perf tools: Save some loops using perf_evlist__id2evsel
[linux-3.10.git] / tools / perf / builtin-sched.c
index 60228d9..6284ed2 100644 (file)
 #include "builtin.h"
+#include "perf.h"
 
 #include "util/util.h"
+#include "util/evlist.h"
 #include "util/cache.h"
+#include "util/evsel.h"
 #include "util/symbol.h"
 #include "util/thread.h"
 #include "util/header.h"
+#include "util/session.h"
+#include "util/tool.h"
 
 #include "util/parse-options.h"
+#include "util/trace-event.h"
 
-#include "perf.h"
 #include "util/debug.h"
 
-#include "util/trace-event.h"
+#include <sys/prctl.h>
 
-static char            const *input_name = "perf.data";
-static int             input;
-static unsigned long   page_size;
-static unsigned long   mmap_window = 32;
+#include <semaphore.h>
+#include <pthread.h>
+#include <math.h>
 
-static unsigned long   total = 0;
-static unsigned long   total_comm = 0;
+static char                    const *input_name = "perf.data";
 
-static struct rb_root  threads;
-static struct thread   *last_match;
+static char                    default_sort_order[] = "avg, max, switch, runtime";
+static const char              *sort_order = default_sort_order;
 
-static struct perf_header *header;
-static u64             sample_type;
+static int                     profile_cpu = -1;
 
+#define PR_SET_NAME            15               /* Set process name */
+#define MAX_CPUS               4096
 
-static int
-process_comm_event(event_t *event, unsigned long offset, unsigned long head)
+static u64                     run_measurement_overhead;
+static u64                     sleep_measurement_overhead;
+
+#define COMM_LEN               20
+#define SYM_LEN                        129
+
+#define MAX_PID                        65536
+
+static unsigned long           nr_tasks;
+
+struct sched_atom;
+
+struct task_desc {
+       unsigned long           nr;
+       unsigned long           pid;
+       char                    comm[COMM_LEN];
+
+       unsigned long           nr_events;
+       unsigned long           curr_event;
+       struct sched_atom       **atoms;
+
+       pthread_t               thread;
+       sem_t                   sleep_sem;
+
+       sem_t                   ready_for_work;
+       sem_t                   work_done_sem;
+
+       u64                     cpu_usage;
+};
+
+enum sched_event_type {
+       SCHED_EVENT_RUN,
+       SCHED_EVENT_SLEEP,
+       SCHED_EVENT_WAKEUP,
+       SCHED_EVENT_MIGRATION,
+};
+
+struct sched_atom {
+       enum sched_event_type   type;
+       int                     specific_wait;
+       u64                     timestamp;
+       u64                     duration;
+       unsigned long           nr;
+       sem_t                   *wait_sem;
+       struct task_desc        *wakee;
+};
+
+static struct task_desc                *pid_to_task[MAX_PID];
+
+static struct task_desc                **tasks;
+
+static pthread_mutex_t         start_work_mutex = PTHREAD_MUTEX_INITIALIZER;
+static u64                     start_time;
+
+static pthread_mutex_t         work_done_wait_mutex = PTHREAD_MUTEX_INITIALIZER;
+
+static unsigned long           nr_run_events;
+static unsigned long           nr_sleep_events;
+static unsigned long           nr_wakeup_events;
+
+static unsigned long           nr_sleep_corrections;
+static unsigned long           nr_run_events_optimized;
+
+static unsigned long           targetless_wakeups;
+static unsigned long           multitarget_wakeups;
+
+static u64                     cpu_usage;
+static u64                     runavg_cpu_usage;
+static u64                     parent_cpu_usage;
+static u64                     runavg_parent_cpu_usage;
+
+static unsigned long           nr_runs;
+static u64                     sum_runtime;
+static u64                     sum_fluct;
+static u64                     run_avg;
+
+static unsigned int            replay_repeat = 10;
+static unsigned long           nr_timestamps;
+static unsigned long           nr_unordered_timestamps;
+static unsigned long           nr_state_machine_bugs;
+static unsigned long           nr_context_switch_bugs;
+static unsigned long           nr_events;
+static unsigned long           nr_lost_chunks;
+static unsigned long           nr_lost_events;
+
+#define TASK_STATE_TO_CHAR_STR "RSDTtZX"
+
+enum thread_state {
+       THREAD_SLEEPING = 0,
+       THREAD_WAIT_CPU,
+       THREAD_SCHED_IN,
+       THREAD_IGNORE
+};
+
+struct work_atom {
+       struct list_head        list;
+       enum thread_state       state;
+       u64                     sched_out_time;
+       u64                     wake_up_time;
+       u64                     sched_in_time;
+       u64                     runtime;
+};
+
+struct work_atoms {
+       struct list_head        work_list;
+       struct thread           *thread;
+       struct rb_node          node;
+       u64                     max_lat;
+       u64                     max_lat_at;
+       u64                     total_lat;
+       u64                     nb_atoms;
+       u64                     total_runtime;
+};
+
+typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
+
+static struct rb_root          atom_root, sorted_atom_root;
+
+static u64                     all_runtime;
+static u64                     all_count;
+
+
+static u64 get_nsecs(void)
 {
-       struct thread *thread;
+       struct timespec ts;
 
-       thread = threads__findnew(event->comm.pid, &threads, &last_match);
+       clock_gettime(CLOCK_MONOTONIC, &ts);
 
-       dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
-               (void *)(offset + head),
-               (void *)(long)(event->header.size),
-               event->comm.comm, event->comm.pid);
+       return ts.tv_sec * 1000000000ULL + ts.tv_nsec;
+}
 
-       if (thread == NULL ||
-           thread__set_comm(thread, event->comm.comm)) {
-               dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
-               return -1;
+static void burn_nsecs(u64 nsecs)
+{
+       u64 T0 = get_nsecs(), T1;
+
+       do {
+               T1 = get_nsecs();
+       } while (T1 + run_measurement_overhead < T0 + nsecs);
+}
+
+static void sleep_nsecs(u64 nsecs)
+{
+       struct timespec ts;
+
+       ts.tv_nsec = nsecs % 999999999;
+       ts.tv_sec = nsecs / 999999999;
+
+       nanosleep(&ts, NULL);
+}
+
+static void calibrate_run_measurement_overhead(void)
+{
+       u64 T0, T1, delta, min_delta = 1000000000ULL;
+       int i;
+
+       for (i = 0; i < 10; i++) {
+               T0 = get_nsecs();
+               burn_nsecs(0);
+               T1 = get_nsecs();
+               delta = T1-T0;
+               min_delta = min(min_delta, delta);
        }
-       total_comm++;
+       run_measurement_overhead = min_delta;
 
