timestamps++;
1186
1187    if (atom->sched_out_time > timestamp)
1188        nr_unordered_timestamps++;
1189}
1190
1191static struct trace_sched_handler lat_ops = {
1192    .wakeup_event = latency_wakeup_event,
1193    .switch_event = latency_switch_event,
1194    .runtime_event = latency_runtime_event,
1195    .fork_event = latency_fork_event,
1196    .migrate_task_event = latency_migrate_task_event,
1197};
1198
1199static void output_lat_thread(struct work_atoms *work_list)
1200{
1201    int i;
1202    int ret;
1203    u64 avg;
1204
1205    if (!work_list->nb_atoms)
1206        return;
1207    /*
1208     * Ignore idle threads:
1209     */
1210    if (!strcmp(work_list->thread->comm, "swapper"))
1211        return;
1212
1213    all_runtime += work_list->total_runtime;
1214    all_count += work_list->nb_atoms;
1215
1216    ret = printf(" %s:%d ", work_list->thread->comm, work_list->thread->pid);
1217
1218    for (i = 0; i < 24 - ret; i++)
1219        printf(" ");
1220
1221    avg = work_list->total_lat / work_list->nb_atoms;
1222
1223    printf("|%11.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
1224          (double)work_list->total_runtime / 1e6,
1225         work_list->nb_atoms, (double)avg / 1e6,
1226         (double)work_list->max_lat / 1e6,
1227         (double)work_list->max_lat_at / 1e9);
1228}
1229
1230static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
1231{
1232    if (l->thread->pid < r->thread->pid)
1233        return -1;
1234    if (l->thread->pid > r->thread->pid)
1235        return 1;
1236
1237    return 0;
1238}
1239
1240static struct sort_dimension pid_sort_dimension = {
1241    .name = "pid",
1242    .cmp = pid_cmp,
1243};
1244
1245static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
1246{
1247    u64 avgl, avgr;
1248
1249    if (!l->nb_atoms)
1250        return -1;
1251
1252    if (!r->nb_atoms)
1253        return 1;
1254
1255    avgl = l->total_lat / l->nb_atoms;
1256    avgr = r->total_lat / r->nb_atoms;
1257
1258    if (avgl < avgr)
1259        return -1;
1260    if (avgl > avgr)
1261        return 1;
1262
1263    return 0;
1264}
1265
1266static struct sort_dimension avg_sort_dimension = {
1267    .name = "avg",
1268    .cmp = avg_cmp,
1269};
1270
1271static int max_cmp(struct work_atoms *l, struct work_atoms *r)
1272{
1273    if (l->max_lat < r->max_lat)
1274        return -1;
1275    if (l->max_lat > r->max_lat)
1276        return 1;
1277
1278    return 0;
1279}
1280
1281static struct sort_dimension max_sort_dimension = {
1282    .name = "max",
1283    .cmp = max_cmp,
1284};
1285
1286static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
1287{
1288    if (l->nb_atoms < r->nb_atoms)
1289        return -1;
1290    if (l->nb_atoms > r->nb_atoms)
1291        return 1;
1292
1293    return 0;
1294}
1295
1296static struct sort_dimension switch_sort_dimension = {
1297    .name = "switch",
1298    .cmp = switch_cmp,
1299};
1300
1301static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
1302{
1303    if (l->total_runtime < r->total_runtime)
1304        return -1;
1305    if (l->total_runtime > r->total_runtime)
1306        return 1;
1307
1308    return 0;
1309}
1310
1311static struct sort_dimension runtime_sort_dimension = {
1312    .name = "runtime",
1313    .cmp = runtime_cmp,
1314};
1315
1316static struct sort_dimension *available_sorts[] = {
1317    &pid_sort_dimension,
1318    &avg_sort_dimension,
1319    &max_sort_dimension,
1320    &switch_sort_dimension,
1321    &runtime_sort_dimension,
1322};
1323
1324#define NB_AVAILABLE_SORTS (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
1325
1326static LIST_HEAD(sort_list);
1327
1328static int sort_dimension__add(const char *tok, struct list_head *list)
1329{
1330    int i;
1331
1332    for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
1333        if (!strcmp(available_sorts[i]->name, tok)) {
1334            list_add_tail(&available_sorts[i]->list, list);
1335
1336            return 0;
