perf_event: Fix raw event processing
[linux-3.10.git] / tools / perf / builtin-sched.c
index f67e351..4655e16 100644 (file)
@@ -11,6 +11,7 @@
 #include "util/trace-event.h"
 
 #include "util/debug.h"
+#include "util/data_map.h"
 
 #include <sys/types.h>
 #include <sys/prctl.h>
 #include <math.h>
 
 static char                    const *input_name = "perf.data";
-static int                     input;
-static unsigned long           page_size;
-static unsigned long           mmap_window = 32;
-
-static unsigned long           total_comm = 0;
-
-static struct rb_root          threads;
-static struct thread           *last_match;
 
 static struct perf_header      *header;
 static u64                     sample_type;
@@ -35,11 +28,11 @@ static u64                  sample_type;
 static char                    default_sort_order[] = "avg, max, switch, runtime";
 static char                    *sort_order = default_sort_order;
 
+static int                     profile_cpu = -1;
+
 #define PR_SET_NAME            15               /* Set process name */
 #define MAX_CPUS               4096
 
-#define BUG_ON(x)              assert(!(x))
-
 static u64                     run_measurement_overhead;
 static u64                     sleep_measurement_overhead;
 
@@ -74,6 +67,7 @@ enum sched_event_type {
        SCHED_EVENT_RUN,
        SCHED_EVENT_SLEEP,
        SCHED_EVENT_WAKEUP,
+       SCHED_EVENT_MIGRATION,
 };
 
 struct sched_atom {
@@ -226,7 +220,7 @@ static void calibrate_sleep_measurement_overhead(void)
 static struct sched_atom *
 get_new_event(struct task_desc *task, u64 timestamp)
 {
-       struct sched_atom *event = calloc(1, sizeof(*event));
+       struct sched_atom *event = zalloc(sizeof(*event));
        unsigned long idx = task->nr_events;
        size_t size;
 
@@ -294,7 +288,7 @@ add_sched_event_wakeup(struct task_desc *task, u64 timestamp,
                return;
        }
 
-       wakee_event->wait_sem = calloc(1, sizeof(*wakee_event->wait_sem));
+       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;
@@ -324,7 +318,7 @@ static struct task_desc *register_pid(unsigned long pid, const char *comm)
        if (task)
                return task;
 
-       task = calloc(1, sizeof(*task));
+       task = zalloc(sizeof(*task));
        task->pid = pid;
        task->nr = nr_tasks;
        strcpy(task->comm, comm);
@@ -398,6 +392,8 @@ process_sched_event(struct task_desc *this_task __used, struct sched_atom *atom)
                                ret = sem_post(atom->wait_sem);
                        BUG_ON(ret);
                        break;
+               case SCHED_EVENT_MIGRATION:
+                       break;
                default:
                        BUG_ON(1);
        }
@@ -632,29 +628,6 @@ static void test_calibrations(void)
        printf("the sleep test took %Ld nsecs\n", T1-T0);
 }
 
-static int
-process_comm_event(event_t *event, unsigned long offset, unsigned long head)
-{
-       struct thread *thread;
-
-       thread = threads__findnew(event->comm.pid, &threads, &last_match);
-
-       dump_printf("%p [%p]: perf_event_comm: %s:%d\n",
-               (void *)(offset + head),
-               (void *)(long)(event->header.size),
-               event->comm.comm, event->comm.pid);
-
-       if (thread == NULL ||
-           thread__set_comm(thread, event->comm.comm)) {
-               dump_printf("problem processing perf_event_comm, skipping event.\n");
-               return -1;
-       }
-       total_comm++;
-
-       return 0;
-}
-
-
 struct raw_event_sample {
        u32 size;
        char data[0];
@@ -745,6 +718,22 @@ struct trace_fork_event {
        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 event *,
@@ -769,6 +758,12 @@ struct trace_sched_handler {
                           int cpu,
                           u64 timestamp,
                           struct thread *thread);
+
+       void (*migrate_task_event)(struct trace_migrate_task_event *,
+                          struct event *,
+                          int cpu,
+                          u64 timestamp,
+                          struct thread *thread);
 };
 
 
@@ -941,9 +936,7 @@ __thread_latency_insert(struct rb_root *root, struct work_atoms *data,
 
 static void thread_atoms_insert(struct thread *thread)
 {
-       struct work_atoms *atoms;
-
-       atoms = calloc(sizeof(*atoms), 1);
+       struct work_atoms *atoms = zalloc(sizeof(*atoms));
        if (!atoms)
                die("No memory");
 
