linux/tools/perf/builtin-report.c
<<
>>
Prefs
   1/*
   2 * builtin-report.c
   3 *
   4 * Builtin report command: Analyze the perf.data input file,
   5 * look up and read DSOs and symbol information and display
   6 * a histogram of results, along various sorting keys.
   7 */
   8#include "builtin.h"
   9
  10#include "util/util.h"
  11
  12#include "util/color.h"
  13#include <linux/list.h>
  14#include "util/cache.h"
  15#include <linux/rbtree.h>
  16#include "util/symbol.h"
  17#include "util/callchain.h"
  18#include "util/strlist.h"
  19#include "util/values.h"
  20
  21#include "perf.h"
  22#include "util/debug.h"
  23#include "util/header.h"
  24#include "util/session.h"
  25
  26#include "util/parse-options.h"
  27#include "util/parse-events.h"
  28
  29#include "util/thread.h"
  30#include "util/sort.h"
  31#include "util/hist.h"
  32
  33static char             const *input_name = "perf.data";
  34
  35static bool             force, use_tui, use_stdio;
  36static bool             hide_unresolved;
  37static bool             dont_use_callchains;
  38
  39static bool             show_threads;
  40static struct perf_read_values  show_threads_values;
  41
  42static const char       default_pretty_printing_style[] = "normal";
  43static const char       *pretty_printing_style = default_pretty_printing_style;
  44
  45static char             callchain_default_opt[] = "fractal,0.5";
  46
  47static struct hists *perf_session__hists_findnew(struct perf_session *self,
  48                                                 u64 event_stream, u32 type,
  49                                                 u64 config)
  50{
  51        struct rb_node **p = &self->hists_tree.rb_node;
  52        struct rb_node *parent = NULL;
  53        struct hists *iter, *new;
  54
  55        while (*p != NULL) {
  56                parent = *p;
  57                iter = rb_entry(parent, struct hists, rb_node);
  58                if (iter->config == config)
  59                        return iter;
  60
  61
  62                if (config > iter->config)
  63                        p = &(*p)->rb_right;
  64                else
  65                        p = &(*p)->rb_left;
  66        }
  67
  68        new = malloc(sizeof(struct hists));
  69        if (new == NULL)
  70                return NULL;
  71        memset(new, 0, sizeof(struct hists));
  72        new->event_stream = event_stream;
  73        new->config = config;
  74        new->type = type;
  75        rb_link_node(&new->rb_node, parent, p);
  76        rb_insert_color(&new->rb_node, &self->hists_tree);
  77        return new;
  78}
  79
  80static int perf_session__add_hist_entry(struct perf_session *self,
  81                                        struct addr_location *al,
  82                                        struct sample_data *data)
  83{
  84        struct map_symbol *syms = NULL;
  85        struct symbol *parent = NULL;
  86        int err = -ENOMEM;
  87        struct hist_entry *he;
  88        struct hists *hists;
  89        struct perf_event_attr *attr;
  90
  91        if ((sort__has_parent || symbol_conf.use_callchain) && data->callchain) {
  92                syms = perf_session__resolve_callchain(self, al->thread,
  93                                                       data->callchain, &parent);
  94                if (syms == NULL)
  95                        return -ENOMEM;
  96        }
  97
  98        attr = perf_header__find_attr(data->id, &self->header);
  99        if (attr)
 100                hists = perf_session__hists_findnew(self, data->id, attr->type, attr->config);
 101        else
 102                hists = perf_session__hists_findnew(self, data->id, 0, 0);
 103        if (hists == NULL)
 104                goto out_free_syms;
 105        he = __hists__add_entry(hists, al, parent, data->period);
 106        if (he == NULL)
 107                goto out_free_syms;
 108        err = 0;
 109        if (symbol_conf.use_callchain) {
 110                err = callchain_append(he->callchain, data->callchain, syms,
 111                                       data->period);
 112                if (err)
 113                        goto out_free_syms;
 114        }
 115        /*
 116         * Only in the newt browser we are doing integrated annotation,
 117         * so we don't allocated the extra space needed because the stdio
 118         * code will not use it.
 119         */
 120        if (use_browser > 0)
 121                err = hist_entry__inc_addr_samples(he, al->addr);
 122out_free_syms:
 123        free(syms);
 124        return err;
 125}
 126
 127static int add_event_total(struct perf_session *session,
 128                           struct sample_data *data,
 129                           struct perf_event_attr *attr)
 130{
 131        struct hists *hists;
 132
 133        if (attr)
 134                hists = perf_session__hists_findnew(session, data->id,
 135                                                    attr->type, attr->config);
 136        else
 137                hists = perf_session__hists_findnew(session, data->id, 0, 0);
 138
 139        if (!hists)
 140                return -ENOMEM;
 141
 142        hists->stats.total_period += data->period;
 143        /*
 144         * FIXME: add_event_total should be moved from here to
 145         * perf_session__process_event so that the proper hist is passed to
 146         * the event_op methods.