1337        }
1338    }
1339
1340    return -1;
1341}
1342
1343static void setup_sorting(void);
1344
1345static void sort_lat(void)
1346{
1347    struct rb_node *node;
1348
1349    for (;;) {
1350        struct work_atoms *data;
1351        node = rb_first(&atom_root);
1352        if (!node)
1353            break;
1354
1355        rb_erase(node, &atom_root);
1356        data = rb_entry(node, struct work_atoms, node);
1357        __thread_latency_insert(&sorted_atom_root, data, &sort_list);
1358    }
1359}
1360
1361static struct trace_sched_handler *trace_handler;
1362
1363static void
1364process_sched_wakeup_event(void *data, struct perf_session *session,
1365               struct event *event,
1366               int cpu __used,
1367               u64 timestamp __used,
1368               struct thread *thread __used)
1369{
1370    struct trace_wakeup_event wakeup_event;
1371
1372    FILL_COMMON_FIELDS(wakeup_event, event, data);
1373
1374    FILL_ARRAY(wakeup_event, comm, event, data);
1375    FILL_FIELD(wakeup_event, pid, event, data);
1376    FILL_FIELD(wakeup_event, prio, event, data);
1377    FILL_FIELD(wakeup_event, success, event, data);
1378    FILL_FIELD(wakeup_event, cpu, event, data);
1379
1380    if (trace_handler->wakeup_event)
1381        trace_handler->wakeup_event(&wakeup_event, session, event,
1382                        cpu, timestamp, thread);
1383}
1384
1385/*
1386 * Track the current task - that way we can know whether there's any
1387 * weird events, such as a task being switched away that is not current.
1388 */
1389static int max_cpu;
1390
1391static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
1392
1393static struct thread *curr_thread[MAX_CPUS];
1394
1395static char next_shortname1 = 'A';
1396static char next_shortname2 = '0';
1397
1398static void
1399map_switch_event(struct trace_switch_event *switch_event,
1400         struct perf_session *session,
1401         struct event *event __used,
1402         int this_cpu,
1403         u64 timestamp,
1404         struct thread *thread __used)
1405{
1406    struct thread *sched_out, *sched_in;
1407    int new_shortname;
1408    u64 timestamp0;
1409    s64 delta;
1410    int cpu;
1411
1412    BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
1413
1414    if (this_cpu > max_cpu)
1415        max_cpu = this_cpu;
1416
1417    timestamp0 = cpu_last_switched[this_cpu];
1418    cpu_last_switched[this_cpu] = timestamp;
1419    if (timestamp0)
1420        delta = timestamp - timestamp0;
1421    else
1422        delta = 0;
1423
1424    if (delta < 0)
1425        die("hm, delta: %Ld < 0 ?\n", delta);
1426
1427
1428    sched_out = perf_session__findnew(session, switch_event->prev_pid);
1429    sched_in = perf_session__findnew(session, switch_event->next_pid);
1430
1431    curr_thread[this_cpu] = sched_in;
1432
1433    printf(" ");
1434
1435    new_shortname = 0;
1436    if (!sched_in->shortname[0]) {
1437        sched_in->shortname[0] = next_shortname1;
1438        sched_in->shortname[1] = next_shortname2;
1439
1440        if (next_shortname1 < 'Z') {
1441            next_shortname1++;
1442        } else {
1443            next_shortname1='A';
1444            if (next_shortname2 < '9') {
1445                next_shortname2++;
1446            } else {
1447                next_shortname2='0';
1448            }
1449        }
1450        new_shortname = 1;
1451    }
1452
1453    for (cpu = 0; cpu <= max_cpu; cpu++) {
1454        if (cpu != this_cpu)
1455            printf(" ");
1456        else
1457            printf("*");
1458
1459        if (curr_thread[cpu]) {
1460            if (curr_thread[cpu]->pid)
1461                printf("%2s ", curr_thread[cpu]->shortname);
1462            else
1463                printf(". ");
1464        } else
1465            printf(" ");
1466    }
1467
1468    printf(" %12.6f secs ", (double)timestamp/1e9);
1469    if (new_shortname) {
1470        printf("%s => %s:%d\n",
