perf report: Add infrastructure for a cycles histogram
[linux-block.git] / tools / perf / builtin-annotate.c
1 /*
2  * builtin-annotate.c
3  *
4  * Builtin annotate 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 #include "util/color.h"
12 #include <linux/list.h>
13 #include "util/cache.h"
14 #include <linux/rbtree.h>
15 #include "util/symbol.h"
16
17 #include "perf.h"
18 #include "util/debug.h"
19
20 #include "util/evlist.h"
21 #include "util/evsel.h"
22 #include "util/annotate.h"
23 #include "util/event.h"
24 #include "util/parse-options.h"
25 #include "util/parse-events.h"
26 #include "util/thread.h"
27 #include "util/sort.h"
28 #include "util/hist.h"
29 #include "util/session.h"
30 #include "util/tool.h"
31 #include "util/data.h"
32 #include "arch/common.h"
33
34 #include <dlfcn.h>
35 #include <linux/bitmap.h>
36
37 struct perf_annotate {
38         struct perf_tool tool;
39         struct perf_session *session;
40         bool       use_tui, use_stdio, use_gtk;
41         bool       full_paths;
42         bool       print_line;
43         bool       skip_missing;
44         const char *sym_hist_filter;
45         const char *cpu_list;
46         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
47 };
48
49 static int perf_evsel__add_sample(struct perf_evsel *evsel,
50                                   struct perf_sample *sample __maybe_unused,
51                                   struct addr_location *al,
52                                   struct perf_annotate *ann)
53 {
54         struct hists *hists = evsel__hists(evsel);
55         struct hist_entry *he;
56         int ret;
57
58         if (ann->sym_hist_filter != NULL &&
59             (al->sym == NULL ||
60              strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
61                 /* We're only interested in a symbol named sym_hist_filter */
62                 /*
63                  * FIXME: why isn't this done in the symbol_filter when loading
64                  * the DSO?
65                  */
66                 if (al->sym != NULL) {
67                         rb_erase(&al->sym->rb_node,
68                                  &al->map->dso->symbols[al->map->type]);
69                         symbol__delete(al->sym);
70                 }
71                 return 0;
72         }
73
74         he = __hists__add_entry(hists, al, NULL, NULL, NULL, 1, 1, 0, true);
75         if (he == NULL)
76                 return -ENOMEM;
77
78         ret = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
79         hists__inc_nr_samples(hists, true);
80         return ret;
81 }
82
83 static int process_sample_event(struct perf_tool *tool,
84                                 union perf_event *event,
85                                 struct perf_sample *sample,
86                                 struct perf_evsel *evsel,
87                                 struct machine *machine)
88 {
89         struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
90         struct addr_location al;
91         int ret = 0;
92
93         if (perf_event__preprocess_sample(event, machine, &al, sample) < 0) {
94                 pr_warning("problem processing %d event, skipping it.\n",
95                            event->header.type);
96                 return -1;
97         }
98
99         if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
100                 goto out_put;
101
102         if (!al.filtered && perf_evsel__add_sample(evsel, sample, &al, ann)) {
103                 pr_warning("problem incrementing symbol count, "
104                            "skipping event\n");
105                 ret = -1;
106         }
107 out_put:
108         addr_location__put(&al);
109         return ret;
110 }
111
112 static int hist_entry__tty_annotate(struct hist_entry *he,
113                                     struct perf_evsel *evsel,
114                                     struct perf_annotate *ann)
115 {
116         return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel,
117                                     ann->print_line, ann->full_paths, 0, 0);
118 }
119
120 static void hists__find_annotations(struct hists *hists,
121                                     struct perf_evsel *evsel,
122                                     struct perf_annotate *ann)
123 {
124         struct rb_node *nd = rb_first(&hists->entries), *next;
125         int key = K_RIGHT;
126
127         while (nd) {
128                 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
129                 struct annotation *notes;
130
131                 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
132                         goto find_next;
133
134                 notes = symbol__annotation(he->ms.sym);
135                 if (notes->src == NULL) {
136 find_next:
137                         if (key == K_LEFT)
138                                 nd = rb_prev(nd);
139                         else
140                                 nd = rb_next(nd);
141                         continue;
142                 }
143
144                 if (use_browser == 2) {
145                         int ret;
146                         int (*annotate)(struct hist_entry *he,