-       return 0;
+       printf("run measurement overhead: %" PRIu64 " nsecs\n", min_delta);
+}
+
+static void calibrate_sleep_measurement_overhead(void)
+{
+       u64 T0, T1, delta, min_delta = 1000000000ULL;
+       int i;
+
+       for (i = 0; i < 10; i++) {
+               T0 = get_nsecs();
+               sleep_nsecs(10000);
+               T1 = get_nsecs();
+               delta = T1-T0;
+               min_delta = min(min_delta, delta);
+       }
+       min_delta -= 10000;
+       sleep_measurement_overhead = min_delta;
+
+       printf("sleep measurement overhead: %" PRIu64 " nsecs\n", min_delta);
+}
+
+static struct sched_atom *
+get_new_event(struct task_desc *task, u64 timestamp)
+{
+       struct sched_atom *event = zalloc(sizeof(*event));
+       unsigned long idx = task->nr_events;
+       size_t size;
+
+       event->timestamp = timestamp;
+       event->nr = idx;
+
+       task->nr_events++;
+       size = sizeof(struct sched_atom *) * task->nr_events;
+       task->atoms = realloc(task->atoms, size);
+       BUG_ON(!task->atoms);
+
+       task->atoms[idx] = event;
+
+       return event;
+}
+
+static struct sched_atom *last_event(struct task_desc *task)
+{
+       if (!task->nr_events)
+               return NULL;
+
+       return task->atoms[task->nr_events - 1];
+}
+
+static void
+add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
+{
+       struct sched_atom *event, *curr_event = last_event(task);
+
+       /*
+        * optimize an existing RUN event by merging this one
+        * to it:
+        */
+       if (curr_event && curr_event->type == SCHED_EVENT_RUN) {
+               nr_run_events_optimized++;
+               curr_event->duration += duration;
+               return;
+       }
+
+       event = get_new_event(task, timestamp);
+
+       event->type = SCHED_EVENT_RUN;
+       event->duration = duration;
+
+       nr_run_events++;
+}
+
+static void
+add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
+                      struct task_desc *wakee)
+{
+       struct sched_atom *event, *wakee_event;
+
+       event = get_new_event(task, timestamp);
+       event->type = SCHED_EVENT_WAKEUP;
+       event->wakee = wakee;
+
+       wakee_event = last_event(wakee);
+       if (!wakee_event || wakee_event->type != SCHED_EVENT_SLEEP) {
+               targetless_wakeups++;
+               return;
+       }
+       if (wakee_event->wait_sem) {
+               multitarget_wakeups++;
+               return;
+       }
+
+       wakee_event->wait_sem = zalloc(sizeof(*wakee_event->wait_sem));
+       sem_init(wakee_event->wait_sem, 0, 0);
+       wakee_event->specific_wait = 1;
+       event->wait_sem = wakee_event->wait_sem;
+
+       nr_wakeup_events++;
+}
+
+static void
+add_sched_event_sleep(struct task_desc *task, u64 timestamp,
+                     u64 task_state __used)
+{
+       struct sched_atom *event = get_new_event(task, timestamp);
+
+       event->type = SCHED_EVENT_SLEEP;
+
+       nr_sleep_events++;
+}
+
+static struct task_desc *register_pid(unsigned long pid, const char *comm)
+{
+       struct task_desc *task;
+
+       BUG_ON(pid >= MAX_PID);
+
+       task = pid_to_task[pid];
+
+       if (task)
+               return task;
+
+       task = zalloc(sizeof(*task));
+       task->pid = pid;
+       task->nr = nr_tasks;
+       strcpy(task->comm, comm);
+       /*
+        * every task starts in sleeping state - this gets ignored
+        * if there's no wakeup pointing to this sleep state:
+        */
+       add_sched_event_sleep(task, 0, 0);
+
+       pid_to_task[pid] = task;
+       nr_tasks++;
+       tasks = realloc(tasks, nr_tasks*sizeof(struct task_task *));
+       BUG_ON(!tasks);
+       tasks[task->nr] = task;
+
+       if (verbose)
+               printf("registered task #%ld, PID %ld (%s)\n", nr_tasks, pid, comm);
+
+       return task;
+}
+
+
+static void print_task_traces(void)
+{
+       struct task_desc *task;
+       unsigned long i;
+
+       for (i = 0; i < nr_tasks; i++) {
+               task = tasks[i];
+               printf("task %6ld (%20s:%10ld), nr_events: %ld\n",
+                       task->nr, task->comm, task->pid, task->nr_events);
+       }
+}
+
+static void add_cross_task_wakeups(void)
+{
+       struct task_desc *task1, *task2;
+       unsigned long i, j;
+
+       for (i = 0; i < nr_tasks; i++) {
+               task1 = tasks[i];
+               j = i + 1;
+               if (j == nr_tasks)
+                       j = 0;
+               task2 = tasks[j];
+               add_sched_event_wakeup(task1, 0, task2);
+       }
+}
+
+static void
+process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
+{
+       int ret = 0;
+
+       switch (atom->type) {
+               case SCHED_EVENT_RUN:
+                       burn_nsecs(atom->duration);
+                       break;
+               case SCHED_EVENT_SLEEP:
+                       if (atom->wait_sem)
+                               ret = sem_wait(atom->wait_sem);
+                       BUG_ON(ret);
+                       break;
+               case SCHED_EVENT_WAKEUP:
+                       if (atom->wait_sem)
+                               ret = sem_post(atom->wait_sem);
+                       BUG_ON(ret);
+                       break;
+               case SCHED_EVENT_MIGRATION:
+                       break;
+               default:
+                       BUG_ON(1);
+       }
+}
+
+static u64 get_cpu_usage_nsec_parent(void)
+{
+       struct rusage ru;
+       u64 sum;
+       int err;
+
+       err = getrusage(RUSAGE_SELF, &ru);
+       BUG_ON(err);
+
+       sum =  ru.ru_utime.tv_sec*1e9 + ru.ru_utime.tv_usec*1e3;
+       sum += ru.ru_stime.tv_sec*1e9 + ru.ru_stime.tv_usec*1e3;
+
+       return sum;
+}
+
+static int self_open_counters(void)
+{
+       struct perf_event_attr attr;
+       int fd;
+
+       memset(&attr, 0, sizeof(attr));
+
+       attr.type = PERF_TYPE_SOFTWARE;
+       attr.config = PERF_COUNT_SW_TASK_CLOCK;
+
+       fd = sys_perf_event_open(&attr, 0, -1, -1, 0);
+
+       if (fd < 0)
+               die("Error: sys_perf_event_open() syscall returned"
+                   "with %d (%s)\n", fd, strerror(errno));
+       return fd;
+}
+
+static u64 get_cpu_usage_nsec_self(int fd)
+{
+       u64 runtime;
+       int ret;
+
+       ret = read(fd, &runtime, sizeof(runtime));
+       BUG_ON(ret != sizeof(runtime));
+
+       return runtime;
+}
+
+static void *thread_func(void *ctx)
+{
+       struct task_desc *this_task = ctx;
+       u64 cpu_usage_0, cpu_usage_1;
+       unsigned long i, ret;
+       char comm2[22];
+       int fd;
+
+       sprintf(comm2, ":%s", this_task->comm);
+       prctl(PR_SET_NAME, comm2);
+       fd = self_open_counters();
+
+again:
+       ret = sem_post(&this_task->ready_for_work);
+       BUG_ON(ret);
+       ret = pthread_mutex_lock(&start_work_mutex);
+       BUG_ON(ret);
+       ret = pthread_mutex_unlock(&start_work_mutex);
+       BUG_ON(ret);
+
+       cpu_usage_0 = get_cpu_usage_nsec_self(fd);
+
+       for (i = 0; i < this_task->nr_events; i++) {
+               this_task->curr_event = i;
+               process_sched_event(this_task, this_task->atoms[i]);
+       }
+
+       cpu_usage_1 = get_cpu_usage_nsec_self(fd);
+       this_task->cpu_usage = cpu_usage_1 - cpu_usage_0;
+       ret = sem_post(&this_task->work_done_sem);
+       BUG_ON(ret);
+
+       ret = pthread_mutex_lock(&work_done_wait_mutex);
+       BUG_ON(ret);
+       ret = pthread_mutex_unlock(&work_done_wait_mutex);
+       BUG_ON(ret);
+
+       goto again;
+}
+
+static void create_tasks(void)
+{
+       struct task_desc *task;
+       pthread_attr_t attr;
+       unsigned long i;
+       int err;
+
+       err = pthread_attr_init(&attr);
+       BUG_ON(err);
+       err = pthread_attr_setstacksize(&attr,
+                       (size_t) max(16 * 1024, PTHREAD_STACK_MIN));
+       BUG_ON(err);
+       err = pthread_mutex_lock(&start_work_mutex);
+       BUG_ON(err);
+       err = pthread_mutex_lock(&work_done_wait_mutex);
+       BUG_ON(err);
+       for (i = 0; i < nr_tasks; i++) {
+               task = tasks[i];
+               sem_init(&task->sleep_sem, 0, 0);
+               sem_init(&task->ready_for_work, 0, 0);
+               sem_init(&task->work_done_sem, 0, 0);
+               task->curr_event = 0;
+               err = pthread_create(&task->thread, &attr, thread_func, task);
+               BUG_ON(err);
+       }
+}
+
+static void wait_for_tasks(void)
+{
+       u64 cpu_usage_0, cpu_usage_1;
+       struct task_desc *task;
+       unsigned long i, ret;
+
+       start_time = get_nsecs();
+       cpu_usage = 0;
+       pthread_mutex_unlock(&work_done_wait_mutex);
+
+       for (i = 0; i < nr_tasks; i++) {
+               task = tasks[i];
+               ret = sem_wait(&task->ready_for_work);
+               BUG_ON(ret);
+               sem_init(&task->ready_for_work, 0, 0);
+       }
+       ret = pthread_mutex_lock(&work_done_wait_mutex);
+       BUG_ON(ret);
+
+       cpu_usage_0 = get_cpu_usage_nsec_parent();
+
+       pthread_mutex_unlock(&start_work_mutex);
+
+       for (i = 0; i < nr_tasks; i++) {
+               task = tasks[i];
+               ret = sem_wait(&task->work_done_sem);
+               BUG_ON(ret);
+               sem_init(&task->work_done_sem, 0, 0);
+               cpu_usage += task->cpu_usage;
+               task->cpu_usage = 0;
+       }
+
+       cpu_usage_1 = get_cpu_usage_nsec_parent();
+       if (!runavg_cpu_usage)
+               runavg_cpu_usage = cpu_usage;
+       runavg_cpu_usage = (runavg_cpu_usage*9 + cpu_usage)/10;
+
+       parent_cpu_usage = cpu_usage_1 - cpu_usage_0;
+       if (!runavg_parent_cpu_usage)
+               runavg_parent_cpu_usage = parent_cpu_usage;
+       runavg_parent_cpu_usage = (runavg_parent_cpu_usage*9 +
+                                  parent_cpu_usage)/10;
+
+       ret = pthread_mutex_lock(&start_work_mutex);
+       BUG_ON(ret);
+
+       for (i = 0; i < nr_tasks; i++) {
+               task = tasks[i];
+               sem_init(&task->sleep_sem, 0, 0);
+               task->curr_event = 0;
+       }
+}
+
+static void run_one_test(void)
+{
+       u64 T0, T1, delta, avg_delta, fluct;
+
+       T0 = get_nsecs();
+       wait_for_tasks();
+       T1 = get_nsecs();
+
+       delta = T1 - T0;
+       sum_runtime += delta;
+       nr_runs++;
+
+       avg_delta = sum_runtime / nr_runs;
+       if (delta < avg_delta)
+               fluct = avg_delta - delta;
+       else
+               fluct = delta - avg_delta;
+       sum_fluct += fluct;
+       if (!run_avg)
+               run_avg = delta;
+       run_avg = (run_avg*9 + delta)/10;
+
+       printf("#%-3ld: %0.3f, ",
+               nr_runs, (double)delta/1000000.0);
+
+       printf("ravg: %0.2f, ",
+               (double)run_avg/1e6);
+
+       printf("cpu: %0.2f / %0.2f",
+               (double)cpu_usage/1e6, (double)runavg_cpu_usage/1e6);
+
+#if 0
+       /*
+        * rusage statistics done by the parent, these are less
+        * accurate than the sum_exec_runtime based statistics:
+        */
+       printf(" [%0.2f / %0.2f]",
+               (double)parent_cpu_usage/1e6,
+               (double)runavg_parent_cpu_usage/1e6);
+#endif
+
+       printf("\n");
+
+       if (nr_sleep_corrections)
+               printf(" (%ld sleep corrections)\n", nr_sleep_corrections);
+       nr_sleep_corrections = 0;
+}
+
+static void test_calibrations(void)
+{
+       u64 T0, T1;
+
+       T0 = get_nsecs();
+       burn_nsecs(1e6);
+       T1 = get_nsecs();
+
+       printf("the run test took %" PRIu64 " nsecs\n", T1 - T0);
+
+       T0 = get_nsecs();
+       sleep_nsecs(1e6);
+       T1 = get_nsecs();
+
+       printf("the sleep test took %" PRIu64 " nsecs\n", T1 - T0);
+}
+
+#define FILL_FIELD(ptr, field, event, data)    \
+       ptr.field = (typeof(ptr.field)) raw_field_value(event, #field, data)
+
+#define FILL_ARRAY(ptr, array, event, data)                    \
+do {                                                           \
+       void *__array = raw_field_ptr(event, #array, data);     \
+       memcpy(ptr.array, __array, sizeof(ptr.array));  \
+} while(0)
+
+#define FILL_COMMON_FIELDS(ptr, event, data)                   \
+do {                                                           \
+       FILL_FIELD(ptr, common_type, event, data);              \
+       FILL_FIELD(ptr, common_flags, event, data);             \
+       FILL_FIELD(ptr, common_preempt_count, event, data);     \
+       FILL_FIELD(ptr, common_pid, event, data);               \
+       FILL_FIELD(ptr, common_tgid, event, data);              \
+} while (0)
+
+
+
+struct trace_switch_event {
+       u32 size;
+
+       u16 common_type;
+       u8 common_flags;
+       u8 common_preempt_count;
+       u32 common_pid;
+       u32 common_tgid;
+
+       char prev_comm[16];
+       u32 prev_pid;
+       u32 prev_prio;
+       u64 prev_state;
+       char next_comm[16];
+       u32 next_pid;
+       u32 next_prio;
+};
+
+struct trace_runtime_event {
+       u32 size;
+
+       u16 common_type;
+       u8 common_flags;
+       u8 common_preempt_count;
+       u32 common_pid;
+       u32 common_tgid;
+
+       char comm[16];
+       u32 pid;
+       u64 runtime;
+       u64 vruntime;
+};
+
+struct trace_wakeup_event {
+       u32 size;
+
+       u16 common_type;
+       u8 common_flags;
+       u8 common_preempt_count;
+       u32 common_pid;
+       u32 common_tgid;
+
+       char comm[16];
+       u32 pid;
+
+       u32 prio;
+       u32 success;
+       u32 cpu;
+};
+
+struct trace_fork_event {
+       u32 size;
+
+       u16 common_type;
+       u8 common_flags;
+       u8 common_preempt_count;
+       u32 common_pid;
+       u32 common_tgid;
+
+       char parent_comm[16];
+       u32 parent_pid;
+       char child_comm[16];
+       u32 child_pid;
+};
+
+struct trace_migrate_task_event {
+       u32 size;
+
+       u16 common_type;
+       u8 common_flags;
+       u8 common_preempt_count;
+       u32 common_pid;
+       u32 common_tgid;
+
+       char comm[16];
+       u32 pid;
+
+       u32 prio;
+       u32 cpu;
+};
+
+struct trace_sched_handler {
+       void (*switch_event)(struct trace_switch_event *,
+                            struct machine *,
+                            struct event *,
+                            int cpu,
+                            u64 timestamp,
+                            struct thread *thread);
+
+       void (*runtime_event)(struct trace_runtime_event *,
+                             struct machine *,
+                             struct event *,
+                             int cpu,
+                             u64 timestamp,
+                             struct thread *thread);
+
+       void (*wakeup_event)(struct trace_wakeup_event *,
+                            struct machine *,
+                            struct event *,
+                            int cpu,
+                            u64 timestamp,
+                            struct thread *thread);
+
+       void (*fork_event)(struct trace_fork_event *,
+                          struct event *,
+                          int cpu,
+                          u64 timestamp,
+                          struct thread *thread);
+
+       void (*migrate_task_event)(struct trace_migrate_task_event *,
+                          struct machine *machine,
+                          struct event *,
+                          int cpu,
+                          u64 timestamp,
+                          struct thread *thread);
+};
+
+
+static void
+replay_wakeup_event(struct trace_wakeup_event *wakeup_event,
+                   struct machine *machine __used,
+                   struct event *event,
+                   int cpu __used,
+                   u64 timestamp __used,
+                   struct thread *thread __used)
+{
+       struct task_desc *waker, *wakee;
+
+       if (verbose) {
+               printf("sched_wakeup event %p\n", event);
+
+               printf(" ... pid %d woke up %s/%d\n",
+                       wakeup_event->common_pid,
+                       wakeup_event->comm,
+                       wakeup_event->pid);
+       }
+
+       waker = register_pid(wakeup_event->common_pid, "<unknown>");
+       wakee = register_pid(wakeup_event->pid, wakeup_event->comm);
+
+       add_sched_event_wakeup(waker, timestamp, wakee);
 }
 
