Loading...
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 <subcmd/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#include "util/block-range.h"
34
35#include <dlfcn.h>
36#include <linux/bitmap.h>
37
38struct perf_annotate {
39 struct perf_tool tool;
40 struct perf_session *session;
41 bool use_tui, use_stdio, use_gtk;
42 bool full_paths;
43 bool print_line;
44 bool skip_missing;
45 const char *sym_hist_filter;
46 const char *cpu_list;
47 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
48};
49
50/*
51 * Given one basic block:
52 *
53 * from to branch_i
54 * * ----> *
55 * |
56 * | block
57 * v
58 * * ----> *
59 * from to branch_i+1
60 *
61 * where the horizontal are the branches and the vertical is the executed
62 * block of instructions.
63 *
64 * We count, for each 'instruction', the number of blocks that covered it as
65 * well as count the ratio each branch is taken.
66 *
67 * We can do this without knowing the actual instruction stream by keeping
68 * track of the address ranges. We break down ranges such that there is no
69 * overlap and iterate from the start until the end.
70 *
71 * @acme: once we parse the objdump output _before_ processing the samples,
72 * we can easily fold the branch.cycles IPC bits in.
73 */
74static void process_basic_block(struct addr_map_symbol *start,
75 struct addr_map_symbol *end,
76 struct branch_flags *flags)
77{
78 struct symbol *sym = start->sym;
79 struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
80 struct block_range_iter iter;
81 struct block_range *entry;
82
83 /*
84 * Sanity; NULL isn't executable and the CPU cannot execute backwards
85 */
86 if (!start->addr || start->addr > end->addr)
87 return;
88
89 iter = block_range__create(start->addr, end->addr);
90 if (!block_range_iter__valid(&iter))
91 return;
92
93 /*
94 * First block in range is a branch target.
95 */
96 entry = block_range_iter(&iter);
97 assert(entry->is_target);
98 entry->entry++;
99
100 do {
101 entry = block_range_iter(&iter);
102
103 entry->coverage++;
104 entry->sym = sym;
105
106 if (notes)
107 notes->max_coverage = max(notes->max_coverage, entry->coverage);
108
109 } while (block_range_iter__next(&iter));
110
111 /*
112 * Last block in rage is a branch.
113 */
114 entry = block_range_iter(&iter);
115 assert(entry->is_branch);
116 entry->taken++;
117 if (flags->predicted)
118 entry->pred++;
119}
120
121static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
122 struct perf_sample *sample)
123{
124 struct addr_map_symbol *prev = NULL;
125 struct branch_info *bi;
126 int i;
127
128 if (!bs || !bs->nr)
129 return;
130
131 bi = sample__resolve_bstack(sample, al);
132 if (!bi)
133 return;
134
135 for (i = bs->nr - 1; i >= 0; i--) {
136 /*
137 * XXX filter against symbol
138 */
139 if (prev)
140 process_basic_block(prev, &bi[i].from, &bi[i].flags);
141 prev = &bi[i].to;
142 }
143
144 free(bi);
145}
146
147static int perf_evsel__add_sample(struct perf_evsel *evsel,
148 struct perf_sample *sample,
149 struct addr_location *al,
150 struct perf_annotate *ann)
151{
152 struct hists *hists = evsel__hists(evsel);
153 struct hist_entry *he;
154 int ret;
155
156 if (ann->sym_hist_filter != NULL &&
157 (al->sym == NULL ||
158 strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
159 /* We're only interested in a symbol named sym_hist_filter */
160 /*
161 * FIXME: why isn't this done in the symbol_filter when loading
162 * the DSO?
163 */
164 if (al->sym != NULL) {
165 rb_erase(&al->sym->rb_node,
166 &al->map->dso->symbols[al->map->type]);
167 symbol__delete(al->sym);
168 dso__reset_find_symbol_cache(al->map->dso);
169 }
170 return 0;
171 }
172
173 /*
174 * XXX filtered samples can still have branch entires pointing into our
175 * symbol and are missed.