1471            sched_in->shortname, sched_in->comm, sched_in->pid);
1472    } else {
1473        printf("\n");
1474    }
1475}
1476
1477
1478static void
1479process_sched_switch_event(void *data, struct perf_session *session,
1480               struct event *event,
1481               int this_cpu,
1482               u64 timestamp __used,
1483               struct thread *thread __used)
1484{
1485    struct trace_switch_event switch_event;
1486
1487    FILL_COMMON_FIELDS(switch_event, event, data);
1488
1489    FILL_ARRAY(switch_event, prev_comm, event, data);
1490    FILL_FIELD(switch_event, prev_pid, event, data);
1491    FILL_FIELD(switch_event, prev_prio, event, data);
1492    FILL_FIELD(switch_event, prev_state, event, data);
1493    FILL_ARRAY(switch_event, next_comm, event, data);
1494    FILL_FIELD(switch_event, next_pid, event, data);
1495    FILL_FIELD(switch_event, next_prio, event, data);
1496
1497    if (curr_pid[this_cpu] != (u32)-1) {
1498        /*
1499         * Are we trying to switch away a PID that is
1500         * not current?
1501         */
1502        if (curr_pid[this_cpu] != switch_event.prev_pid)
1503            nr_context_switch_bugs++;
1504    }
1505    if (trace_handler->switch_event)
1506        trace_handler->switch_event(&switch_event, session, event,
1507                        this_cpu, timestamp, thread);
1508
1509    curr_pid[this_cpu] = switch_event.next_pid;
1510}
1511
1512static void
1513process_sched_runtime_event(void *data, struct perf_session *session,
1514               struct event *event,
1515               int cpu __used,
1516               u64 timestamp __used,
1517               struct thread *thread __used)
1518{
1519    struct trace_runtime_event runtime_event;
1520
1521    FILL_ARRAY(runtime_event, comm, event, data);
1522    FILL_FIELD(runtime_event, pid, event, data);
1523    FILL_FIELD(runtime_event, runtime, event, data);
1524    FILL_FIELD(runtime_event, vruntime, event, data);
1525
1526    if (trace_handler->runtime_event)
1527        trace_handler->runtime_event(&runtime_event, session, event, cpu, timestamp, thread);
1528}
1529
1530static void
1531process_sched_fork_event(void *data,
1532             struct event *event,
1533             int cpu __used,
1534             u64 timestamp __used,
1535             struct thread *thread __used)
1536{
1537    struct trace_fork_event fork_event;
1538
1539    FILL_COMMON_FIELDS(fork_event, event, data);
1540
1541    FILL_ARRAY(fork_event, parent_comm, event, data);
1542    FILL_FIELD(fork_event, parent_pid, event, data);
1543    FILL_ARRAY(fork_event, child_comm, event, data);
1544    FILL_FIELD(fork_event, child_pid, event, data);
1545
1546    if (trace_handler->fork_event)
1547        trace_handler->fork_event(&fork_event, event,
1548                      cpu, timestamp, thread);
1549}
1550
1551static void
1552process_sched_exit_event(struct event *event,
1553             int cpu __used,
1554             u64 timestamp __used,
1555             struct thread *thread __used)
1556{
1557    if (verbose)
1558        printf("sched_exit event %p\n", event);
1559}
1560
1561static void
1562process_sched_migrate_task_event(void *data, struct perf_session *session,
1563               struct event *event,
1564               int cpu __used,
1565               u64 timestamp __used,
1566               struct thread *thread __used)
1567{
1568    struct trace_migrate_task_event migrate_task_event;
1569
1570    FILL_COMMON_FIELDS(migrate_task_event, event, data);
1571
1572    FILL_ARRAY(migrate_task_event, comm, event, data);
1573    FILL_FIELD(migrate_task_event, pid, event, data);
1574    FILL_FIELD(migrate_task_event, prio, event, data);
1575    FILL_FIELD(migrate_task_event, cpu, event, data);
1576
1577    if (trace_handler->migrate_task_event)
1578        trace_handler->migrate_task_event(&migrate_task_event, session,
1579                         event, cpu, timestamp, thread);
1580}
1581
1582static void