 147         */
 148        hists__inc_nr_events(hists, PERF_RECORD_SAMPLE);
 149        session->hists.stats.total_period += data->period;
 150        return 0;
 151}
 152
 153static int process_sample_event(event_t *event, struct sample_data *sample,
 154                                struct perf_session *session)
 155{
 156        struct addr_location al;
 157        struct perf_event_attr *attr;
 158
 159        if (event__preprocess_sample(event, session, &al, sample, NULL) < 0) {
 160                fprintf(stderr, "problem processing %d event, skipping it.\n",
 161                        event->header.type);
 162                return -1;
 163        }
 164
 165        if (al.filtered || (hide_unresolved && al.sym == NULL))
 166                return 0;
 167
 168        if (perf_session__add_hist_entry(session, &al, sample)) {
 169                pr_debug("problem incrementing symbol period, skipping event\n");
 170                return -1;
 171        }
 172
 173        attr = perf_header__find_attr(sample->id, &session->header);
 174
 175        if (add_event_total(session, sample, attr)) {
 176                pr_debug("problem adding event period\n");
 177                return -1;
 178        }
 179
 180        return 0;
 181}
 182
 183static int process_read_event(event_t *event, struct sample_data *sample __used,
 184                              struct perf_session *session __used)
 185{
 186        struct perf_event_attr *attr;
 187
 188        attr = perf_header__find_attr(event->read.id, &session->header);
 189
 190        if (show_threads) {
 191                const char *name = attr ? __event_name(attr->type, attr->config)
 192                                   : "unknown";
 193                perf_read_values_add_value(&show_threads_values,
 194                                           event->read.pid, event->read.tid,
 195                                           event->read.id,
 196                                           name,
 197                                           event->read.value);
 198        }
 199
 200        dump_printf(": %d %d %s %" PRIu64 "\n", event->read.pid, event->read.tid,
 201                    attr ? __event_name(attr->type, attr->config) : "FAIL",
 202                    event->read.value);
 203
 204        return 0;
 205}
 206
 207static int perf_session__setup_sample_type(struct perf_session *self)
 208{
 209        if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
 210                if (sort__has_parent) {
 211                        fprintf(stderr, "selected --sort parent, but no"
 212                                        " callchain data. Did you call"
 213                                        " perf record without -g?\n");
 214                        return -EINVAL;
 215                }
 216                if (symbol_conf.use_callchain) {
 217                        fprintf(stderr, "selected -g but no callchain data."
 218                                        " Did you call perf record without"
 219                                        " -g?\n");
 220                        return -1;
 221                }
 222        } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
 223                   !symbol_conf.use_callchain) {
 224                        symbol_conf.use_callchain = true;
 225                        if (register_callchain_param(&callchain_param) < 0) {
 226                                fprintf(stderr, "Can't register callchain"
 227                                                " params\n");
 228                                return -EINVAL;
 229                        }
 230        }
 231
 232        return 0;
 233}
 234
 235static struct perf_event_ops event_ops = {
 236        .sample = process_sample_event,
 237        .mmap   = event__process_mmap,
 238        .comm   = event__process_comm,
 239        .exit   = event__process_task,
 240        .fork   = event__process_task,
 241        .lost   = event__process_lost,
 242        .read   = process_read_event,
 243        .attr   = event__process_attr,
 244        .event_type = event__process_event_type,
 245        .tracing_data = event__process_tracing_data,
 246        .build_id = event__process_build_id,
 247        .ordered_samples = true,
 248        .ordering_requires_timestamps = true,
 249};
 250
 251extern volatile int session_done;
 252
 253static void sig_handler(int sig __used)
 254{
 255        session_done = 1;
 256}
 257
 258static size_t hists__fprintf_nr_sample_events(struct hists *self,
 259                                              const char *evname, FILE *fp)
 260{
 261        size_t ret;
 262        char unit;
 263        unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
 264
 265        nr_events = convert_unit(nr_events, &unit);
 266        ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
 267        if (evname != NULL)
 268                ret += fprintf(fp, " %s", evname);
 269        return ret + fprintf(fp, "\n#\n");
 270}
 271
 272static int hists__tty_browse_tree(struct rb_root *tree, const char *help)
 273{
 274        struct rb_node *next = rb_first(tree);
 275
 276        while (next) {
 277                struct hists *hists = rb_entry(next, struct hists, rb_node);
 278                const char *evname = NULL;
 279
 280                if (rb_first(&hists->entries) != rb_last(&hists->entries))