176 */
177 process_branch_stack(sample->branch_stack, al, sample);
178
179 sample->period = 1;
180 sample->weight = 1;
181
182 he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
183 if (he == NULL)
184 return -ENOMEM;
185
186 ret = hist_entry__inc_addr_samples(he, evsel->idx, al->addr);
187 hists__inc_nr_samples(hists, true);
188 return ret;
189}
190
191static int process_sample_event(struct perf_tool *tool,
192 union perf_event *event,
193 struct perf_sample *sample,
194 struct perf_evsel *evsel,
195 struct machine *machine)
196{
197 struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
198 struct addr_location al;
199 int ret = 0;
200
201 if (machine__resolve(machine, &al, sample) < 0) {
202 pr_warning("problem processing %d event, skipping it.\n",
203 event->header.type);
204 return -1;
205 }
206
207 if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
208 goto out_put;
209
210 if (!al.filtered && perf_evsel__add_sample(evsel, sample, &al, ann)) {
211 pr_warning("problem incrementing symbol count, "
212 "skipping event\n");
213 ret = -1;
214 }
215out_put:
216 addr_location__put(&al);
217 return ret;
218}
219
220static int hist_entry__tty_annotate(struct hist_entry *he,
221 struct perf_evsel *evsel,
222 struct perf_annotate *ann)
223{
224 return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel,
225 ann->print_line, ann->full_paths, 0, 0);
226}
227
228static void hists__find_annotations(struct hists *hists,
229 struct perf_evsel *evsel,
230 struct perf_annotate *ann)
231{
232 struct rb_node *nd = rb_first(&hists->entries), *next;
233 int key = K_RIGHT;
234
235 while (nd) {
236 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
237 struct annotation *notes;
238
239 if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
240 goto find_next;
241
242 notes = symbol__annotation(he->ms.sym);
243 if (notes->src == NULL) {
244find_next:
245 if (key == K_LEFT)
246 nd = rb_prev(nd);
247 else
248 nd = rb_next(nd);
249 continue;
250 }
251
252 if (use_browser == 2) {
253 int ret;
254 int (*annotate)(struct hist_entry *he,
255 struct perf_evsel *evsel,
256 struct hist_browser_timer *hbt);
257
258 annotate = dlsym(perf_gtk_handle,
259 "hist_entry__gtk_annotate");
260 if (annotate == NULL) {
261 ui__error("GTK browser not found!\n");
262 return;
263 }
264
265 ret = annotate(he, evsel, NULL);
266 if (!ret || !ann->skip_missing)
267 return;
268
269 /* skip missing symbols */
270 nd = rb_next(nd);
271 } else if (use_browser == 1) {
272 key = hist_entry__tui_annotate(he, evsel, NULL);
273 switch (key) {
274 case -1:
275 if (!ann->skip_missing)
276 return;
277 /* fall through */
278 case K_RIGHT:
279 next = rb_next(nd);
280 break;
281 case K_LEFT:
282 next = rb_prev(nd);
283 break;
284 default:
285 return;
286 }
287
288 if (next != NULL)
289 nd = next;
290 } else {
291 hist_entry__tty_annotate(he, evsel, ann);
292 nd = rb_next(nd);
293 /*
294 * Since we have a hist_entry per IP for the same
295 * symbol, free he->ms.sym->src to signal we already
296 * processed this symbol.