+static u64 cpu_last_switched[MAX_CPUS];
+
+static void
+replay_switch_event(struct trace_switch_event *switch_event,
+                   struct machine *machine __used,
+                   struct event *event,
+                   int cpu,
+                   u64 timestamp,
+                   struct thread *thread __used)
+{
+       struct task_desc *prev, __used *next;
+       u64 timestamp0;
+       s64 delta;
+
+       if (verbose)
+               printf("sched_switch event %p\n", event);
+
+       if (cpu >= MAX_CPUS || cpu < 0)
+               return;
+
+       timestamp0 = cpu_last_switched[cpu];
+       if (timestamp0)
+               delta = timestamp - timestamp0;
+       else
+               delta = 0;
+
+       if (delta < 0)
+               die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
+
+       if (verbose) {
+               printf(" ... switch from %s/%d to %s/%d [ran %" PRIu64 " nsecs]\n",
+                       switch_event->prev_comm, switch_event->prev_pid,
+                       switch_event->next_comm, switch_event->next_pid,
+                       delta);
+       }
+
+       prev = register_pid(switch_event->prev_pid, switch_event->prev_comm);
+       next = register_pid(switch_event->next_pid, switch_event->next_comm);
+
+       cpu_last_switched[cpu] = timestamp;
+
+       add_sched_event_run(prev, timestamp, delta);
+       add_sched_event_sleep(prev, timestamp, switch_event->prev_state);
+}
+
+
+static void
+replay_fork_event(struct trace_fork_event *fork_event,
+                 struct event *event,
+                 int cpu __used,
+                 u64 timestamp __used,
+                 struct thread *thread __used)
+{
+       if (verbose) {
+               printf("sched_fork event %p\n", event);
+               printf("... parent: %s/%d\n", fork_event->parent_comm, fork_event->parent_pid);
+               printf("...  child: %s/%d\n", fork_event->child_comm, fork_event->child_pid);
+       }
+       register_pid(fork_event->parent_pid, fork_event->parent_comm);
+       register_pid(fork_event->child_pid, fork_event->child_comm);
+}
+
+static struct trace_sched_handler replay_ops  = {
+       .wakeup_event           = replay_wakeup_event,
+       .switch_event           = replay_switch_event,
+       .fork_event             = replay_fork_event,
+};
+
+struct sort_dimension {
+       const char              *name;
+       sort_fn_t               cmp;
+       struct list_head        list;
+};
+
+static LIST_HEAD(cmp_pid);
+
 static int
-process_sample_event(event_t *event, unsigned long offset, unsigned long head)
+thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 {
-       char level;
-       int show = 0;
-       struct dso *dso = NULL;
-       struct thread *thread;
-       u64 ip = event->ip.ip;
-       u64 timestamp = -1;
-       u32 cpu = -1;
-       u64 period = 1;
-       void *more_data = event->ip.__more_data;
-       int cpumode;
+       struct sort_dimension *sort;
+       int ret = 0;
 
