perf sched: Add support for sched:sched_stat_runtime events
authormingo <mingo@europe.(none)>
Mon, 14 Sep 2009 18:04:48 +0000 (20:04 +0200)
committerIngo Molnar <mingo@elte.hu>
Mon, 14 Sep 2009 18:08:23 +0000 (20:08 +0200)
This allows more precise 'perf sched latency' output:

 ---------------------------------------------------------------------------------------
  Task                  |  Runtime ms | Switches | Average delay ms | Maximum delay ms |
 ---------------------------------------------------------------------------------------
  ksoftirqd/0-4         |    0.010 ms |        2 | avg:    2.476 ms | max:    2.977 ms |
  perf-12328            |   15.844 ms |       66 | avg:    1.118 ms | max:    9.979 ms |
  bdi-default-235       |    0.009 ms |        1 | avg:    0.998 ms | max:    0.998 ms |
  events/1-8            |    0.020 ms |        2 | avg:    0.998 ms | max:    0.998 ms |
  events/0-7            |    0.018 ms |        2 | avg:    0.992 ms | max:    0.996 ms |
  sleep-12329           |    0.742 ms |        3 | avg:    0.906 ms | max:    2.289 ms |
  sshd-12122            |    0.163 ms |        2 | avg:    0.283 ms | max:    0.562 ms |
  loop-getpid-lon-12322 | 1023.636 ms |       69 | avg:    0.208 ms | max:    5.996 ms |
  loop-getpid-lon-12321 | 1038.638 ms |        5 | avg:    0.073 ms | max:    0.171 ms |
  migration/1-5         |    0.000 ms |        1 | avg:    0.006 ms | max:    0.006 ms |
 ---------------------------------------------------------------------------------------
  TOTAL:                | 2079.078 ms |      153 |
 -------------------------------------------------

Also, streamline the code a bit more, add asserts for various state
machine failures (they should be debugged if they occur) and fix
a few odd ends.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
LKML-Reference: <new-submission>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
tools/perf/builtin-sched.c

index 93ef7b215aba7568f872f68040933bcf7c0c8a0c..adcb563ec4d28e7999c89bb17aa107bfe2ead549 100644 (file)
@@ -50,7 +50,7 @@ static u64                    sleep_measurement_overhead;
 
 static unsigned long           nr_tasks;
 
-struct sched_event;
+struct sched_atom;
 
 struct task_desc {
        unsigned long           nr;
@@ -59,7 +59,7 @@ struct task_desc {
 
        unsigned long           nr_events;
        unsigned long           curr_event;
-       struct sched_event      **events;
+       struct sched_atom       **atoms;
 
        pthread_t               thread;
        sem_t                   sleep_sem;
@@ -76,7 +76,7 @@ enum sched_event_type {
        SCHED_EVENT_WAKEUP,
 };
 
-struct sched_event {
+struct sched_atom {
        enum sched_event_type   type;
        u64                     timestamp;
        u64                     duration;
@@ -137,8 +137,8 @@ struct work_atom {
        u64                     runtime;
 };
 
-struct task_atoms {
-       struct list_head        atom_list;
+struct work_atoms {
+       struct list_head        work_list;
        struct thread           *thread;
        struct rb_node          node;
        u64                     max_lat;
@@ -147,7 +147,7 @@ struct task_atoms {
        u64                     total_runtime;
 };
 
-typedef int (*sort_fn_t)(struct task_atoms *, struct task_atoms *);
+typedef int (*sort_fn_t)(struct work_atoms *, struct work_atoms *);
 
 static struct rb_root          atom_root, sorted_atom_root;
 
@@ -220,10 +220,10 @@ static void calibrate_sleep_measurement_overhead(void)
        printf("sleep measurement overhead: %Ld nsecs\n", min_delta);
 }
 
-static struct sched_event *
+static struct sched_atom *
 get_new_event(struct task_desc *task, u64 timestamp)
 {
-       struct sched_event *event = calloc(1, sizeof(*event));
+       struct sched_atom *event = calloc(1, sizeof(*event));
        unsigned long idx = task->nr_events;
        size_t size;
 
@@ -231,27 +231,27 @@ get_new_event(struct task_desc *task, u64 timestamp)
        event->nr = idx;
 
        task->nr_events++;
-       size = sizeof(struct sched_event *) * task->nr_events;
-       task->events = realloc(task->events, size);
-       BUG_ON(!task->events);
+       size = sizeof(struct sched_atom *) * task->nr_events;
+       task->atoms = realloc(task->atoms, size);
+       BUG_ON(!task->atoms);
 