@@ -975,9 +968,7 @@ add_sched_out_event(struct work_atoms *atoms,
                    char run_state,
                    u64 timestamp)
 {
-       struct work_atom *atom;
-
-       atom = calloc(sizeof(*atom), 1);
+       struct work_atom *atom = zalloc(sizeof(*atom));
        if (!atom)
                die("Non memory");
 
@@ -1058,8 +1049,8 @@ latency_switch_event(struct trace_switch_event *switch_event,
                die("hm, delta: %Ld < 0 ?\n", delta);
 
 
-       sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
-       sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
+       sched_out = threads__findnew(switch_event->prev_pid);
+       sched_in = threads__findnew(switch_event->next_pid);
 
        out_events = thread_atoms_search(&atom_root, sched_out, &cmp_pid);
        if (!out_events) {
@@ -1092,13 +1083,10 @@ latency_runtime_event(struct trace_runtime_event *runtime_event,
                     u64 timestamp,
                     struct thread *this_thread __used)
 {
-       struct work_atoms *atoms;
-       struct thread *thread;
+       struct thread *thread = threads__findnew(runtime_event->pid);
+       struct work_atoms *atoms = thread_atoms_search(&atom_root, thread, &cmp_pid);
 
        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);
@@ -1125,7 +1113,7 @@ latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
        if (!wakeup_event->success)
                return;
 
-       wakee = threads__findnew(wakeup_event->pid, &threads, &last_match);
+       wakee = threads__findnew(wakeup_event->pid);
        atoms = thread_atoms_search(&atom_root, wakee, &cmp_pid);
        if (!atoms) {
                thread_atoms_insert(wakee);
@@ -1139,7 +1127,12 @@ latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
 
        atom = list_entry(atoms->work_list.prev, struct work_atom, list);
 
-       if (atom->state != THREAD_SLEEPING)
+       /*
+        * 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++;
@@ -1152,11 +1145,51 @@ latency_wakeup_event(struct trace_wakeup_event *wakeup_event,
        atom->wake_up_time = timestamp;
 }
 
+static void
+latency_migrate_task_event(struct trace_migrate_task_event *migrate_task_event,
+                    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 = threads__findnew(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)
@@ -1287,7 +1320,7 @@ static struct sort_dimension *available_sorts[] = {
 
 static LIST_HEAD(sort_list);
 
-static int sort_dimension__add(char *tok, struct list_head *list)
+static int sort_dimension__add(const char *tok, struct list_head *list)
 {
        int i;
 
@@ -1347,7 +1380,7 @@ process_sched_wakeup_event(struct raw_event_sample *raw,
  * 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 = 15;
+static int max_cpu;
 
 static u32 curr_pid[MAX_CPUS] = { [0 ... MAX_CPUS-1] = -1 };
 
@@ -1385,8 +1418,8 @@ map_switch_event(struct trace_switch_event *switch_event,
                die("hm, delta: %Ld < 0 ?\n", delta);
 
 
-       sched_out = threads__findnew(switch_event->prev_pid, &threads, &last_match);
-       sched_in = threads__findnew(switch_event->next_pid, &threads, &last_match);
+       sched_out = threads__findnew(switch_event->prev_pid);
+       sched_in = threads__findnew(switch_event->next_pid);
 
        curr_thread[this_cpu] = sched_in;
 
@@ -1517,13 +1550,37 @@ process_sched_exit_event(struct event *event,
 }
 
 static void
-process_raw_event(event_t *raw_event __used, void *more_data,
+process_sched_migrate_task_event(struct raw_event_sample *raw,
+                          struct event *event,
+                          int cpu __used,
+                          u64 timestamp __used,
+                          struct thread *thread __used)
+{
+       struct trace_migrate_task_event migrate_task_event;
+
+       FILL_COMMON_FIELDS(migrate_task_event, event, raw->data);
+
+       FILL_ARRAY(migrate_task_event, comm, event, raw->data);
+       FILL_FIELD(migrate_task_event, pid, event, raw->data);
+       FILL_FIELD(migrate_task_event, prio, event, raw->data);
+       FILL_FIELD(migrate_task_event, cpu, event, raw->data);
+
+       if (trace_handler->migrate_task_event)
+               trace_handler->migrate_task_event(&migrate_task_event, event, cpu, timestamp, thread);
+}
+
+static void
+process_raw_event(event_t *raw_event __used, u32 size, void *data,
                  int cpu, u64 timestamp, struct thread *thread)
 {
-       struct raw_event_sample *raw = more_data;
+       struct raw_event_sample *raw;
        struct event *event;
        int type;
 