-       thread = threads__findnew(event->ip.pid, &threads, &last_match);
+       BUG_ON(list_empty(list));
 
-       if (sample_type & PERF_SAMPLE_TIME) {
-               timestamp = *(u64 *)more_data;
-               more_data += sizeof(u64);
+       list_for_each_entry(sort, list, list) {
+               ret = sort->cmp(l, r);
+               if (ret)
+                       return ret;
        }
 
-       if (sample_type & PERF_SAMPLE_CPU) {
-               cpu = *(u32 *)more_data;
-               more_data += sizeof(u32);
-               more_data += sizeof(u32); /* reserved */
+       return ret;
+}
+
+static struct work_atoms *
+thread_atoms_search(struct rb_root *root, struct thread *thread,
+                        struct list_head *sort_list)
+{
+       struct rb_node *node = root->rb_node;
+       struct work_atoms key = { .thread = thread };
+
+       while (node) {
+               struct work_atoms *atoms;
+               int cmp;
+
+               atoms = container_of(node, struct work_atoms, node);
+
+               cmp = thread_lat_cmp(sort_list, &key, atoms);
+               if (cmp > 0)
+                       node = node->rb_left;
+               else if (cmp < 0)
+                       node = node->rb_right;
+               else {
+                       BUG_ON(thread != atoms->thread);
+                       return atoms;
+               }
        }
+       return NULL;
+}
+
+static void
+__thread_latency_insert(struct rb_root *root, struct work_atoms *data,
+                        struct list_head *sort_list)
+{
+       struct rb_node **new = &(root->rb_node), *parent = NULL;
 
-       if (sample_type & PERF_SAMPLE_PERIOD) {
-               period = *(u64 *)more_data;
-               more_data += sizeof(u64);
+       while (*new) {
+               struct work_atoms *this;
+               int cmp;
+
+               this = container_of(*new, struct work_atoms, node);
+               parent = *new;
+
+               cmp = thread_lat_cmp(sort_list, data, this);
+
+               if (cmp > 0)
+                       new = &((*new)->rb_left);
+               else
+                       new = &((*new)->rb_right);
        }
 
-       dump_printf("%p [%p]: PERF_EVENT_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
-               (void *)(offset + head),
-               (void *)(long)(event->header.size),
-               event->header.misc,
-               event->ip.pid, event->ip.tid,
-               (void *)(long)ip,
-               (long long)period);
+       rb_link_node(&data->node, parent, new);
+       rb_insert_color(&data->node, root);
+}
 
-       dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
+static void thread_atoms_insert(struct thread *thread)
+{
+       struct work_atoms *atoms = zalloc(sizeof(*atoms));
+       if (!atoms)
+               die("No memory");
 
-       if (thread == NULL) {
-               eprintf("problem processing %d event, skipping it.\n",
-                       event->header.type);
-               return -1;
+       atoms->thread = thread;
+       INIT_LIST_HEAD(&atoms->work_list);
+       __thread_latency_insert(&atom_root, atoms, &cmp_pid);
+}
+
+static void
+latency_fork_event(struct trace_fork_event *fork_event __used,
+                  struct event *event __used,
+                  int cpu __used,
+                  u64 timestamp __used,
+                  struct thread *thread __used)
+{
+       /* should insert the newcomer */
+}
+
+__used
+static char sched_out_state(struct trace_switch_event *switch_event)
+{
+       const char *str = TASK_STATE_TO_CHAR_STR;
+
+       return str[switch_event->prev_state];
+}
+
+static void
+add_sched_out_event(struct work_atoms *atoms,
+                   char run_state,
+                   u64 timestamp)
+{
+       struct work_atom *atom = zalloc(sizeof(*atom));
+       if (!atom)
+               die("Non memory");
+
+       atom->sched_out_time = timestamp;
+
+       if (run_state == 'R') {
+               atom->state = THREAD_WAIT_CPU;
+               atom->wake_up_time = atom->sched_out_time;
        }
 
-       cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
+       list_add_tail(&atom->list, &atoms->work_list);
+}
 
-       if (cpumode == PERF_EVENT_MISC_KERNEL) {
-               show = SHOW_KERNEL;
-               level = 'k';
+static void
+add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
+{
+       struct work_atom *atom;
 
-               dso = kernel_dso;
+       BUG_ON(list_empty(&atoms->work_list));
 
-               dump_printf(" ...... dso: %s\n", dso->name);
+       atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 
-       } else if (cpumode == PERF_EVENT_MISC_USER) {
+       atom->runtime += delta;
+       atoms->total_runtime += delta;
+}
 
-               show = SHOW_USER;
-               level = '.';
+static void
+add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
+{
+       struct work_atom *atom;
+       u64 delta;
 
-       } else {
-               show = SHOW_HV;
-               level = 'H';
+       if (list_empty(&atoms->work_list))
+               return;
 
-               dso = hypervisor_dso;
+       atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 
-               dump_printf(" ...... dso: [hypervisor]\n");
+       if (atom->state != THREAD_WAIT_CPU)
+               return;
+
+       if (timestamp < atom->wake_up_time) {
+               atom->state = THREAD_IGNORE;
+               return;
        }
 