1583process_raw_event(event_t *raw_event __used, struct perf_session *session,
1584          void *data, int cpu, u64 timestamp, struct thread *thread)
1585{
1586    struct event *event;
1587    int type;
1588
1589
1590    type = trace_parse_common_type(data);
1591    event = trace_find_event(type);
1592
1593    if (!strcmp(event->name, "sched_switch"))
1594        process_sched_switch_event(data, session, event, cpu, timestamp, thread);
1595    if (!strcmp(event->name, "sched_stat_runtime"))
1596        process_sched_runtime_event(data, session, event, cpu, timestamp, thread);
1597    if (!strcmp(event->name, "sched_wakeup"))
1598        process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1599    if (!strcmp(event->name, "sched_wakeup_new"))
1600        process_sched_wakeup_event(data, session, event, cpu, timestamp, thread);
1601    if (!strcmp(event->name, "sched_process_fork"))
1602        process_sched_fork_event(data, event, cpu, timestamp, thread);
1603    if (!strcmp(event->name, "sched_process_exit"))
1604        process_sched_exit_event(event, cpu, timestamp, thread);
1605    if (!strcmp(event->name, "sched_migrate_task"))
1606        process_sched_migrate_task_event(data, session, event, cpu, timestamp, thread);
1607}
1608
1609static int process_sample_event(event_t *event, struct perf_session *session)
1610{
1611    struct sample_data data;
1612    struct thread *thread;
1613
1614    if (!(session->sample_type & PERF_SAMPLE_RAW))
1615        return 0;
1616
1617    memset(&data, 0, sizeof(data));
1618    data.time = -1;
1619    data.cpu = -1;
1620    data.period = -1;
1621
1622    event__parse_sample(event, session->sample_type, &data);
1623
1624    dump_printf("(IP, %d): %d/%d: %#Lx period: %Ld\n", event->header.misc,
1625            data.pid, data.tid, data.ip, data.period);
1626
1627    thread = perf_session__findnew(session, data.pid);
1628    if (thread == NULL) {
1629        pr_debug("problem processing %d event, skipping it.\n",
1630             event->header.type);
1631        return -1;
1632    }
1633
1634    dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
1635
1636    if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
1637        return 0;
1638
1639    process_raw_event(event, session, data.raw_data, data.cpu, data.time, thread);
1640
1641    return 0;
1642}
1643
1644static int process_lost_event(event_t *event __used,
1645                  struct perf_session *session __used)
1646{
1647    nr_lost_chunks++;
1648    nr_lost_events += event->lost.lost;
1649
1650    return 0;
1651}
1652
1653static struct perf_event_ops event_ops = {
1654    .sample = process_sample_event,
1655    .comm = event__process_comm,
1656    .lost = process_lost_event,
1657};
1658
1659static int read_events(void)
1660{
1661    int err = -EINVAL;
1662    struct perf_session *session = perf_session__new(input_name, O_RDONLY, 0);
1663    if (session == NULL)
1664        return -ENOMEM;
1665
1666    if (perf_session__has_traces(session, "record -R"))
1667        err = perf_session__process_events(session, &event_ops);
1668
1669    perf_session__delete(session);
1670    return err;
1671}
1672
1673static void print_bad_events(void)
1674{
1675    if (nr_unordered_timestamps && nr_timestamps) {
1676        printf(" INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
1677            (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
1678            nr_unordered_timestamps, nr_timestamps);
1679    }
1680    if (nr_lost_events && nr_events) {
1681        printf(" INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
1682            (double)nr_lost_events/(double)nr_events*100.0,
1683            nr_lost_events, nr_events, nr_lost_chunks);
1684    }
1685    if (nr_state_machine_bugs && nr_timestamps) {
1686        printf(" INFO: %.3f%% state machine bugs (%ld out of %ld)",
1687            (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