+       raw = malloc_or_die(sizeof(*raw)+size);
+       raw->size = size;
+       memcpy(raw->data, data, size);
+
        type = trace_parse_common_type(raw->data);
        event = trace_find_event(type);
 
@@ -1539,211 +1596,85 @@ process_raw_event(event_t *raw_event __used, void *more_data,
                process_sched_fork_event(raw, event, cpu, timestamp, thread);
        if (!strcmp(event->name, "sched_process_exit"))
                process_sched_exit_event(event, cpu, timestamp, thread);
+       if (!strcmp(event->name, "sched_migrate_task"))
+               process_sched_migrate_task_event(raw, event, cpu, timestamp, thread);
 }
 
-static int
-process_sample_event(event_t *event, unsigned long offset, unsigned long head)
+static int process_sample_event(event_t *event)
 {
-       char level;
-       int show = 0;
-       struct dso *dso = NULL;
+       struct sample_data data;
        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;
-
-       thread = threads__findnew(event->ip.pid, &threads, &last_match);
-
-       if (sample_type & PERF_SAMPLE_TIME) {
-               timestamp = *(u64 *)more_data;
-               more_data += sizeof(u64);
-       }
 
-       if (sample_type & PERF_SAMPLE_CPU) {
-               cpu = *(u32 *)more_data;
-               more_data += sizeof(u32);
-               more_data += sizeof(u32); /* reserved */
-       }
+       if (!(sample_type & PERF_SAMPLE_RAW))
+               return 0;
 
-       if (sample_type & PERF_SAMPLE_PERIOD) {
-               period = *(u64 *)more_data;
-               more_data += sizeof(u64);
-       }
+       memset(&data, 0, sizeof(data));
+       data.time = -1;
+       data.cpu = -1;
+       data.period = -1;
 
-       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);
+       event__parse_sample(event, sample_type, &data);
 
-       dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
+       dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
+               event->header.misc,
+               data.pid, data.tid,
+               (void *)(long)data.ip,
+               (long long)data.period);
 
+       thread = threads__findnew(data.pid);
        if (thread == NULL) {
-               eprintf("problem processing %d event, skipping it.\n",
-                       event->header.type);
+               pr_debug("problem processing %d event, skipping it.\n",
+                        event->header.type);
                return -1;
        }
 
-       cpumode = event->header.misc & PERF_EVENT_MISC_CPUMODE_MASK;
-
-       if (cpumode == PERF_EVENT_MISC_KERNEL) {
-               show = SHOW_KERNEL;
-               level = 'k';
-
-               dso = kernel_dso;
-
-               dump_printf(" ...... dso: %s\n", dso->name);
-
-       } else if (cpumode == PERF_EVENT_MISC_USER) {
-
-               show = SHOW_USER;
-               level = '.';
-
-       } else {
-               show = SHOW_HV;
-               level = 'H';
-
-               dso = hypervisor_dso;
+       dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
 
-               dump_printf(" ...... dso: [hypervisor]\n");
-       }
+       if (profile_cpu != -1 && profile_cpu != (int)data.cpu)
+               return 0;
 
-       if (sample_type & PERF_SAMPLE_RAW)
-               process_raw_event(event, more_data, cpu, timestamp, thread);
+       process_raw_event(event, data.raw_size, data.raw_data, data.cpu,
+                         data.time, thread);
 
        return 0;
 }
 
-static int
-process_event(event_t *event, unsigned long offset, unsigned long head)
+static int process_lost_event(event_t *event __used)
 {
-       trace_event(event);
+       nr_lost_chunks++;
+       nr_lost_events += event->lost.lost;
 
-       nr_events++;
-       switch (event->header.type) {
-       case PERF_EVENT_MMAP:
-               return 0;
-       case PERF_EVENT_LOST:
-               nr_lost_chunks++;
-               nr_lost_events += event->lost.lost;
-               return 0;
-
-       case PERF_EVENT_COMM:
-               return process_comm_event(event, offset, head);
-
-       case PERF_EVENT_EXIT ... PERF_EVENT_READ:
-               return 0;
+       return 0;
+}
 
-       case PERF_EVENT_SAMPLE:
-               return process_sample_event(event, offset, head);
+static int sample_type_check(u64 type)
+{
+       sample_type = type;
 