-       if (sample_type & PERF_SAMPLE_RAW) {
-               struct {
-                       u32 size;
-                       char data[0];
-               } *raw = more_data;
+       atom->state = THREAD_SCHED_IN;
+       atom->sched_in_time = timestamp;
+
+       delta = atom->sched_in_time - atom->wake_up_time;
+       atoms->total_lat += delta;
+       if (delta > atoms->max_lat) {
+               atoms->max_lat = delta;
+               atoms->max_lat_at = timestamp;
+       }
+       atoms->nb_atoms++;
+}
+
+static void
+latency_switch_event(struct trace_switch_event *switch_event,
+                    struct machine *machine,
+                    struct event *event __used,
+                    int cpu,
+                    u64 timestamp,
+                    struct thread *thread __used)
+{
+       struct work_atoms *out_events, *in_events;
+       struct thread *sched_out, *sched_in;
+       u64 timestamp0;
+       s64 delta;
+
+       BUG_ON(cpu >= MAX_CPUS || cpu < 0);
+
+       timestamp0 = cpu_last_switched[cpu];
+       cpu_last_switched[cpu] = timestamp;
+       if (timestamp0)
+               delta = timestamp - timestamp0;
+       else
+               delta = 0;
+
+       if (delta < 0)
+               die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
 
+
+       sched_out = machine__findnew_thread(machine, switch_event->prev_pid);
+       sched_in = machine__findnew_thread(machine, switch_event->next_pid);
+
+       out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
+       if (!out_events) {
+               thread_atoms_insert(sched_out);
+               out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
+               if (!out_events)
+                       die("out-event: Internal tree error");
+       }
+       add_sched_out_event(out_events, sched_out_state(switch_event), timestamp);
+
+       in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
+       if (!in_events) {
+               thread_atoms_insert(sched_in);
+               in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
+               if (!in_events)
+                       die("in-event: Internal tree error");
                /*
-                * FIXME: better resolve from pid from the struct trace_entry
-                * field, although it should be the same than this perf
-                * event pid
+                * Take came in we have not heard about yet,
+                * add in an initial atom in runnable state:
                 */
-               print_event(cpu, raw->data, raw->size, timestamp, thread->comm);
+               add_sched_out_event(in_events, 'R', timestamp);
        }
-       total += period;
+       add_sched_in_event(in_events, timestamp);
+}
+
+static void
+latency_runtime_event(struct trace_runtime_event *runtime_event,
+                    struct machine *machine,
+                    struct event *event __used,
+                    int cpu,
+                    u64 timestamp,
+                    struct thread *this_thread __used)
+{
+       struct thread *thread = machine__findnew_thread(machine, runtime_event->pid);
+       struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
+
+       BUG_ON(cpu >= MAX_CPUS || cpu < 0);
+       if (!atoms) {
+               thread_atoms_insert(thread);
+               atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
+               if (!atoms)
+                       die("in-event: Internal tree error");
+               add_sched_out_event(atoms, 'R', timestamp);
+       }
+
+       add_runtime_event(atoms, runtime_event->runtime, timestamp);
+}
+
+static void
+latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
+                    struct machine *machine,
+                    struct event *__event __used,
+                    int cpu __used,
+                    u64 timestamp,
+                    struct thread *thread __used)
+{
+       struct work_atoms *atoms;
+       struct work_atom *atom;
+       struct thread *wakee;
+
+       /* Note for later, it may be interesting to observe the failing cases */
+       if (!wakeup_event->success)
+               return;
+
+       wakee = machine__findnew_thread(machine, wakeup_event->pid);
+       atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
+       if (!atoms) {
+               thread_atoms_insert(wakee);
+               atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
+               if (!atoms)
+                       die("wakeup-event: Internal tree error");
+               add_sched_out_event(atoms, 'S', timestamp);
+       }
+
+       BUG_ON(list_empty(&atoms->work_list));
+
+       atom = list_entry(atoms->work_list.prev, struct work_atom, list);
+
+       /*
+        * You WILL be missing events if you've recorded only
+        * one CPU, or are only looking at only one, so don't
+        * make useless noise.
+        */
+       if (profile_cpu == -1 && atom->state != THREAD_SLEEPING)
+               nr_state_machine_bugs++;
+
+       nr_timestamps++;
+       if (atom->sched_out_time > timestamp) {
+               nr_unordered_timestamps++;
+               return;
+       }
+
+       atom->state = THREAD_WAIT_CPU;
+       atom->wake_up_time = timestamp;
+}
+
+static void
+latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
+                    struct machine *machine,
+                    struct event *__event __used,
+                    int cpu __used,
+                    u64 timestamp,
+                    struct thread *thread __used)
+{
+       struct work_atoms *atoms;
+       struct work_atom *atom;
+       struct thread *migrant;
+
+       /*
+        * Only need to worry about migration when profiling one CPU.
+        */
+       if (profile_cpu == -1)
+               return;
+
+       migrant = machine__findnew_thread(machine, migrate_task_event->pid);
+       atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
+       if (!atoms) {
+               thread_atoms_insert(migrant);
+               register_pid(migrant->pid, migrant->comm);
+               atoms = thread_atoms_search(&atom_root, migrant, &cmp_pid);
+               if (!atoms)
+                       die("migration-event: Internal tree error");
+               add_sched_out_event(atoms, 'R', timestamp);
+       }
+
+       BUG_ON(list_empty(&atoms->work_list));
+
+       atom = list_entry(atoms->work_list.prev, struct work_atom, list);
+       atom->sched_in_time = atom->sched_out_time = atom->wake_up_time = timestamp;
+
+       nr_timestamps++;
+
+       if (atom->sched_out_time > timestamp)
+               nr_unordered_timestamps++;
+}
+
+static struct trace_sched_handler lat_ops  = {
+       .wakeup_event           = latency_wakeup_event,
+       .switch_event           = latency_switch_event,
+       .runtime_event          = latency_runtime_event,
+       .fork_event             = latency_fork_event,
+       .migrate_task_event     = latency_migrate_task_event,
+};
+
+static void output_lat_thread(struct work_atoms *work_list)
+{
+       int i;
+       int ret;
+       u64 avg;
+
+       if (!work_list->nb_atoms)
+               return;
+       /*
+        * Ignore idle threads:
+        */
+       if (!strcmp(work_list->thread->comm, "swapper"))
+               return;
+
+       all_runtime += work_list->total_runtime;
+       all_count += work_list->nb_atoms;
+
+       ret = printf("  %s:%d ", work_list->thread->comm, work_list->thread->pid);
+
+       for (i = 0; i < 24 - ret; i++)
+               printf(" ");
+
+       avg = work_list->total_lat / work_list->nb_atoms;
+
+       printf("|%11.3f ms |%9" PRIu64 " | avg:%9.3f ms | max:%9.3f ms | max at: %9.6f s\n",
+             (double)work_list->total_runtime / 1e6,
+                work_list->nb_atoms, (double)avg / 1e6,
+                (double)work_list->max_lat / 1e6,
+                (double)work_list->max_lat_at / 1e9);
+}
+
+static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
+{
+       if (l->thread->pid < r->thread->pid)
+               return -1;
+       if (l->thread->pid > r->thread->pid)
+               return 1;
 
        return 0;
 }
 
-static int
-process_event(event_t *event, unsigned long offset, unsigned long head)
+static struct sort_dimension pid_sort_dimension = {
+       .name                   = "pid",
+       .cmp                    = pid_cmp,
+};
+
+static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
 {
-       trace_event(event);
+       u64 avgl, avgr;
 
-       switch (event->header.type) {
-       case PERF_EVENT_MMAP ... PERF_EVENT_LOST:
-               return 0;
+       if (!l->nb_atoms)
+               return -1;
 
-       case PERF_EVENT_COMM:
-               return process_comm_event(event, offset, head);
+       if (!r->nb_atoms)
+               return 1;
 
-       case PERF_EVENT_EXIT ... PERF_EVENT_READ:
-               return 0;
+       avgl = l->total_lat / l->nb_atoms;
+       avgr = r->total_lat / r->nb_atoms;
 
-       case PERF_EVENT_SAMPLE:
-               return process_sample_event(event, offset, head);
+       if (avgl < avgr)
+               return -1;
+       if (avgl > avgr)
+               return 1;
 