1688            nr_state_machine_bugs, nr_timestamps);
1689        if (nr_lost_events)
1690            printf(" (due to lost events?)");
1691        printf("\n");
1692    }
1693    if (nr_context_switch_bugs && nr_timestamps) {
1694        printf(" INFO: %.3f%% context switch bugs (%ld out of %ld)",
1695            (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
1696            nr_context_switch_bugs, nr_timestamps);
1697        if (nr_lost_events)
1698            printf(" (due to lost events?)");
1699        printf("\n");
1700    }
1701}
1702
1703static void __cmd_lat(void)
1704{
1705    struct rb_node *next;
1706
1707    setup_pager();
1708    read_events();
1709    sort_lat();
1710
1711    printf("\n ---------------------------------------------------------------------------------------------------------------\n");
1712    printf(" Task | Runtime ms | Switches | Average delay ms | Maximum delay ms | Maximum delay at |\n");
1713    printf(" ---------------------------------------------------------------------------------------------------------------\n");
1714
1715    next = rb_first(&sorted_atom_root);
1716
1717    while (next) {
1718        struct work_atoms *work_list;
1719
1720        work_list = rb_entry(next, struct work_atoms, node);
1721        output_lat_thread(work_list);
1722        next = rb_next(next);
1723    }
1724
1725    printf(" -----------------------------------------------------------------------------------------\n");
1726    printf(" TOTAL: |%11.3f ms |%9Ld |\n",
1727        (double)all_runtime/1e6, all_count);
1728
1729    printf(" ---------------------------------------------------\n");
1730
1731    print_bad_events();
1732    printf("\n");
1733
1734}
1735
1736static struct trace_sched_handler map_ops = {
1737    .wakeup_event = NULL,
1738    .switch_event = map_switch_event,
1739    .runtime_event = NULL,
1740    .fork_event = NULL,
1741};
1742
1743static void __cmd_map(void)
1744{
1745    max_cpu = sysconf(_SC_NPROCESSORS_CONF);
1746
1747    setup_pager();
1748    read_events();
1749    print_bad_events();
1750}
1751
1752static void __cmd_replay(void)
1753{
1754    unsigned long i;
1755
1756    calibrate_run_measurement_overhead();
1757    calibrate_sleep_measurement_overhead();
1758
1759    test_calibrations();
1760
1761    read_events();
1762
1763    printf("nr_run_events: %ld\n", nr_run_events);
1764    printf("nr_sleep_events: %ld\n", nr_sleep_events);
1765    printf("nr_wakeup_events: %ld\n", nr_wakeup_events);
1766
1767    if (targetless_wakeups)
1768        printf("target-less wakeups: %ld\n", targetless_wakeups);
1769    if (multitarget_wakeups)
1770        printf("multi-target wakeups: %ld\n", multitarget_wakeups);
1771    if (nr_run_events_optimized)
1772        printf("run atoms optimized: %ld\n",
1773            nr_run_events_optimized);
1774
1775    print_task_traces();
1776    add_cross_task_wakeups();
1777
1778    create_tasks();
1779    printf("------------------------------------------------------------\n");
1780    for (i = 0; i < replay_repeat; i++)
1781        run_one_test();
1782}
1783
1784
1785static const char * const sched_usage[] = {
1786    "perf sched [<options>] {record|latency|map|replay|trace}",
1787    NULL
1788};
1789
1790static const struct option sched_options[] = {
1791    OPT_STRING('i', "input", &input_name, "file",
1792            "input file name"),
1793    OPT_BOOLEAN('v', "verbose", &verbose,
1794            "be more verbose (show symbol address, etc)"),
1795    OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1796            "dump raw trace in ASCII"),
1797    OPT_END()
1798};
1799
1800static const char * const latency_usage[] = {
1801    "perf sched latency [<options>]",
1802    NULL
1803};
1804
1805static const struct option latency_options[] = {
1806    OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1807           "sort by key(s): runtime, switch, avg, max"),
1808    OPT_BOOLEAN('v', "verbose", &verbose,