297 */
298 zfree(¬es->src->cycles_hist);
299 zfree(¬es->src);
300 }
301 }
302}
303
304static int __cmd_annotate(struct perf_annotate *ann)
305{
306 int ret;
307 struct perf_session *session = ann->session;
308 struct perf_evsel *pos;
309 u64 total_nr_samples;
310
311 if (ann->cpu_list) {
312 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
313 ann->cpu_bitmap);
314 if (ret)
315 goto out;
316 }
317
318 if (!objdump_path) {
319 ret = perf_env__lookup_objdump(&session->header.env);
320 if (ret)
321 goto out;
322 }
323
324 ret = perf_session__process_events(session);
325 if (ret)
326 goto out;
327
328 if (dump_trace) {
329 perf_session__fprintf_nr_events(session, stdout);
330 perf_evlist__fprintf_nr_events(session->evlist, stdout);
331 goto out;
332 }
333
334 if (verbose > 3)
335 perf_session__fprintf(session, stdout);
336
337 if (verbose > 2)
338 perf_session__fprintf_dsos(session, stdout);
339
340 total_nr_samples = 0;
341 evlist__for_each_entry(session->evlist, pos) {
342 struct hists *hists = evsel__hists(pos);
343 u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
344
345 if (nr_samples > 0) {
346 total_nr_samples += nr_samples;
347 hists__collapse_resort(hists, NULL);
348 /* Don't sort callchain */
349 perf_evsel__reset_sample_bit(pos, CALLCHAIN);
350 perf_evsel__output_resort(pos, NULL);
351
352 if (symbol_conf.event_group &&
353 !perf_evsel__is_group_leader(pos))
354 continue;
355
356 hists__find_annotations(hists, pos, ann);
357 }
358 }
359
360 if (total_nr_samples == 0) {
361 ui__error("The %s file has no samples!\n", session->file->path);
362 goto out;
363 }
364
365 if (use_browser == 2) {
366 void (*show_annotations)(void);
367
368 show_annotations = dlsym(perf_gtk_handle,
369 "perf_gtk__show_annotations");
370 if (show_annotations == NULL) {
371 ui__error("GTK browser not found!\n");
372 goto out;
373 }
374 show_annotations();
375 }
376
377out:
378 return ret;
379}
380
381static const char * const annotate_usage[] = {
382 "perf annotate [<options>]",
383 NULL
384};
385
386int cmd_annotate(int argc, const char **argv, const char *prefix __maybe_unused)
387{
388 struct perf_annotate annotate = {
389 .tool = {
390 .sample = process_sample_event,
391 .mmap = perf_event__process_mmap,
392 .mmap2 = perf_event__process_mmap2,
393 .comm = perf_event__process_comm,
394 .exit = perf_event__process_exit,
395 .fork = perf_event__process_fork,
396 .ordered_events = true,
397 .ordering_requires_timestamps = true,
398 },
399 };
400 struct perf_data_file file = {
401 .mode = PERF_DATA_MODE_READ,
402 };
403 const struct option options[] = {
404 OPT_STRING('i', "input", &input_name, "file",
405 "input file name"),
406 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
407 "only consider symbols in these dsos"),
408 OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
409 "symbol to annotate"),
410 OPT_BOOLEAN('f', "force", &file.force, "don't complain, do it"),
411 OPT_INCR('v', "verbose", &verbose,
412 "be more verbose (show symbol address, etc)"),
413 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
414 "dump raw trace in ASCII"),
415 OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
416 OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
417 OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
418 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
419 "file", "vmlinux pathname"),
420 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
421 "load module symbols - WARNING: use only with -k and LIVE kernel"),
422 OPT_BOOLEAN('l', "print-line", &annotate.print_line,
423 "print matching source lines (may be slow)"),
424 OPT_BOOLEAN('P', "full-paths", &annotate.full_paths,
425 "Don't shorten the displayed pathnames"),
426 OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
427 "Skip symbols that cannot be annotated"),
428 OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
429 OPT_CALLBACK(0, "symfs", NULL, "directory",
430 "Look for files with symbols relative to this directory",
431 symbol__config_symfs),
432 OPT_BOOLEAN(0, "source", &symbol_conf.annotate_src,
433 "Interleave source code with assembly code (default)"),
434 OPT_BOOLEAN(0, "asm-raw", &symbol_conf.annotate_asm_raw,
435 "Display raw encoding of assembly instructions (default)"),
436 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
437 "Specify disassembler style (e.g. -M intel for intel syntax)"),
438 OPT_STRING(0, "objdump", &objdump_path, "path",
439 "objdump binary to use for disassembly and annotations"),
440 OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
441 "Show event group information together"),
442 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
443 "Show a column with the sum of periods"),
444 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
445 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
446 stdio__config_color, "always"),
447 OPT_END()
448 };
449 int ret = hists__init();
450
451 if (ret < 0)
452 return ret;
453
454 argc = parse_options(argc, argv, options, annotate_usage, 0);
455 if (argc) {
456 /*
457 * Special case: if there's an argument left then assume that
458 * it's a symbol filter:
459 */
460 if (argc > 1)
461 usage_with_options(annotate_usage, options);
462
463 annotate.sym_hist_filter = argv[0];
464 }
465
466 file.path = input_name;
467
468 annotate.session = perf_session__new(&file, false, &annotate.tool);
469 if (annotate.session == NULL)
470 return -1;
471
472 ret = symbol__annotation_init();
473 if (ret < 0)
474 goto out_delete;
475
476 symbol_conf.try_vmlinux_path = true;
477
478 ret = symbol__init(&annotate.session->header.env);
479 if (ret < 0)
480 goto out_delete;
481
482 if (setup_sorting(NULL) < 0)
483 usage_with_options(annotate_usage, options);
484
485 if (annotate.use_stdio)
486 use_browser = 0;
487 else if (annotate.use_tui)
488 use_browser = 1;
489 else if (annotate.use_gtk)
490 use_browser = 2;
491
492 setup_browser(true);
493
494 ret = __cmd_annotate(&annotate);
495
496out_delete:
497 /*
498 * Speed up the exit process, for large files this can
499 * take quite a while.
