perf report: Speed up exit path
[linux-2.6.git] / tools / perf / builtin-report.c
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
33 static char             const *input_name = "perf.data";
34
35 static bool             force;
36 static bool             hide_unresolved;
37 static bool             dont_use_callchains;
38
39 static bool             show_threads;
40 static struct perf_read_values  show_threads_values;
41
42 static const char       default_pretty_printing_style[] = "normal";
43 static const char       *pretty_printing_style = default_pretty_printing_style;
44
45 static char             callchain_default_opt[] = "fractal,0.5";
46
47 static 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
80 static 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 = append_chain(he->callchain, data->callchain, syms, data->period);
111                 if (err)
112                         goto out_free_syms;
113         }
114         /*
115          * Only in the newt browser we are doing integrated annotation,
116          * so we don't allocated the extra space needed because the stdio
117          * code will not use it.
118          */
119         if (use_browser > 0)
120                 err = hist_entry__inc_addr_samples(he, al->addr);
121 out_free_syms:
122         free(syms);
123         return err;
124 }
125
126 static int add_event_total(struct perf_session *session,
127                            struct sample_data *data,
128                            struct perf_event_attr *attr)
129 {
130         struct hists *hists;
131
132         if (attr)
133                 hists = perf_session__hists_findnew(session, data->id,
134                                                     attr->type, attr->config);
135         else
136                 hists = perf_session__hists_findnew(session, data->id, 0, 0);
137
138         if (!hists)
139                 return -ENOMEM;
140
141         hists->stats.total_period += data->period;
142         /*
143          * FIXME: add_event_total should be moved from here to
144          * perf_session__process_event so that the proper hist is passed to
145          * the event_op methods.
146          */
147         hists__inc_nr_events(hists, PERF_RECORD_SAMPLE);
148         session->hists.stats.total_period += data->period;
149         return 0;
150 }
151
152 static int process_sample_event(event_t *event, struct perf_session *session)
153 {
154         struct sample_data data = { .period = 1, };
155         struct addr_location al;
156         struct perf_event_attr *attr;
157
158         if (event__preprocess_sample(event, session, &al, &data, NULL) < 0) {
159                 fprintf(stderr, "problem processing %d event, skipping it.\n",
160                         event->header.type);
161                 return -1;
162         }
163
164         if (al.filtered || (hide_unresolved && al.sym == NULL))
165                 return 0;
166
167         if (perf_session__add_hist_entry(session, &al, &data)) {
168                 pr_debug("problem incrementing symbol period, skipping event\n");
169                 return -1;
170         }
171
172         attr = perf_header__find_attr(data.id, &session->header);
173
174         if (add_event_total(session, &data, attr)) {
175                 pr_debug("problem adding event period\n");
176                 return -1;
177         }
178
179         return 0;
180 }
181
182 static int process_read_event(event_t *event, struct perf_session *session __used)
183 {
184         struct perf_event_attr *attr;
185
186         attr = perf_header__find_attr(event->read.id, &session->header);
187
188         if (show_threads) {
189                 const char *name = attr ? __event_name(attr->type, attr->config)
190                                    : "unknown";
191                 perf_read_values_add_value(&show_threads_values,
192                                            event->read.pid, event->read.tid,
193                                            event->read.id,
194                                            name,
195                                            event->read.value);
196         }
197
198         dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
199                     attr ? __event_name(attr->type, attr->config) : "FAIL",
200                     event->read.value);
201
202         return 0;
203 }
204
205 static int perf_session__setup_sample_type(struct perf_session *self)
206 {
207         if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
208                 if (sort__has_parent) {
209                         fprintf(stderr, "selected --sort parent, but no"
210                                         " callchain data. Did you call"
211                                         " perf record without -g?\n");
212                         return -EINVAL;
213                 }
214                 if (symbol_conf.use_callchain) {
215                         fprintf(stderr, "selected -g but no callchain data."