-       case PERF_EVENT_MAX:
-       default:
+       return 0;
+}
+
+static struct sort_dimension avg_sort_dimension = {
+       .name                   = "avg",
+       .cmp                    = avg_cmp,
+};
+
+static int max_cmp(struct work_atoms *l, struct work_atoms *r)
+{
+       if (l->max_lat < r->max_lat)
                return -1;
-       }
+       if (l->max_lat > r->max_lat)
+               return 1;
+
+       return 0;
+}
+
+static struct sort_dimension max_sort_dimension = {
+       .name                   = "max",
+       .cmp                    = max_cmp,
+};
+
+static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
+{
+       if (l->nb_atoms < r->nb_atoms)
+               return -1;
+       if (l->nb_atoms > r->nb_atoms)
+               return 1;
 
        return 0;
 }
 
-static int __cmd_sched(void)
+static struct sort_dimension switch_sort_dimension = {
+       .name                   = "switch",
+       .cmp                    = switch_cmp,
+};
+
+static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
 {
-       int ret, rc = EXIT_FAILURE;
-       unsigned long offset = 0;
-       unsigned long head = 0;
-       struct stat perf_stat;
-       event_t *event;
-       uint32_t size;
-       char *buf;
+       if (l->total_runtime < r->total_runtime)
+               return -1;
+       if (l->total_runtime > r->total_runtime)
+               return 1;
+
+       return 0;
+}
 
-       trace_report();
-       register_idle_thread(&threads, &last_match);
+static struct sort_dimension runtime_sort_dimension = {
+       .name                   = "runtime",
+       .cmp                    = runtime_cmp,
+};
+
+static struct sort_dimension *available_sorts[] = {
+       &pid_sort_dimension,
+       &avg_sort_dimension,
+       &max_sort_dimension,
+       &switch_sort_dimension,
+       &runtime_sort_dimension,
+};
 
-       input = open(input_name, O_RDONLY);
-       if (input < 0) {
-               perror("failed to open file");
-               exit(-1);
+#define NB_AVAILABLE_SORTS     (int)(sizeof(available_sorts) / sizeof(struct sort_dimension *))
+
+static LIST_HEAD(sort_list);
+
+static int sort_dimension__add(const char *tok, struct list_head *list)
+{
+       int i;
+
+       for (i = 0; i < NB_AVAILABLE_SORTS; i++) {
+               if (!strcmp(available_sorts[i]->name, tok)) {
+                       list_add_tail(&available_sorts[i]->list, list);
+
+                       return 0;
+               }
        }
 
-       ret = fstat(input, &perf_stat);
-       if (ret < 0) {
-               perror("failed to stat file");
-               exit(-1);
+       return -1;
+}
+
+static void setup_sorting(void);
+
+static void sort_lat(void)
+{
+       struct rb_node *node;
+
+       for (;;) {
+               struct work_atoms *data;
+               node = rb_first(&atom_root);
+               if (!node)
+                       break;
+
+               rb_erase(node, &atom_root);
+               data = rb_entry(node, struct work_atoms, node);
+               __thread_latency_insert(&sorted_atom_root, data, &sort_list);
        }
+}
+
+static struct trace_sched_handler *trace_handler;
+
+static void
+process_sched_wakeup_event(struct perf_tool *tool __used,
+                          struct event *event,
+                          struct perf_sample *sample,
+                          struct machine *machine,
+                          struct thread *thread)
+{
+       void *data = sample->raw_data;
+       struct trace_wakeup_event wakeup_event;
 
-       if (!perf_stat.st_size) {
-               fprintf(stderr, "zero-sized file, nothing to do!\n");
-               exit(0);
+       FILL_COMMON_FIELDS(wakeup_event, event, data);
+
+       FILL_ARRAY(wakeup_event, comm, event, data);
+       FILL_FIELD(wakeup_event, pid, event, data);
+       FILL_FIELD(wakeup_event, prio, event, data);
+       FILL_FIELD(wakeup_event, success, event, data);
+       FILL_FIELD(wakeup_event, cpu, event, data);
+
+       if (trace_handler->wakeup_event)
+               trace_handler->wakeup_event(&wakeup_event, machine, event,
+                                           sample->cpu, sample->time, thread);
+}
+
+/*
+ * Track the current task - that way we can know whether there's any
+ * weird events, such as a task being switched away that is not current.
+ */
+static int max_cpu;
+
+static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
+
+static struct thread *curr_thread[MAX_CPUS];
+
+static char next_shortname1 = 'A';
+static char next_shortname2 = '0';
+
+static void
+map_switch_event(struct trace_switch_event *switch_event,
+                struct machine *machine,
+                struct event *event __used,
+                int this_cpu,
+                u64 timestamp,
+                struct thread *thread __used)
+{
+       struct thread *sched_out __used, *sched_in;
+       int new_shortname;
+       u64 timestamp0;
+       s64 delta;
+       int cpu;
+
+       BUG_ON(this_cpu >= MAX_CPUS || this_cpu < 0);
+
+       if (this_cpu > max_cpu)
+               max_cpu = this_cpu;
+
+       timestamp0 = cpu_last_switched[this_cpu];
+       cpu_last_switched[this_cpu] = timestamp;
+       if (timestamp0)
+               delta = timestamp - timestamp0;
+       else
+               delta = 0;
+
+       if (delta < 0)
+               die("hm, delta: %" PRIu64 " < 0 ?\n", delta);
+
+
+       sched_out = machine__findnew_thread(machine, switch_event->prev_pid);
+       sched_in = machine__findnew_thread(machine, switch_event->next_pid);
+
+       curr_thread[this_cpu] = sched_in;
+
+       printf("  ");
+
+       new_shortname = 0;
+       if (!sched_in->shortname[0]) {
+               sched_in->shortname[0] = next_shortname1;
+               sched_in->shortname[1] = next_shortname2;
+
+               if (next_shortname1 < 'Z') {
+                       next_shortname1++;
+               } else {
+                       next_shortname1='A';
+                       if (next_shortname2 < '9') {
+                               next_shortname2++;
+                       } else {
+                               next_shortname2='0';
+                       }
+               }
+               new_shortname = 1;
        }
-       header = perf_header__read(input);
-       head = header->data_offset;
-       sample_type = perf_header__sample_type(header);
 
-       if (!(sample_type & PERF_SAMPLE_RAW))
-               die("No trace sample to read. Did you call perf record "
-                   "without -R?");
+       for (cpu = 0; cpu <= max_cpu; cpu++) {
+               if (cpu != this_cpu)
+                       printf(" ");
+               else
+                       printf("*");
+
+               if (curr_thread[cpu]) {
+                       if (curr_thread[cpu]->pid)
+                               printf("%2s ", curr_thread[cpu]->shortname);
+                       else
+                               printf(".  ");
+               } else
+                       printf("   ");
+       }
 
-       if (load_kernel() < 0) {
-               perror("failed to load kernel symbols");
-               return EXIT_FAILURE;
+       printf("  %12.6f secs ", (double)timestamp/1e9);
+       if (new_shortname) {
+               printf("%s => %s:%d\n",
+                       sched_in->shortname, sched_in->comm, sched_in->pid);
+       } else {
+               printf("\n");
        }
+}
+
+static void
+process_sched_switch_event(struct perf_tool *tool __used,
+                          struct event *event,
+                          struct perf_sample *sample,
+                          struct machine *machine,
+                          struct thread *thread)
+{
+       int this_cpu = sample->cpu;
+       void *data = sample->raw_data;
+       struct trace_switch_event switch_event;
 