500 *
501 * XXX Enable this when using valgrind or if we ever
502 * librarize this command.
503 *
504 * Also experiment with obstacks to see how much speed
505 * up we'll get here.
506 *
507 * perf_session__delete(session);
508 */
509 return ret;
510}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * builtin-annotate.c
4 *
5 * Builtin annotate command: Analyze the perf.data input file,
6 * look up and read DSOs and symbol information and display
7 * a histogram of results, along various sorting keys.
8 */
9#include "builtin.h"
10
11#include "util/color.h"
12#include <linux/list.h>
13#include "util/cache.h"
14#include <linux/rbtree.h>
15#include <linux/zalloc.h>
16#include "util/symbol.h"
17
18#include "util/debug.h"
19
20#include "util/evlist.h"
21#include "util/evsel.h"
22#include "util/annotate.h"
23#include "util/annotate-data.h"
24#include "util/event.h"
25#include <subcmd/parse-options.h>
26#include "util/parse-events.h"
27#include "util/sort.h"
28#include "util/hist.h"
29#include "util/dso.h"
30#include "util/machine.h"
31#include "util/map.h"
32#include "util/session.h"
33#include "util/tool.h"
34#include "util/data.h"
35#include "arch/common.h"
36#include "util/block-range.h"
37#include "util/map_symbol.h"
38#include "util/branch.h"
39#include "util/util.h"
40
41#include <dlfcn.h>
42#include <errno.h>
43#include <linux/bitmap.h>
44#include <linux/err.h>
45
46struct perf_annotate {
47 struct perf_tool tool;
48 struct perf_session *session;
49#ifdef HAVE_SLANG_SUPPORT
50 bool use_tui;
51#endif
52 bool use_stdio, use_stdio2;
53#ifdef HAVE_GTK2_SUPPORT
54 bool use_gtk;
55#endif
56 bool skip_missing;
57 bool has_br_stack;
58 bool group_set;
59 bool data_type;
60 bool type_stat;
61 bool insn_stat;
62 float min_percent;
63 const char *sym_hist_filter;
64 const char *cpu_list;
65 const char *target_data_type;
66 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
67};
68
69/*
70 * Given one basic block:
71 *
72 * from to branch_i
73 * * ----> *
74 * |
75 * | block
76 * v
77 * * ----> *
78 * from to branch_i+1
79 *
80 * where the horizontal are the branches and the vertical is the executed
81 * block of instructions.
82 *
83 * We count, for each 'instruction', the number of blocks that covered it as
84 * well as count the ratio each branch is taken.
85 *
86 * We can do this without knowing the actual instruction stream by keeping
87 * track of the address ranges. We break down ranges such that there is no
88 * overlap and iterate from the start until the end.
89 *
90 * @acme: once we parse the objdump output _before_ processing the samples,
91 * we can easily fold the branch.cycles IPC bits in.