-       task->events[idx] = event;
+       task->atoms[idx] = event;
 
        return event;
 }
 
-static struct sched_event *last_event(struct task_desc *task)
+static struct sched_atom *last_event(struct task_desc *task)
 {
        if (!task->nr_events)
                return NULL;
 
-       return task->events[task->nr_events - 1];
+       return task->atoms[task->nr_events - 1];
 }
 
 static void
 add_sched_event_run(struct task_desc *task, u64 timestamp, u64 duration)
 {
-       struct sched_event *event, *curr_event = last_event(task);
+       struct sched_atom *event, *curr_event = last_event(task);
 
        /*
         * optimize an existing RUN event by merging this one
@@ -275,7 +275,7 @@ static void
 add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
                       struct task_desc *wakee)
 {
-       struct sched_event *event, *wakee_event;
+       struct sched_atom *event, *wakee_event;
 
        event = get_new_event(task, timestamp);
        event->type = SCHED_EVENT_WAKEUP;
@@ -303,7 +303,7 @@ static void
 add_sched_event_sleep(struct task_desc *task, u64 timestamp,
                      u64 task_state __used)
 {
-       struct sched_event *event = get_new_event(task, timestamp);
+       struct sched_atom *event = get_new_event(task, timestamp);
 
        event->type = SCHED_EVENT_SLEEP;
 
@@ -372,27 +372,27 @@ static void add_cross_task_wakeups(void)
 }
 
 static void
-process_sched_event(struct task_desc *this_task __used, struct sched_event *event)
+process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
 {
        int ret = 0;
        u64 now;
        long long delta;
 
        now = get_nsecs();
-       delta = start_time + event->timestamp - now;
+       delta = start_time + atom->timestamp - now;
 
-       switch (event->type) {
+       switch (atom->type) {
                case SCHED_EVENT_RUN:
-                       burn_nsecs(event->duration);
+                       burn_nsecs(atom->duration);
                        break;
                case SCHED_EVENT_SLEEP:
-                       if (event->wait_sem)
-                               ret = sem_wait(event->wait_sem);
+                       if (atom->wait_sem)
+                               ret = sem_wait(atom->wait_sem);
                        BUG_ON(ret);
                        break;
                case SCHED_EVENT_WAKEUP:
-                       if (event->wait_sem)
-                               ret = sem_post(event->wait_sem);
+                       if (atom->wait_sem)
+                               ret = sem_post(atom->wait_sem);
                        BUG_ON(ret);
                        break;
                default:
@@ -467,7 +467,7 @@ again:
 
        for (i = 0; i < this_task->nr_events; i++) {
                this_task->curr_event = i;
-               process_sched_event(this_task, this_task->events[i]);
+               process_sched_event(this_task, this_task->atoms[i]);
        }
 
        cpu_usage_1 = get_cpu_usage_nsec_self();
@@ -649,7 +649,7 @@ static void __cmd_replay(void)
        if (multitarget_wakeups)
                printf("multi-target wakeups: %ld\n", multitarget_wakeups);
        if (nr_run_events_optimized)
-               printf("run events optimized: %ld\n",
+               printf("run atoms optimized: %ld\n",
                        nr_run_events_optimized);
 
        print_task_traces();
@@ -727,6 +727,20 @@ struct trace_switch_event {
        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;
@@ -767,6 +781,12 @@ struct trace_sched_handler {
                             u64 timestamp,
                             struct thread *thread);
 
+       void (*runtime_event)(struct trace_runtime_event *,
+                             struct event *,
+                             int cpu,
+                             u64 timestamp,
+                             struct thread *thread);
+
        void (*wakeup_event)(struct trace_wakeup_event *,
                             struct event *,
                             int cpu,
@@ -881,7 +901,7 @@ struct sort_dimension {
 static LIST_HEAD(cmp_pid);
 
 static int
-thread_lat_cmp(struct list_head *list, struct task_atoms *l, struct task_atoms *r)
+thread_lat_cmp(struct list_head *list, struct work_atoms *l, struct work_atoms *r)
 {
        struct sort_dimension *sort;
        int ret = 0;
@@ -897,18 +917,18 @@ thread_lat_cmp(struct list_head *list, struct task_atoms *l, struct task_atoms *
        return ret;
 }
 