147                                         struct perf_evsel *evsel,
148                                         struct hist_browser_timer *hbt);
149
150                         annotate = dlsym(perf_gtk_handle,
151                                          "hist_entry__gtk_annotate");
152                         if (annotate == NULL) {
153                                 ui__error("GTK browser not found!\n");
154                                 return;
155                         }
156
157                         ret = annotate(he, evsel, NULL);
158                         if (!ret || !ann->skip_missing)
159                                 return;
160
161                         /* skip missing symbols */
162                         nd = rb_next(nd);
163                 } else if (use_browser == 1) {
164                         key = hist_entry__tui_annotate(he, evsel, NULL);
165                         switch (key) {
166                         case -1:
167                                 if (!ann->skip_missing)
168                                         return;
169                                 /* fall through */
170                         case K_RIGHT:
171                                 next = rb_next(nd);
172                                 break;
173                         case K_LEFT:
174                                 next = rb_prev(nd);
175                                 break;
176                         default:
177                                 return;
178                         }
179
180                         if (next != NULL)
181                                 nd = next;
182                 } else {
183                         hist_entry__tty_annotate(he, evsel, ann);
184                         nd = rb_next(nd);
185                         /*
186                          * Since we have a hist_entry per IP for the same
187                          * symbol, free he->ms.sym->src to signal we already
188                          * processed this symbol.
189                          */
190                         zfree(&notes->src->cycles_hist);
191                         zfree(&notes->src);
192                 }
193         }
194 }
195
196 static int __cmd_annotate(struct perf_annotate *ann)
197 {
198         int ret;
199         struct perf_session *session = ann->session;
200         struct perf_evsel *pos;
201         u64 total_nr_samples;
202
203         machines__set_symbol_filter(&session->machines, symbol__annotate_init);
204
205         if (ann->cpu_list) {
206                 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
207                                                ann->cpu_bitmap);
208                 if (ret)
209                         goto out;
210         }
211
212         if (!objdump_path) {
213                 ret = perf_session_env__lookup_objdump(&session->header.env);
214                 if (ret)
215                         goto out;
216         }
217
218         ret = perf_session__process_events(session);
219         if (ret)
220                 goto out;
221
222         if (dump_trace) {
223                 perf_session__fprintf_nr_events(session, stdout);
224                 perf_evlist__fprintf_nr_events(session->evlist, stdout);
225                 goto out;
226         }
227
228         if (verbose > 3)
229                 perf_session__fprintf(session, stdout);
230
231         if (verbose > 2)
232                 perf_session__fprintf_dsos(session, stdout);
233
234         total_nr_samples = 0;
235         evlist__for_each(session->evlist, pos) {
236                 struct hists *hists = evsel__hists(pos);
237                 u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
238
239                 if (nr_samples > 0) {
240                         total_nr_samples += nr_samples;
241                         hists__collapse_resort(hists, NULL);
242                         hists__output_resort(hists, NULL);
243
244                         if (symbol_conf.event_group &&
245                             !perf_evsel__is_group_leader(pos))
246                                 continue;
247
248                         hists__find_annotations(hists, pos, ann);
249                 }
250         }
251
252         if (total_nr_samples == 0) {
253                 ui__error("The %s file has no samples!\n", session->file->path);
254                 goto out;
255         }
256
257         if (use_browser == 2) {
258                 void (*show_annotations)(void);
259
260                 show_annotations = dlsym(perf_gtk_handle,
261                                          "perf_gtk__show_annotations");
262                 if (show_annotations == NULL) {
263                         ui__error("GTK browser not found!\n");
264                         goto out;
265                 }
266                 show_annotations();
267         }
268
269 out:
270         return ret;
271 }
272
273 static const char * const annotate_usage[] = {
274         "perf annotate [<options>]",
275         NULL
276 };
277
278 int cmd_annotate(int argc, const char **argv, const char *prefix __maybe_unused)
279 {
280         struct perf_annotate annotate = {
281                 .tool = {