92 */
93static void process_basic_block(struct addr_map_symbol *start,
94 struct addr_map_symbol *end,
95 struct branch_flags *flags)
96{
97 struct symbol *sym = start->ms.sym;
98 struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
99 struct block_range_iter iter;
100 struct block_range *entry;
101 struct annotated_branch *branch;
102
103 /*
104 * Sanity; NULL isn't executable and the CPU cannot execute backwards
105 */
106 if (!start->addr || start->addr > end->addr)
107 return;
108
109 iter = block_range__create(start->addr, end->addr);
110 if (!block_range_iter__valid(&iter))
111 return;
112
113 branch = annotation__get_branch(notes);
114
115 /*
116 * First block in range is a branch target.
117 */
118 entry = block_range_iter(&iter);
119 assert(entry->is_target);
120 entry->entry++;
121
122 do {
123 entry = block_range_iter(&iter);
124
125 entry->coverage++;
126 entry->sym = sym;
127
128 if (branch)
129 branch->max_coverage = max(branch->max_coverage, entry->coverage);
130
131 } while (block_range_iter__next(&iter));
132
133 /*
134 * Last block in rage is a branch.
135 */
136 entry = block_range_iter(&iter);
137 assert(entry->is_branch);
138 entry->taken++;
139 if (flags->predicted)
140 entry->pred++;
141}
142
143static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
144 struct perf_sample *sample)
145{
146 struct addr_map_symbol *prev = NULL;
147 struct branch_info *bi;
148 int i;
149
150 if (!bs || !bs->nr)
151 return;
152
153 bi = sample__resolve_bstack(sample, al);
154 if (!bi)
155 return;
156
157 for (i = bs->nr - 1; i >= 0; i--) {
158 /*
159 * XXX filter against symbol
160 */
161 if (prev)
162 process_basic_block(prev, &bi[i].from, &bi[i].flags);
163 prev = &bi[i].to;
164 }
165
166 free(bi);
167}
168
169static int hist_iter__branch_callback(struct hist_entry_iter *iter,
170 struct addr_location *al __maybe_unused,
171 bool single __maybe_unused,
172 void *arg __maybe_unused)
173{
174 struct hist_entry *he = iter->he;
175 struct branch_info *bi;
176 struct perf_sample *sample = iter->sample;
177 struct evsel *evsel = iter->evsel;
178 int err;
179
180 bi = he->branch_info;
181 err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
182
183 if (err)
184 goto out;
185
186 err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
187
188out:
189 return err;
190}
191
192static int process_branch_callback(struct evsel *evsel,
193 struct perf_sample *sample,
194 struct addr_location *al,
195 struct perf_annotate *ann,
196 struct machine *machine)
197{
198 struct hist_entry_iter iter = {
199 .evsel = evsel,
200 .sample = sample,
201 .add_entry_cb = hist_iter__branch_callback,
202 .hide_unresolved = symbol_conf.hide_unresolved,
203 .ops = &hist_iter_branch,
204 };
205 struct addr_location a;
206 int ret;
207
208 addr_location__init(&a);
209 if (machine__resolve(machine, &a, sample) < 0) {
210 ret = -1;
211 goto out;
212 }
213
214 if (a.sym == NULL) {
215 ret = 0;
216 goto out;
217 }
218
219 if (a.map != NULL)
220 map__dso(a.map)->hit = 1;
221
222 hist__account_cycles(sample->branch_stack, al, sample, false, NULL);
223
224 ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
225out:
226 addr_location__exit(&a);
227 return ret;
228}
229
230static bool has_annotation(struct perf_annotate *ann)
231{
232 return ui__has_annotation() || ann->use_stdio2;
233}
234
235static int evsel__add_sample(struct evsel *evsel, struct perf_sample *sample,
236 struct addr_location *al, struct perf_annotate *ann,
237 struct machine *machine)
238{
239 struct hists *hists = evsel__hists(evsel);
240 struct hist_entry *he;
241 int ret;
242
243 if ((!ann->has_br_stack || !has_annotation(ann)) &&
244 ann->sym_hist_filter != NULL &&
245 (al->sym == NULL ||
246 strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
247 /* We're only interested in a symbol named sym_hist_filter */
248 /*
249 * FIXME: why isn't this done in the symbol_filter when loading
250 * the DSO?