-remap:
-       buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
-                          MAP_SHARED, input, offset);
-       if (buf == MAP_FAILED) {
-               perror("failed to mmap file");
-               exit(-1);
+       FILL_COMMON_FIELDS(switch_event, event, data);
+
+       FILL_ARRAY(switch_event, prev_comm, event, data);
+       FILL_FIELD(switch_event, prev_pid, event, data);
+       FILL_FIELD(switch_event, prev_prio, event, data);
+       FILL_FIELD(switch_event, prev_state, event, data);
+       FILL_ARRAY(switch_event, next_comm, event, data);
+       FILL_FIELD(switch_event, next_pid, event, data);
+       FILL_FIELD(switch_event, next_prio, event, data);
+
+       if (curr_pid[this_cpu] != (u32)-1) {
+               /*
+                * Are we trying to switch away a PID that is
+                * not current?
+                */
+               if (curr_pid[this_cpu] != switch_event.prev_pid)
+                       nr_context_switch_bugs++;
        }
+       if (trace_handler->switch_event)
+               trace_handler->switch_event(&switch_event, machine, event,
+                                           this_cpu, sample->time, thread);
+
+       curr_pid[this_cpu] = switch_event.next_pid;
+}
+
+static void
+process_sched_runtime_event(struct perf_tool *tool __used,
+                           struct event *event,
+                           struct perf_sample *sample,
+                           struct machine *machine,
+                           struct thread *thread)
+{
+       void *data = sample->raw_data;
+       struct trace_runtime_event runtime_event;
+
+       FILL_ARRAY(runtime_event, comm, event, data);
+       FILL_FIELD(runtime_event, pid, event, data);
+       FILL_FIELD(runtime_event, runtime, event, data);
+       FILL_FIELD(runtime_event, vruntime, event, data);
+
+       if (trace_handler->runtime_event)
+               trace_handler->runtime_event(&runtime_event, machine, event,
+                                            sample->cpu, sample->time, thread);
+}
+
+static void
+process_sched_fork_event(struct perf_tool *tool __used,
+                        struct event *event,
+                        struct perf_sample *sample,
+                        struct machine *machine __used,
+                        struct thread *thread)
+{
+       void *data = sample->raw_data;
+       struct trace_fork_event fork_event;
+
+       FILL_COMMON_FIELDS(fork_event, event, data);
+
+       FILL_ARRAY(fork_event, parent_comm, event, data);
+       FILL_FIELD(fork_event, parent_pid, event, data);
+       FILL_ARRAY(fork_event, child_comm, event, data);
+       FILL_FIELD(fork_event, child_pid, event, data);
 
-more:
-       event = (event_t *)(buf + head);
+       if (trace_handler->fork_event)
+               trace_handler->fork_event(&fork_event, event,
+                                         sample->cpu, sample->time, thread);
+}
+
+static void
+process_sched_exit_event(struct perf_tool *tool __used,
+                        struct event *event,
+                        struct perf_sample *sample __used,
+                        struct machine *machine __used,
+                        struct thread *thread __used)
+{
+       if (verbose)
+               printf("sched_exit event %p\n", event);
+}
 
-       size = event->header.size;
-       if (!size)
-               size = 8;
+static void
+process_sched_migrate_task_event(struct perf_tool *tool __used,
+                                struct event *event,
+                                struct perf_sample *sample,
+                                struct machine *machine,
+                                struct thread *thread)
+{
+       void *data = sample->raw_data;
+       struct trace_migrate_task_event migrate_task_event;
+
+       FILL_COMMON_FIELDS(migrate_task_event, event, data);
 
-       if (head + event->header.size >= page_size * mmap_window) {
-               unsigned long shift = page_size * (head / page_size);
-               int res;
+       FILL_ARRAY(migrate_task_event, comm, event, data);
+       FILL_FIELD(migrate_task_event, pid, event, data);
+       FILL_FIELD(migrate_task_event, prio, event, data);
+       FILL_FIELD(migrate_task_event, cpu, event, data);
+
+       if (trace_handler->migrate_task_event)
+               trace_handler->migrate_task_event(&migrate_task_event, machine,
+                                                 event, sample->cpu,
+                                                 sample->time, thread);
+}
 
-               res = munmap(buf, page_size * mmap_window);
-               assert(res == 0);
+typedef void (*tracepoint_handler)(struct perf_tool *tool, struct event *event,
+                                  struct perf_sample *sample,
+                                  struct machine *machine,
+                                  struct thread *thread);
 
-               offset += shift;
-               head -= shift;
-               goto remap;
+static int perf_sched__process_tracepoint_sample(struct perf_tool *tool,
+                                                union perf_event *event __used,
+                                                struct perf_sample *sample,
+                                                struct perf_evsel *evsel,
+                                                struct machine *machine)
+{
+       struct thread *thread = machine__findnew_thread(machine, sample->pid);
+
+       if (thread == NULL) {
+               pr_debug("problem processing %s event, skipping it.\n",
+                        evsel->name);
+               return -1;
        }
 
-       size = event->header.size;
+       evsel->hists.stats.total_period += sample->period;
+       hists__inc_nr_events(&evsel->hists, PERF_RECORD_SAMPLE);
 
+       if (evsel->handler.func != NULL) {
+               tracepoint_handler f = evsel->handler.func;
 
-       if (!size || process_event(event, offset, head) < 0) {
+               if (evsel->handler.data == NULL)
+                       evsel->handler.data = trace_find_event(evsel->attr.config);
 
-               /*
-                * assume we lost track of the stream, check alignment, and
-                * increment a single u64 in the hope to catch on again 'soon'.
-                */
+               f(tool, evsel->handler.data, sample, machine, thread);
+       }
 
-               if (unlikely(head & 7))
-                       head &= ~7ULL;
+       return 0;
+}
+
+static struct perf_tool perf_sched = {
+       .sample                 = perf_sched__process_tracepoint_sample,
+       .comm                   = perf_event__process_comm,
+       .lost                   = perf_event__process_lost,
+       .fork                   = perf_event__process_task,
+       .ordered_samples        = true,
+};
 
-               size = 8;
+static void read_events(bool destroy, struct perf_session **psession)
+{
+       int err = -EINVAL;
+       const struct perf_evsel_str_handler handlers[] = {
+               { "sched:sched_switch",       process_sched_switch_event, },
+               { "sched:sched_stat_runtime", process_sched_runtime_event, },
+               { "sched:sched_wakeup",       process_sched_wakeup_event, },
+               { "sched:sched_wakeup_new",   process_sched_wakeup_event, },
+               { "sched:sched_process_fork", process_sched_fork_event, },
+               { "sched:sched_process_exit", process_sched_exit_event, },
+               { "sched:sched_migrate_task", process_sched_migrate_task_event, },
+       };
+       struct perf_session *session = perf_session__new(input_name, O_RDONLY,
+                                                        0, false, &perf_sched);
+       if (session == NULL)
+               die("No Memory");
+
+       err = perf_evlist__set_tracepoints_handlers_array(session->evlist, handlers);
+       assert(err == 0);
+
+       if (perf_session__has_traces(session, "record -R")) {
+               err = perf_session__process_events(session, &perf_sched);
+               if (err)
+                       die("Failed to process events, error %d", err);
+
+               nr_events      = session->hists.stats.nr_events[0];
+               nr_lost_events = session->hists.stats.total_lost;
+               nr_lost_chunks = session->hists.stats.nr_events[PERF_RECORD_LOST];
        }
 
-       head += size;
+       if (destroy)
+               perf_session__delete(session);
 
-       if (offset + head < (unsigned long)perf_stat.st_size)
-               goto more;
+       if (psession)
+               *psession = session;
+}
 
-       rc = EXIT_SUCCESS;
-       close(input);
+static void print_bad_events(void)
+{
+       if (nr_unordered_timestamps && nr_timestamps) {
+               printf("  INFO: %.3f%% unordered timestamps (%ld out of %ld)\n",
+                       (double)nr_unordered_timestamps/(double)nr_timestamps*100.0,
+                       nr_unordered_timestamps, nr_timestamps);
+       }
+       if (nr_lost_events && nr_events) {
+               printf("  INFO: %.3f%% lost events (%ld out of %ld, in %ld chunks)\n",
+                       (double)nr_lost_events/(double)nr_events*100.0,
+                       nr_lost_events, nr_events, nr_lost_chunks);
+       }
+       if (nr_state_machine_bugs && nr_timestamps) {
+               printf("  INFO: %.3f%% state machine bugs (%ld out of %ld)",
+                       (double)nr_state_machine_bugs/(double)nr_timestamps*100.0,
+                       nr_state_machine_bugs, nr_timestamps);
+               if (nr_lost_events)
+                       printf(" (due to lost events?)");
+               printf("\n");
+       }
+       if (nr_context_switch_bugs && nr_timestamps) {
+               printf("  INFO: %.3f%% context switch bugs (%ld out of %ld)",
+                       (double)nr_context_switch_bugs/(double)nr_timestamps*100.0,
+                       nr_context_switch_bugs, nr_timestamps);
+               if (nr_lost_events)
+                       printf(" (due to lost events?)");
+               printf("\n");
+       }
+}
 