-static struct task_atoms *
+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 task_atoms key = { .thread = thread };
+       struct work_atoms key = { .thread = thread };
 
        while (node) {
-               struct task_atoms *atoms;
+               struct work_atoms *atoms;
                int cmp;
 
-               atoms = container_of(node, struct task_atoms, node);
+               atoms = container_of(node, struct work_atoms, node);
 
                cmp = thread_lat_cmp(sort_list, &key, atoms);
                if (cmp > 0)
@@ -924,16 +944,16 @@ thread_atoms_search(struct rb_root *root, struct thread *thread,
 }
 
 static void
-__thread_latency_insert(struct rb_root *root, struct task_atoms *data,
+__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;
 
        while (*new) {
-               struct task_atoms *this;
+               struct work_atoms *this;
                int cmp;
 
-               this = container_of(*new, struct task_atoms, node);
+               this = container_of(*new, struct work_atoms, node);
                parent = *new;
 
                cmp = thread_lat_cmp(sort_list, data, this);
@@ -950,14 +970,14 @@ __thread_latency_insert(struct rb_root *root, struct task_atoms *data,
 
 static void thread_atoms_insert(struct thread *thread)
 {
-       struct task_atoms *atoms;
+       struct work_atoms *atoms;
 
        atoms = calloc(sizeof(*atoms), 1);
        if (!atoms)
                die("No memory");
 
        atoms->thread = thread;
-       INIT_LIST_HEAD(&atoms->atom_list);
+       INIT_LIST_HEAD(&atoms->work_list);
        __thread_latency_insert(&atom_root, atoms, &cmp_pid);
 }
 
@@ -980,10 +1000,9 @@ static char sched_out_state(struct trace_switch_event *switch_event)
 }
 
 static void
-lat_sched_out(struct task_atoms *atoms,
-             struct trace_switch_event *switch_event __used,
-             u64 delta,
-             u64 timestamp)
+add_sched_out_event(struct work_atoms *atoms,
+                   char run_state,
+                   u64 timestamp)
 {
        struct work_atom *atom;
 
@@ -993,25 +1012,37 @@ lat_sched_out(struct task_atoms *atoms,
 
        atom->sched_out_time = timestamp;
 
-       if (sched_out_state(switch_event) == 'R') {
+       if (run_state == 'R') {
                atom->state = THREAD_WAIT_CPU;
                atom->wake_up_time = atom->sched_out_time;
        }
 
-       atom->runtime = delta;
-       list_add_tail(&atom->list, &atoms->atom_list);
+       list_add_tail(&atom->list, &atoms->work_list);
+}
+
+static void
+add_runtime_event(struct work_atoms *atoms, u64 delta, u64 timestamp __used)
+{
+       struct work_atom *atom;
+
+       BUG_ON(list_empty(&atoms->work_list));
+
+       atom = list_entry(atoms->work_list.prev, struct work_atom, list);
+
+       atom->runtime += delta;
+       atoms->total_runtime += delta;
 }
 
 static void
-lat_sched_in(struct task_atoms *atoms, u64 timestamp)
+add_sched_in_event(struct work_atoms *atoms, u64 timestamp)
 {
        struct work_atom *atom;
        u64 delta;
 
-       if (list_empty(&atoms->atom_list))
+       if (list_empty(&atoms->work_list))
                return;
 
-       atom = list_entry(atoms->atom_list.prev, struct work_atom, list);
+       atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 
        if (atom->state != THREAD_WAIT_CPU)
                return;
@@ -1029,7 +1060,6 @@ lat_sched_in(struct task_atoms *atoms, u64 timestamp)
        if (delta > atoms->max_lat)
                atoms->max_lat = delta;
        atoms->nb_atoms++;
-       atoms->total_runtime += atom->runtime;
 }
 
 static void
@@ -1039,13 +1069,12 @@ latency_switch_event(struct trace_switch_event *switch_event,
                     u64 timestamp,
                     struct thread *thread __used)
 {
-       struct task_atoms *out_atoms, *in_atoms;
+       struct work_atoms *out_events, *in_events;
        struct thread *sched_out, *sched_in;
        u64 timestamp0;
        s64 delta;
 