 281                        evname = __event_name(hists->type, hists->config);
 282
 283                hists__fprintf_nr_sample_events(hists, evname, stdout);
 284                hists__fprintf(hists, NULL, false, stdout);
 285                fprintf(stdout, "\n\n");
 286                next = rb_next(&hists->rb_node);
 287        }
 288
 289        if (sort_order == default_sort_order &&
 290            parent_pattern == default_parent_pattern) {
 291                fprintf(stdout, "#\n# (%s)\n#\n", help);
 292
 293                if (show_threads) {
 294                        bool style = !strcmp(pretty_printing_style, "raw");
 295                        perf_read_values_display(stdout, &show_threads_values,
 296                                                 style);
 297                        perf_read_values_destroy(&show_threads_values);
 298                }
 299        }
 300
 301        return 0;
 302}
 303
 304static int __cmd_report(void)
 305{
 306        int ret = -EINVAL;
 307        struct perf_session *session;
 308        struct rb_node *next;
 309        const char *help = "For a higher level overview, try: perf report --sort comm,dso";
 310
 311        signal(SIGINT, sig_handler);
 312
 313        session = perf_session__new(input_name, O_RDONLY, force, false, &event_ops);
 314        if (session == NULL)
 315                return -ENOMEM;
 316
 317        if (show_threads)
 318                perf_read_values_init(&show_threads_values);
 319
 320        ret = perf_session__setup_sample_type(session);
 321        if (ret)
 322                goto out_delete;
 323
 324        ret = perf_session__process_events(session, &event_ops);
 325        if (ret)
 326                goto out_delete;
 327
 328        if (dump_trace) {
 329                perf_session__fprintf_nr_events(session, stdout);
 330                goto out_delete;
 331        }
 332
 333        if (verbose > 3)
 334                perf_session__fprintf(session, stdout);
 335
 336        if (verbose > 2)
 337                perf_session__fprintf_dsos(session, stdout);
 338
 339        next = rb_first(&session->hists_tree);
 340        while (next) {
 341                struct hists *hists;
 342
 343                hists = rb_entry(next, struct hists, rb_node);
 344                hists__collapse_resort(hists);
 345                hists__output_resort(hists);
 346                next = rb_next(&hists->rb_node);
 347        }
 348
 349        if (use_browser > 0)
 350                hists__tui_browse_tree(&session->hists_tree, help);
 351        else
 352                hists__tty_browse_tree(&session->hists_tree, help);
 353
 354out_delete:
 355        /*
 356         * Speed up the exit process, for large files this can
 357         * take quite a while.
 358         *
 359         * XXX Enable this when using valgrind or if we ever
 360         * librarize this command.
 361         *
 362         * Also experiment with obstacks to see how much speed
 363         * up we'll get here.
 364         *
 365         * perf_session__delete(session);
 366         */
 367        return ret;
 368}
 369
 370static int
 371parse_callchain_opt(const struct option *opt __used, const char *arg,
 372                    int unset)
 373{
 374        char *tok, *tok2;
 375        char *endptr;
 376
 377        /*
 378         * --no-call-graph
 379         */
 380        if (unset) {
 381                dont_use_callchains = true;
 382                return 0;
 383        }
 384
 385        symbol_conf.use_callchain = true;
 386
 387        if (!arg)
 388                return 0;
 389
 390        tok = strtok((char *)arg, ",");
 391        if (!tok)
 392                return -1;
 393
 394        /* get the output mode */
 395        if (!strncmp(tok, "graph", strlen(arg)))
 396                callchain_param.mode = CHAIN_GRAPH_ABS;
 397
 398        else if (!strncmp(tok, "flat", strlen(arg)))
 399                callchain_param.mode = CHAIN_FLAT;
 400
 401        else if (!strncmp(tok, "fractal", strlen(arg)))
 402                callchain_param.mode = CHAIN_GRAPH_REL;
 403
 404        else if (!strncmp(tok, "none", strlen(arg))) {
 405                callchain_param.mode = CHAIN_NONE;
 406                symbol_conf.use_callchain = false;
 407
 408                return 0;
 409        }
 410
 411        else
 412                return -1;
 413
 414        /* get the min percentage */
 415        tok = strtok(NULL, ",");
 416        if (!tok)
 417                goto setup;
 418
 419        tok2 = strtok(NULL, ",");
 420        callchain_param.min_percent = strtod(tok, &endptr);
 421        if (tok == endptr)
 422                return -1;
 423
 424        if (tok2)
 425                callchain_param.print_limit = strtod(tok2, &endptr);
 426setup:
 427        if (register_callchain_param(&callchain_param) < 0) {
 428                fprintf(stderr, "Can't register callchain params\n");
 429                return -1;
 430        }
 431        return 0;
 432}
 433
 434static const char * const report_usage[] = {
 435        "perf report [<options>] <command>",
 436        NULL
 437};
 438
 439static const struct option options[] = {
 440        OPT_STRING('i', "input", &input_name, "file",
 441                    "input file name"),
 442        OPT_INCR('v', "verbose", &verbose,
 443                    "be more verbose (show symbol address, etc)"),