-       case PERF_EVENT_MAX:
-       default:
+       if (!(sample_type & PERF_SAMPLE_RAW)) {
+               fprintf(stderr,
+                       "No trace sample to read. Did you call perf record "
+                       "without -R?");
                return -1;
        }
 
        return 0;
 }
 
+static struct perf_file_handler file_handler = {
+       .process_sample_event   = process_sample_event,
+       .process_comm_event     = event__process_comm,
+       .process_lost_event     = process_lost_event,
+       .sample_type_check      = sample_type_check,
+};
+
 static int read_events(void)
 {
-       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;
-
-       trace_report();
-       register_idle_thread(&threads, &last_match);
-
-       input = open(input_name, O_RDONLY);
-       if (input < 0) {
-               perror("failed to open file");
-               exit(-1);
-       }
-
-       ret = fstat(input, &perf_stat);
-       if (ret < 0) {
-               perror("failed to stat file");
-               exit(-1);
-       }
-
-       if (!perf_stat.st_size) {
-               fprintf(stderr, "zero-sized file, nothing to do!\n");
-               exit(0);
-       }
-       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?");
-
-       if (load_kernel() < 0) {
-               perror("failed to load kernel symbols");
-               return EXIT_FAILURE;
-       }
-
-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);
-       }
-
-more:
-       event = (event_t *)(buf + head);
-
-       size = event->header.size;
-       if (!size)
-               size = 8;
-
-       if (head + event->header.size >= page_size * mmap_window) {
-               unsigned long shift = page_size * (head / page_size);
-               int res;
-
-               res = munmap(buf, page_size * mmap_window);
-               assert(res == 0);
-
-               offset += shift;
-               head -= shift;
-               goto remap;
-       }
-
-       size = event->header.size;
-
+       register_idle_thread();
+       register_perf_file_handler(&file_handler);
 
-       if (!size || process_event(event, offset, head) < 0) {
-
-               /*
-                * assume we lost track of the stream, check alignment, and
-                * increment a single u64 in the hope to catch on again 'soon'.
-                */
-
-               if (unlikely(head & 7))
-                       head &= ~7ULL;
-
-               size = 8;
-       }
-
-       head += size;
-
-       if (offset + head < (unsigned long)perf_stat.st_size)
-               goto more;
-
-       rc = EXIT_SUCCESS;
-       close(input);
-
-       return rc;
+       return mmap_dispatch_perf_file(&header, input_name, 0, 0,
+                                      &event__cwdlen, &event__cwd);
 }
 
 static void print_bad_events(void)
@@ -1818,6 +1749,8 @@ static struct trace_sched_handler map_ops  = {
 
 static void __cmd_map(void)
 {
+       max_cpu = sysconf(_SC_NPROCESSORS_CONF);
+
        setup_pager();
        read_events();
        print_bad_events();
@@ -1857,11 +1790,13 @@ static void __cmd_replay(void)
 
 
 static const char * const sched_usage[] = {
-       "perf sched [<options>] {record|latency|replay|trace}",
+       "perf sched [<options>] {record|latency|map|replay|trace}",
        NULL
 };
 
 static const struct option sched_options[] = {
+       OPT_STRING('i', "input", &input_name, "file",
+                   "input file name"),
        OPT_BOOLEAN('v', "verbose", &verbose,
                    "be more verbose (show symbol address, etc)"),
        OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
@@ -1879,6 +1814,8 @@ static const struct option latency_options[] = {
                   "sort by key(s): runtime, switch, avg, max"),
        OPT_BOOLEAN('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()
@@ -1913,7 +1850,7 @@ static void setup_sorting(void)
 
        free(str);
 
-       sort_dimension__add((char *)"pid", &cmp_pid);
+       sort_dimension__add("pid", &cmp_pid);
 }
 
 static const char *record_args[] = {
@@ -1956,14 +1893,18 @@ static int __cmd_record(int argc, const char **argv)
 
 int cmd_sched(int argc, const char **argv, const char *prefix __used)
 {
-       symbol__init();
-       page_size = getpagesize();
-
        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 trace' for now:
+        */
+       if (!strcmp(argv[0], "trace"))
+               return cmd_trace(argc, argv, prefix);
+
+       symbol__init(0);
        if (!strncmp(argv[0], "rec", 3)) {
                return __cmd_record(argc, argv);
        } else if (!strncmp(argv[0], "lat", 3)) {
@@ -1987,11 +1928,6 @@ int cmd_sched(int argc, const char **argv, const char *prefix __used)
                                usage_with_options(replay_usage, replay_options);
                }
                __cmd_replay();
-       } else if (!strcmp(argv[0], "trace")) {
-               /*
-                * Aliased to 'perf trace' for now:
-                */
-               return cmd_trace(argc, argv, prefix);
        } else {
                usage_with_options(sched_usage, sched_options);
        }