-       if (cpu >= MAX_CPUS || cpu < 0)
-               return;
+       BUG_ON(cpu >= MAX_CPUS || cpu < 0);
 
        timestamp0 = cpu_last_switched[cpu];
        cpu_last_switched[cpu] = timestamp;
@@ -1061,34 +1090,63 @@ latency_switch_event(struct trace_switch_event *switch_event,
        sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
        sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
 
-       in_atoms = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
-       if (!in_atoms) {
+       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_atoms = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
-               if (!in_atoms)
-                       die("in-atom: Internal tree error");
+               in_events = thread_atoms_search(&atom_root, sched_in, &cmp_pid);
+               if (!in_events)
+                       die("in-event: Internal tree error");
+               /*
+                * Take came in we have not heard about yet,
+                * add in an initial atom in runnable state:
+                */
+               add_sched_out_event(in_events, 'R', timestamp);
        }
+       add_sched_in_event(in_events, timestamp);
+}
 
-       out_atoms = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
-       if (!out_atoms) {
-               thread_atoms_insert(sched_out);
-               out_atoms = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
-               if (!out_atoms)
-                       die("out-atom: Internal tree error");
+static void
+latency_runtime_event(struct trace_runtime_event *runtime_event,
+                    struct event *event __used,
+                    int cpu,
+                    u64 timestamp,
+                    struct thread *this_thread __used)
+{
+       struct work_atoms *atoms;
+       struct thread *thread;
+
+       BUG_ON(cpu >= MAX_CPUS || cpu < 0);
+
+       thread = threads__findnew(runtime_event->pid, &threads, &last_match);
+       atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
+       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);
        }
 
-       lat_sched_in(in_atoms, timestamp);
-       lat_sched_out(out_atoms, switch_event, delta, timestamp);
+       add_runtime_event(atoms, runtime_event->runtime, timestamp);
 }
 
 static void
 latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
-                    struct event *event __used,
+                    struct event *__event __used,
                     int cpu __used,
                     u64 timestamp,
                     struct thread *thread __used)
 {
-       struct task_atoms *atoms;
+       struct work_atoms *atoms;
        struct work_atom *atom;
        struct thread *wakee;
 
@@ -1100,16 +1158,20 @@ latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
        atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
        if (!atoms) {
                thread_atoms_insert(wakee);
-               return;
+               atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
+               if (!atoms)
+                       die("wakeup-event: Internal tree error");
+               add_sched_out_event(atoms, 'S', timestamp);
        }
 
-       if (list_empty(&atoms->atom_list))
-               return;
+       BUG_ON(list_empty(&atoms->work_list));
 
-       atom = list_entry(atoms->atom_list.prev, struct work_atom, list);
+       atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 
-       if (atom->state != THREAD_SLEEPING)
+       if (atom->state != THREAD_SLEEPING) {
+               printf("boo2\n");
                return;
+       }
 
        nr_timestamps++;
        if (atom->sched_out_time > timestamp) {
@@ -1124,40 +1186,41 @@ latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
 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,
 };
 
-static void output_lat_thread(struct task_atoms *atom_list)
+static void output_lat_thread(struct work_atoms *work_list)
 {
        int i;
        int ret;
        u64 avg;
 
-       if (!atom_list->nb_atoms)
+       if (!work_list->nb_atoms)
                return;
        /*
         * Ignore idle threads:
         */
-       if (!atom_list->thread->pid)
+       if (!work_list->thread->pid)
                return;
 
-       all_runtime += atom_list->total_runtime;
-       all_count += atom_list->nb_atoms;
+       all_runtime += work_list->total_runtime;
+       all_count += work_list->nb_atoms;
 
-       ret = printf("  %s-%d ", atom_list->thread->comm, atom_list->thread->pid);
+       ret = printf("  %s-%d ", work_list->thread->comm, work_list->thread->pid);
 
        for (i = 0; i < 24 - ret; i++)
                printf(" ");
 
-       avg = atom_list->total_lat / atom_list->nb_atoms;
+       avg = work_list->total_lat / work_list->nb_atoms;
 
        printf("|%9.3f ms |%9llu | avg:%9.3f ms | max:%9.3f ms |\n",
-             (double)atom_list->total_runtime / 1e6,
-                atom_list->nb_atoms, (double)avg / 1e6,
-                (double)atom_list->max_lat / 1e6);
+             (double)work_list->total_runtime / 1e6,
+                work_list->nb_atoms, (double)avg / 1e6,
+                (double)work_list->max_lat / 1e6);
 }
 