 444        OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
 445                    "dump raw trace in ASCII"),
 446        OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
 447                   "file", "vmlinux pathname"),
 448        OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
 449                   "file", "kallsyms pathname"),
 450        OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
 451        OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
 452                    "load module symbols - WARNING: use only with -k and LIVE kernel"),
 453        OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
 454                    "Show a column with the number of samples"),
 455        OPT_BOOLEAN('T', "threads", &show_threads,
 456                    "Show per-thread event counters"),
 457        OPT_STRING(0, "pretty", &pretty_printing_style, "key",
 458                   "pretty printing style key: normal raw"),
 459        OPT_BOOLEAN(0, "tui", &use_tui, "Use the TUI interface"),
 460        OPT_BOOLEAN(0, "stdio", &use_stdio, "Use the stdio interface"),
 461        OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
 462                   "sort by key(s): pid, comm, dso, symbol, parent"),
 463        OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
 464                    "Show sample percentage for different cpu modes"),
 465        OPT_STRING('p', "parent", &parent_pattern, "regex",
 466                   "regex filter to identify parent, see: '--sort parent'"),
 467        OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
 468                    "Only display entries with parent-match"),
 469        OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
 470                     "Display callchains using output_type (graph, flat, fractal, or none) and min percent threshold. "
 471                     "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
 472        OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
 473                   "only consider symbols in these dsos"),
 474        OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
 475                   "only consider symbols in these comms"),
 476        OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
 477                   "only consider these symbols"),
 478        OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
 479                   "width[,width...]",
 480                   "don't try to adjust column width, use these fixed values"),
 481        OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
 482                   "separator for columns, no spaces will be added between "
 483                   "columns '.' is reserved."),
 484        OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
 485                    "Only display entries resolved to a symbol"),
 486        OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
 487                    "Look for files with symbols relative to this directory"),
 488        OPT_END()
 489};
 490
 491int cmd_report(int argc, const char **argv, const char *prefix __used)
 492{
 493        argc = parse_options(argc, argv, options, report_usage, 0);
 494
 495        if (use_stdio)
 496                use_browser = 0;
 497        else if (use_tui)
 498                use_browser = 1;
 499
 500        if (strcmp(input_name, "-") != 0)
 501                setup_browser();
 502        else
 503                use_browser = 0;
 504        /*
 505         * Only in the newt browser we are doing integrated annotation,
 506         * so don't allocate extra space that won't be used in the stdio
 507         * implementation.
 508         */
 509        if (use_browser > 0) {
 510                symbol_conf.priv_size = sizeof(struct sym_priv);
 511                /*
 512                 * For searching by name on the "Browse map details".
 513                 * providing it only in verbose mode not to bloat too
 514                 * much struct symbol.
 515                 */
 516                if (verbose) {
 517                        /*
 518                         * XXX: Need to provide a less kludgy way to ask for
 519                         * more space per symbol, the u32 is for the index on
 520                         * the ui browser.
 521                         * See symbol__browser_index.
 522                         */
 523                        symbol_conf.priv_size += sizeof(u32);
 524                        symbol_conf.sort_by_name = true;
 525                }
 526        }
 527
 528        if (symbol__init() < 0)
 529                return -1;
 530
 531        setup_sorting(report_usage, options);
 532
 533        if (parent_pattern != default_parent_pattern) {
 534                if (sort_dimension__add("parent") < 0)
 535                        return -1;
 536                sort_parent.elide = 1;
 537        } else
 538                symbol_conf.exclude_other = false;
 539
 540        /*
 541         * Any (unrecognized) arguments left?
 542         */
 543        if (argc)
 544                usage_with_options(report_usage, options);
 545
 546        sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
 547        sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
 548        sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
 549
 550        return __cmd_report();
 551}
 552