216                                         " Did you call perf record without"
217                                         " -g?\n");
218                         return -1;
219                 }
220         } else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
221                    !symbol_conf.use_callchain) {
222                         symbol_conf.use_callchain = true;
223                         if (register_callchain_param(&callchain_param) < 0) {
224                                 fprintf(stderr, "Can't register callchain"
225                                                 " params\n");
226                                 return -EINVAL;
227                         }
228         }
229
230         return 0;
231 }
232
233 static struct perf_event_ops event_ops = {
234         .sample = process_sample_event,
235         .mmap   = event__process_mmap,
236         .comm   = event__process_comm,
237         .exit   = event__process_task,
238         .fork   = event__process_task,
239         .lost   = event__process_lost,
240         .read   = process_read_event,
241         .attr   = event__process_attr,
242         .event_type = event__process_event_type,
243         .tracing_data = event__process_tracing_data,
244         .build_id = event__process_build_id,
245 };
246
247 extern volatile int session_done;
248
249 static void sig_handler(int sig __used)
250 {
251         session_done = 1;
252 }
253
254 static size_t hists__fprintf_nr_sample_events(struct hists *self,
255                                               const char *evname, FILE *fp)
256 {
257         size_t ret;
258         char unit;
259         unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
260
261         nr_events = convert_unit(nr_events, &unit);
262         ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
263         if (evname != NULL)
264                 ret += fprintf(fp, " %s", evname);
265         return ret + fprintf(fp, "\n#\n");
266 }
267
268 static int hists__tty_browse_tree(struct rb_root *tree, const char *help)
269 {
270         struct rb_node *next = rb_first(tree);
271
272         while (next) {
273                 struct hists *hists = rb_entry(next, struct hists, rb_node);
274                 const char *evname = NULL;
275
276                 if (rb_first(&hists->entries) != rb_last(&hists->entries))
277                         evname = __event_name(hists->type, hists->config);
278
279                 hists__fprintf_nr_sample_events(hists, evname, stdout);
280                 hists__fprintf(hists, NULL, false, stdout);
281                 fprintf(stdout, "\n\n");
282                 next = rb_next(&hists->rb_node);
283         }
284
285         if (sort_order == default_sort_order &&
286             parent_pattern == default_parent_pattern) {
287                 fprintf(stdout, "#\n# (%s)\n#\n", help);
288
289                 if (show_threads) {
290                         bool style = !strcmp(pretty_printing_style, "raw");
291                         perf_read_values_display(stdout, &show_threads_values,
292                                                  style);
293                         perf_read_values_destroy(&show_threads_values);
294                 }
295         }
296
297         return 0;
298 }
299
300 static int __cmd_report(void)
301 {
302         int ret = -EINVAL;
303         struct perf_session *session;
304         struct rb_node *next;
305         const char *help = "For a higher level overview, try: perf report --sort comm,dso";
306
307         signal(SIGINT, sig_handler);
308
309         session = perf_session__new(input_name, O_RDONLY, force, false);
310         if (session == NULL)
311                 return -ENOMEM;
312
313         if (show_threads)
314                 perf_read_values_init(&show_threads_values);
315
316         ret = perf_session__setup_sample_type(session);
317         if (ret)
318                 goto out_delete;
319
320         ret = perf_session__process_events(session, &event_ops);
321         if (ret)
322                 goto out_delete;
323
324         if (dump_trace) {
325                 perf_session__fprintf_nr_events(session, stdout);
326                 goto out_delete;
327         }
328
329         if (verbose > 3)
330                 perf_session__fprintf(session, stdout);
331
332         if (verbose > 2)
333                 perf_session__fprintf_dsos(session, stdout);
334
335         next = rb_first(&session->hists_tree);
336         while (next) {
337                 struct hists *hists;
338
339                 hists = rb_entry(next, struct hists, rb_node);
340                 hists__collapse_resort(hists);
341                 hists__output_resort(hists);
342                 next = rb_next(&hists->rb_node);
343         }
344
345         if (use_browser > 0)
346                 hists__tui_browse_tree(&session->hists_tree, help);
347         else
348                 hists__tty_browse_tree(&session->hists_tree, help);
349
350 out_delete:
351         /*
352          * Speed up the exit process, for large files this can
353          * take quite a while.
354          *
355          * XXX Enable this when using valgrind or if we ever
356          * librarize this command.
357          *
358          * Also experiment with obstacks to see how much speed
359          * up we'll get here.