-static int pid_cmp(struct task_atoms *l, struct task_atoms *r)
+static int pid_cmp(struct work_atoms *l, struct work_atoms *r)
 {
        if (l->thread->pid < r->thread->pid)
                return -1;
@@ -1172,7 +1235,7 @@ static struct sort_dimension pid_sort_dimension = {
        .cmp                    = pid_cmp,
 };
 
-static int avg_cmp(struct task_atoms *l, struct task_atoms *r)
+static int avg_cmp(struct work_atoms *l, struct work_atoms *r)
 {
        u64 avgl, avgr;
 
@@ -1198,7 +1261,7 @@ static struct sort_dimension avg_sort_dimension = {
        .cmp                    = avg_cmp,
 };
 
-static int max_cmp(struct task_atoms *l, struct task_atoms *r)
+static int max_cmp(struct work_atoms *l, struct work_atoms *r)
 {
        if (l->max_lat < r->max_lat)
                return -1;
@@ -1213,7 +1276,7 @@ static struct sort_dimension max_sort_dimension = {
        .cmp                    = max_cmp,
 };
 
-static int switch_cmp(struct task_atoms *l, struct task_atoms *r)
+static int switch_cmp(struct work_atoms *l, struct work_atoms *r)
 {
        if (l->nb_atoms < r->nb_atoms)
                return -1;
@@ -1228,7 +1291,7 @@ static struct sort_dimension switch_sort_dimension = {
        .cmp                    = switch_cmp,
 };
 
-static int runtime_cmp(struct task_atoms *l, struct task_atoms *r)
+static int runtime_cmp(struct work_atoms *l, struct work_atoms *r)
 {
        if (l->total_runtime < r->total_runtime)
                return -1;
@@ -1277,13 +1340,13 @@ static void sort_lat(void)
        struct rb_node *node;
 
        for (;;) {
-               struct task_atoms *data;
+               struct work_atoms *data;
                node = rb_first(&atom_root);
                if (!node)
                        break;
 
                rb_erase(node, &atom_root);
-               data = rb_entry(node, struct task_atoms, node);
+               data = rb_entry(node, struct work_atoms, node);
                __thread_latency_insert(&sorted_atom_root, data, &sort_list);
        }
 }
@@ -1303,10 +1366,10 @@ static void __cmd_lat(void)
        next = rb_first(&sorted_atom_root);
 
        while (next) {
-               struct task_atoms *atom_list;
+               struct work_atoms *work_list;
 
-               atom_list = rb_entry(next, struct task_atoms, node);
-               output_lat_thread(atom_list);
+               work_list = rb_entry(next, struct work_atoms, node);
+               output_lat_thread(work_list);
                next = rb_next(next);
        }
 
@@ -1368,6 +1431,23 @@ process_sched_switch_event(struct raw_event_sample *raw,
        trace_handler->switch_event(&switch_event, event, cpu, timestamp, thread);
 }
 
+static void
+process_sched_runtime_event(struct raw_event_sample *raw,
+                          struct event *event,
+                          int cpu __used,
+                          u64 timestamp __used,
+                          struct thread *thread __used)
+{
+       struct trace_runtime_event runtime_event;
+
+       FILL_ARRAY(runtime_event, comm, event, raw->data);
+       FILL_FIELD(runtime_event, pid, event, raw->data);
+       FILL_FIELD(runtime_event, runtime, event, raw->data);
+       FILL_FIELD(runtime_event, vruntime, event, raw->data);
+
+       trace_handler->runtime_event(&runtime_event, event, cpu, timestamp, thread);
+}
+
 static void
 process_sched_fork_event(struct raw_event_sample *raw,
                         struct event *event,
@@ -1410,6 +1490,8 @@ process_raw_event(event_t *raw_event __used, void *more_data,
 
        if (!strcmp(event->name, "sched_switch"))
                process_sched_switch_event(raw, event, cpu, timestamp, thread);
+       if (!strcmp(event->name, "sched_stat_runtime"))
+               process_sched_runtime_event(raw, event, cpu, timestamp, thread);
        if (!strcmp(event->name, "sched_wakeup"))
                process_sched_wakeup_event(raw, event, cpu, timestamp, thread);
        if (!strcmp(event->name, "sched_wakeup_new"))