1809            "be more verbose (show symbol address, etc)"),
1810    OPT_INTEGER('C', "CPU", &profile_cpu,
1811            "CPU to profile on"),
1812    OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1813            "dump raw trace in ASCII"),
1814    OPT_END()
1815};
1816
1817static const char * const replay_usage[] = {
1818    "perf sched replay [<options>]",
1819    NULL
1820};
1821
1822static const struct option replay_options[] = {
1823    OPT_INTEGER('r', "repeat", &replay_repeat,
1824            "repeat the workload replay N times (-1: infinite)"),
1825    OPT_BOOLEAN('v', "verbose", &verbose,
1826            "be more verbose (show symbol address, etc)"),
1827    OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1828            "dump raw trace in ASCII"),
1829    OPT_END()
1830};
1831
1832static void setup_sorting(void)
1833{
1834    char *tmp, *tok, *str = strdup(sort_order);
1835
1836    for (tok = strtok_r(str, ", ", &tmp);
1837            tok; tok = strtok_r(NULL, ", ", &tmp)) {
1838        if (sort_dimension__add(tok, &sort_list) < 0) {
1839            error("Unknown --sort key: `%s'", tok);
1840            usage_with_options(latency_usage, latency_options);
1841        }
1842    }
1843
1844    free(str);
1845
1846    sort_dimension__add("pid", &cmp_pid);
1847}
1848
1849static const char *record_args[] = {
1850    "record",
1851    "-a",
1852    "-R",
1853    "-M",
1854    "-f",
1855    "-m", "1024",
1856    "-c", "1",
1857    "-e", "sched:sched_switch:r",
1858    "-e", "sched:sched_stat_wait:r",
1859    "-e", "sched:sched_stat_sleep:r",
1860    "-e", "sched:sched_stat_iowait:r",
1861    "-e", "sched:sched_stat_runtime:r",
1862    "-e", "sched:sched_process_exit:r",
1863    "-e", "sched:sched_process_fork:r",
1864    "-e", "sched:sched_wakeup:r",
1865    "-e", "sched:sched_migrate_task:r",
1866};
1867
1868static int __cmd_record(int argc, const char **argv)
1869{
1870    unsigned int rec_argc, i, j;
1871    const char **rec_argv;
1872
1873    rec_argc = ARRAY_SIZE(record_args) + argc - 1;
1874    rec_argv = calloc(rec_argc + 1, sizeof(char *));
1875
1876    for (i = 0; i < ARRAY_SIZE(record_args); i++)
1877        rec_argv[i] = strdup(record_args[i]);
1878
1879    for (j = 1; j < (unsigned int)argc; j++, i++)
1880        rec_argv[i] = argv[j];
1881
1882    BUG_ON(i != rec_argc);
1883
1884    return cmd_record(i, rec_argv, NULL);
1885}
1886
1887int cmd_sched(int argc, const char **argv, const char *prefix __used)
1888{
1889    argc = parse_options(argc, argv, sched_options, sched_usage,
1890                 PARSE_OPT_STOP_AT_NON_OPTION);
1891    if (!argc)
1892        usage_with_options(sched_usage, sched_options);
1893
1894    /*
1895     * Aliased to 'perf trace' for now:
1896     */
1897    if (!strcmp(argv[0], "trace"))
1898        return cmd_trace(argc, argv, prefix);
1899
1900    symbol__init();
1901    if (!strncmp(argv[0], "rec", 3)) {
1902        return __cmd_record(argc, argv);
1903    } else if (!strncmp(argv[0], "lat", 3)) {
1904        trace_handler = &lat_ops;
1905        if (argc > 1) {
1906            argc = parse_options(argc, argv, latency_options, latency_usage, 0);
1907            if (argc)
1908                usage_with_options(latency_usage, latency_options);
1909        }
1910        setup_sorting();
1911        __cmd_lat();
1912    } else if (!strcmp(argv[0], "map")) {
1913        trace_handler = &map_ops;
1914        setup_sorting();
1915        __cmd_map();
1916    } else if (!strncmp(argv[0], "rep", 3)) {
1917        trace_handler = &replay_ops;
1918        if (argc) {
1919            argc = parse_options(argc, argv, replay_options, replay_usage, 0);
1920            if (argc)
1921                usage_with_options(replay_usage, replay_options);
1922        }
1923        __cmd_replay();
1924    } else {
1925        usage_with_options(sched_usage, sched_options);
1926    }
1927
1928    return 0;
1929}
1930

Archive Download this file



interactive