-       return rc;
+static void __cmd_lat(void)
+{
+       struct rb_node *next;
+       struct perf_session *session;
+
+       setup_pager();
+       read_events(false, &session);
+       sort_lat();
+
+       printf("\n ---------------------------------------------------------------------------------------------------------------\n");
+       printf("  Task                  |   Runtime ms  | Switches | Average delay ms | Maximum delay ms | Maximum delay at     |\n");
+       printf(" ---------------------------------------------------------------------------------------------------------------\n");
+
+       next = rb_first(&sorted_atom_root);
+
+       while (next) {
+               struct work_atoms *work_list;
+
+               work_list = rb_entry(next, struct work_atoms, node);
+               output_lat_thread(work_list);
+               next = rb_next(next);
+       }
+
+       printf(" -----------------------------------------------------------------------------------------\n");
+       printf("  TOTAL:                |%11.3f ms |%9" PRIu64 " |\n",
+               (double)all_runtime/1e6, all_count);
+
+       printf(" ---------------------------------------------------\n");
+
+       print_bad_events();
+       printf("\n");
+
+       perf_session__delete(session);
 }
 
-static const char * const annotate_usage[] = {
-       "perf trace [<options>] <command>",
+static struct trace_sched_handler map_ops  = {
+       .wakeup_event           = NULL,
+       .switch_event           = map_switch_event,
+       .runtime_event          = NULL,
+       .fork_event             = NULL,
+};
+
+static void __cmd_map(void)
+{
+       max_cpu = sysconf(_SC_NPROCESSORS_CONF);
+
+       setup_pager();
+       read_events(true, NULL);
+       print_bad_events();
+}
+
+static void __cmd_replay(void)
+{
+       unsigned long i;
+
+       calibrate_run_measurement_overhead();
+       calibrate_sleep_measurement_overhead();
+
+       test_calibrations();
+
+       read_events(true, NULL);
+
+       printf("nr_run_events:        %ld\n", nr_run_events);
+       printf("nr_sleep_events:      %ld\n", nr_sleep_events);
+       printf("nr_wakeup_events:     %ld\n", nr_wakeup_events);
+
+       if (targetless_wakeups)
+               printf("target-less wakeups:  %ld\n", targetless_wakeups);
+       if (multitarget_wakeups)
+               printf("multi-target wakeups: %ld\n", multitarget_wakeups);
+       if (nr_run_events_optimized)
+               printf("run atoms optimized: %ld\n",
+                       nr_run_events_optimized);
+
+       print_task_traces();
+       add_cross_task_wakeups();
+
+       create_tasks();
+       printf("------------------------------------------------------------\n");
+       for (i = 0; i < replay_repeat; i++)
+               run_one_test();
+}
+
+
+static const char * const sched_usage[] = {
+       "perf sched [<options>] {record|latency|map|replay|script}",
        NULL
 };
 
-static const struct option options[] = {
+static const struct option sched_options[] = {
+       OPT_STRING('i', "input", &input_name, "file",
+                   "input file name"),
+       OPT_INCR('v', "verbose", &verbose,
+                   "be more verbose (show symbol address, etc)"),
        OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
                    "dump raw trace in ASCII"),
-       OPT_BOOLEAN('v', "verbose", &verbose,
+       OPT_END()
+};
+
+static const char * const latency_usage[] = {
+       "perf sched latency [<options>]",
+       NULL
+};
+
+static const struct option latency_options[] = {
+       OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
+                  "sort by key(s): runtime, switch, avg, max"),
+       OPT_INCR('v', "verbose", &verbose,
                    "be more verbose (show symbol address, etc)"),
+       OPT_INTEGER('C', "CPU", &profile_cpu,
+                   "CPU to profile on"),
+       OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
+                   "dump raw trace in ASCII"),
        OPT_END()
 };
 
-int cmd_sched(int argc, const char **argv, const char *prefix __used)
-{
-       symbol__init();
-       page_size = getpagesize();
+static const char * const replay_usage[] = {
+       "perf sched replay [<options>]",
+       NULL
+};
 
-       argc = parse_options(argc, argv, options, annotate_usage, 0);
-       if (argc) {
-               /*
-                * Special case: if there's an argument left then assume tha
-                * it's a symbol filter:
-                */
-               if (argc > 1)
-                       usage_with_options(annotate_usage, options);
+static const struct option replay_options[] = {
+       OPT_UINTEGER('r', "repeat", &replay_repeat,
+                    "repeat the workload replay N times (-1: infinite)"),
+       OPT_INCR('v', "verbose", &verbose,
+                   "be more verbose (show symbol address, etc)"),
+       OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
+                   "dump raw trace in ASCII"),
+       OPT_END()
+};
+
+static void setup_sorting(void)
+{
+       char *tmp, *tok, *str = strdup(sort_order);
+
+       for (tok = strtok_r(str, ", ", &tmp);
+                       tok; tok = strtok_r(NULL, ", ", &tmp)) {
+               if (sort_dimension__add(tok, &sort_list) < 0) {
+                       error("Unknown --sort key: `%s'", tok);
+                       usage_with_options(latency_usage, latency_options);
+               }
        }
 
+       free(str);
 
-       setup_pager();
+       sort_dimension__add("pid", &cmp_pid);
+}
+
+static const char *record_args[] = {
+       "record",
+       "-a",
+       "-R",
+       "-f",
+       "-m", "1024",
+       "-c", "1",
+       "-e", "sched:sched_switch",
+       "-e", "sched:sched_stat_wait",
+       "-e", "sched:sched_stat_sleep",
+       "-e", "sched:sched_stat_iowait",
+       "-e", "sched:sched_stat_runtime",
+       "-e", "sched:sched_process_exit",
+       "-e", "sched:sched_process_fork",
+       "-e", "sched:sched_wakeup",
+       "-e", "sched:sched_migrate_task",
+};
+
+static int __cmd_record(int argc, const char **argv)
+{
+       unsigned int rec_argc, i, j;
+       const char **rec_argv;
+
+       rec_argc = ARRAY_SIZE(record_args) + argc - 1;
+       rec_argv = calloc(rec_argc + 1, sizeof(char *));
+
+       if (rec_argv == NULL)
+               return -ENOMEM;
 
-       return __cmd_sched();
+       for (i = 0; i < ARRAY_SIZE(record_args); i++)
+               rec_argv[i] = strdup(record_args[i]);
+
+       for (j = 1; j < (unsigned int)argc; j++, i++)
+               rec_argv[i] = argv[j];
+
+       BUG_ON(i != rec_argc);
+
+       return cmd_record(i, rec_argv, NULL);
+}
+
+int cmd_sched(int argc, const char **argv, const char *prefix __used)
+{
+       argc = parse_options(argc, argv, sched_options, sched_usage,
+                            PARSE_OPT_STOP_AT_NON_OPTION);
+       if (!argc)
+               usage_with_options(sched_usage, sched_options);
+
+       /*
+        * Aliased to 'perf script' for now:
+        */
+       if (!strcmp(argv[0], "script"))
+               return cmd_script(argc, argv, prefix);
+
+       symbol__init();
+       if (!strncmp(argv[0], "rec", 3)) {
+               return __cmd_record(argc, argv);
+       } else if (!strncmp(argv[0], "lat", 3)) {
+               trace_handler = &lat_ops;
+               if (argc > 1) {
+                       argc = parse_options(argc, argv, latency_options, latency_usage, 0);
+                       if (argc)
+                               usage_with_options(latency_usage, latency_options);
+               }
+               setup_sorting();
+               __cmd_lat();
+       } else if (!strcmp(argv[0], "map")) {
+               trace_handler = &map_ops;
+               setup_sorting();
+               __cmd_map();
+       } else if (!strncmp(argv[0], "rep", 3)) {
+               trace_handler = &replay_ops;
+               if (argc) {
+                       argc = parse_options(argc, argv, replay_options, replay_usage, 0);
+                       if (argc)
+                               usage_with_options(replay_usage, replay_options);
+               }
+               __cmd_replay();
+       } else {
+               usage_with_options(sched_usage, sched_options);
+       }
+
+       return 0;
 }