360          *
361          * perf_session__delete(session);
362          */
363         return ret;
364 }
365
366 static int
367 parse_callchain_opt(const struct option *opt __used, const char *arg,
368                     int unset)
369 {
370         char *tok, *tok2;
371         char *endptr;
372
373         /*
374          * --no-call-graph
375          */
376         if (unset) {
377                 dont_use_callchains = true;
378                 return 0;
379         }
380
381         symbol_conf.use_callchain = true;
382
383         if (!arg)
384                 return 0;
385
386         tok = strtok((char *)arg, ",");
387         if (!tok)
388                 return -1;
389
390         /* get the output mode */
391         if (!strncmp(tok, "graph", strlen(arg)))
392                 callchain_param.mode = CHAIN_GRAPH_ABS;
393
394         else if (!strncmp(tok, "flat", strlen(arg)))
395                 callchain_param.mode = CHAIN_FLAT;
396
397         else if (!strncmp(tok, "fractal", strlen(arg)))
398                 callchain_param.mode = CHAIN_GRAPH_REL;
399
400         else if (!strncmp(tok, "none", strlen(arg))) {
401                 callchain_param.mode = CHAIN_NONE;
402                 symbol_conf.use_callchain = false;
403
404                 return 0;
405         }
406
407         else
408                 return -1;
409
410         /* get the min percentage */
411         tok = strtok(NULL, ",");
412         if (!tok)
413                 goto setup;
414
415         tok2 = strtok(NULL, ",");
416         callchain_param.min_percent = strtod(tok, &endptr);
417         if (tok == endptr)
418                 return -1;
419
420         if (tok2)
421                 callchain_param.print_limit = strtod(tok2, &endptr);
422 setup:
423         if (register_callchain_param(&callchain_param) < 0) {
424                 fprintf(stderr, "Can't register callchain params\n");
425                 return -1;
426         }
427         return 0;
428 }
429
430 static const char * const report_usage[] = {
431         "perf report [<options>] <command>",
432         NULL
433 };
434
435 static const struct option options[] = {
436         OPT_STRING('i', "input", &input_name, "file",
437                     "input file name"),
438         OPT_INCR('v', "verbose", &verbose,
439                     "be more verbose (show symbol address, etc)"),
440         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
441                     "dump raw trace in ASCII"),
442         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
443                    "file", "vmlinux pathname"),
444         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
445         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
446                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
447         OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
448                     "Show a column with the number of samples"),
449         OPT_BOOLEAN('T', "threads", &show_threads,
450                     "Show per-thread event counters"),
451         OPT_STRING(0, "pretty", &pretty_printing_style, "key",
452                    "pretty printing style key: normal raw"),
453         OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
454                    "sort by key(s): pid, comm, dso, symbol, parent"),
455         OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
456                     "Show sample percentage for different cpu modes"),
457         OPT_STRING('p', "parent", &parent_pattern, "regex",
458                    "regex filter to identify parent, see: '--sort parent'"),
459         OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
460                     "Only display entries with parent-match"),
461         OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
462                      "Display callchains using output_type (graph, flat, fractal, or none) and min percent threshold. "
463                      "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
464         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
465                    "only consider symbols in these dsos"),
466         OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
467                    "only consider symbols in these comms"),
468         OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
469                    "only consider these symbols"),
470         OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
471                    "width[,width...]",
472                    "don't try to adjust column width, use these fixed values"),
473         OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
474                    "separator for columns, no spaces will be added between "
475                    "columns '.' is reserved."),
476         OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
477                     "Only display entries resolved to a symbol"),
478         OPT_END()
479 };
480
481 int cmd_report(int argc, const char **argv, const char *prefix __used)
482 {
483         argc = parse_options(argc, argv, options, report_usage, 0);
484
485         if (strcmp(input_name, "-") != 0)
486                 setup_browser();
487         /*
488          * Only in the newt browser we are doing integrated annotation,
489          * so don't allocate extra space that won't be used in the stdio
490          * implementation.
491          */
492         if (use_browser > 0) {
493                 symbol_conf.priv_size = sizeof(struct sym_priv);
494                 /*
495                  * For searching by name on the "Browse map details".
496                  * providing it only in verbose mode not to bloat too
497                  * much struct symbol.
498                  */
499                 if (verbose) {
500                         /*
501                          * XXX: Need to provide a less kludgy way to ask for
502                          * more space per symbol, the u32 is for the index on
503                          * the ui browser.
504                          * See symbol__browser_index.
505                          */
506                         symbol_conf.priv_size += sizeof(u32);
507                         symbol_conf.sort_by_name = true;
508                 }
509         }
510
511         if (symbol__init() < 0)
512                 return -1;
513
514         setup_sorting(report_usage, options);
515
516         if (parent_pattern != default_parent_pattern) {
517                 if (sort_dimension__add("parent") < 0)
518                         return -1;
519                 sort_parent.elide = 1;
520         } else
521                 symbol_conf.exclude_other = false;
522
523         /*
524          * Any (unrecognized) arguments left?
525          */
526         if (argc)
527                 usage_with_options(report_usage, options);
528
529         sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
530         sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
531         sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
532
533         return __cmd_report();
534 }