282                         .sample = process_sample_event,
283                         .mmap   = perf_event__process_mmap,
284                         .mmap2  = perf_event__process_mmap2,
285                         .comm   = perf_event__process_comm,
286                         .exit   = perf_event__process_exit,
287                         .fork   = perf_event__process_fork,
288                         .ordered_events = true,
289                         .ordering_requires_timestamps = true,
290                 },
291         };
292         struct perf_data_file file = {
293                 .mode  = PERF_DATA_MODE_READ,
294         };
295         const struct option options[] = {
296         OPT_STRING('i', "input", &input_name, "file",
297                     "input file name"),
298         OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
299                    "only consider symbols in these dsos"),
300         OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
301                     "symbol to annotate"),
302         OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
303         OPT_INCR('v', "verbose", &verbose,
304                     "be more verbose (show symbol address, etc)"),
305         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
306                     "dump raw trace in ASCII"),
307         OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
308         OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
309         OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
310         OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
311                    "file", "vmlinux pathname"),
312         OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
313                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
314         OPT_BOOLEAN('l', "print-line", &annotate.print_line,
315                     "print matching source lines (may be slow)"),
316         OPT_BOOLEAN('P', "full-paths", &annotate.full_paths,
317                     "Don't shorten the displayed pathnames"),
318         OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
319                     "Skip symbols that cannot be annotated"),
320         OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
321         OPT_STRING(0, "symfs", &symbol_conf.symfs, "directory",
322                    "Look for files with symbols relative to this directory"),
323         OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
324                     "Interleave source code with assembly code (default)"),
325         OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
326                     "Display raw encoding of assembly instructions (default)"),
327         OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
328                    "Specify disassembler style (e.g. -M intel for intel syntax)"),
329         OPT_STRING(0, "objdump", &objdump_path, "path",
330                    "objdump binary to use for disassembly and annotations"),
331         OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
332                     "Show event group information together"),
333         OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
334                     "Show a column with the sum of periods"),
335         OPT_END()
336         };
337         int ret = hists__init();
338
339         if (ret < 0)
340                 return ret;
341
342         argc = parse_options(argc, argv, options, annotate_usage, 0);
343
344         if (annotate.use_stdio)
345                 use_browser = 0;
346         else if (annotate.use_tui)
347                 use_browser = 1;
348         else if (annotate.use_gtk)
349                 use_browser = 2;
350
351         file.path  = input_name;
352
353         setup_browser(true);
354
355         annotate.session = perf_session__new(&file, false, &annotate.tool);
356         if (annotate.session == NULL)
357                 return -1;
358
359         symbol_conf.priv_size = sizeof(struct annotation);
360         symbol_conf.try_vmlinux_path = true;
361
362         ret = symbol__init(&annotate.session->header.env);
363         if (ret < 0)
364                 goto out_delete;
365
366         if (setup_sorting() < 0)
367                 usage_with_options(annotate_usage, options);
368
369         if (argc) {
370                 /*
371                  * Special case: if there's an argument left then assume that
372                  * it's a symbol filter:
373                  */
374                 if (argc > 1)
375                         usage_with_options(annotate_usage, options);
376
377                 annotate.sym_hist_filter = argv[0];
378         }
379
380         ret = __cmd_annotate(&annotate);
381
382 out_delete:
383         /*
384          * Speed up the exit process, for large files this can
385          * take quite a while.
386          *
387          * XXX Enable this when using valgrind or if we ever
388          * librarize this command.
389          *
390          * Also experiment with obstacks to see how much speed
391          * up we'll get here.
392          *
393          * perf_session__delete(session);
394          */
395         return ret;
396 }