251 */
252 if (al->sym != NULL) {
253 struct dso *dso = map__dso(al->map);
254
255 rb_erase_cached(&al->sym->rb_node, &dso->symbols);
256 symbol__delete(al->sym);
257 dso__reset_find_symbol_cache(dso);
258 }
259 return 0;
260 }
261
262 /*
263 * XXX filtered samples can still have branch entries pointing into our
264 * symbol and are missed.
265 */
266 process_branch_stack(sample->branch_stack, al, sample);
267
268 if (ann->has_br_stack && has_annotation(ann))
269 return process_branch_callback(evsel, sample, al, ann, machine);
270
271 he = hists__add_entry(hists, al, NULL, NULL, NULL, NULL, sample, true);
272 if (he == NULL)
273 return -ENOMEM;
274
275 ret = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
276 hists__inc_nr_samples(hists, true);
277 return ret;
278}
279
280static int process_sample_event(struct perf_tool *tool,
281 union perf_event *event,
282 struct perf_sample *sample,
283 struct evsel *evsel,
284 struct machine *machine)
285{
286 struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
287 struct addr_location al;
288 int ret = 0;
289
290 addr_location__init(&al);
291 if (machine__resolve(machine, &al, sample) < 0) {
292 pr_warning("problem processing %d event, skipping it.\n",
293 event->header.type);
294 ret = -1;
295 goto out_put;
296 }
297
298 if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
299 goto out_put;
300
301 if (!al.filtered &&
302 evsel__add_sample(evsel, sample, &al, ann, machine)) {
303 pr_warning("problem incrementing symbol count, "
304 "skipping event\n");
305 ret = -1;
306 }
307out_put:
308 addr_location__exit(&al);
309 return ret;
310}
311
312static int process_feature_event(struct perf_session *session,
313 union perf_event *event)
314{
315 if (event->feat.feat_id < HEADER_LAST_FEATURE)
316 return perf_event__process_feature(session, event);
317 return 0;
318}
319
320static int hist_entry__tty_annotate(struct hist_entry *he,
321 struct evsel *evsel,
322 struct perf_annotate *ann)
323{
324 if (!ann->use_stdio2)
325 return symbol__tty_annotate(&he->ms, evsel);
326
327 return symbol__tty_annotate2(&he->ms, evsel);
328}
329
330static void print_annotated_data_header(struct hist_entry *he, struct evsel *evsel)
331{
332 struct dso *dso = map__dso(he->ms.map);
333 int nr_members = 1;
334 int nr_samples = he->stat.nr_events;
335
336 if (evsel__is_group_event(evsel)) {
337 struct hist_entry *pair;
338
339 list_for_each_entry(pair, &he->pairs.head, pairs.node)
340 nr_samples += pair->stat.nr_events;
341 }
342
343 printf("Annotate type: '%s' in %s (%d samples):\n",
344 he->mem_type->self.type_name, dso->name, nr_samples);
345
346 if (evsel__is_group_event(evsel)) {
347 struct evsel *pos;
348 int i = 0;
349
350 for_each_group_evsel(pos, evsel)
351 printf(" event[%d] = %s\n", i++, pos->name);
352
353 nr_members = evsel->core.nr_members;
354 }
355
356 printf("============================================================================\n");
357 printf("%*s %10s %10s %s\n", 11 * nr_members, "samples", "offset", "size", "field");
358}
359
360static void print_annotated_data_type(struct annotated_data_type *mem_type,
361 struct annotated_member *member,
362 struct evsel *evsel, int indent)
363{
364 struct annotated_member *child;
365 struct type_hist *h = mem_type->histograms[evsel->core.idx];
366 int i, nr_events = 1, samples = 0;
367
368 for (i = 0; i < member->size; i++)
369 samples += h->addr[member->offset + i].nr_samples;
370 printf(" %10d", samples);
371
372 if (evsel__is_group_event(evsel)) {
373 struct evsel *pos;
374
375 for_each_group_member(pos, evsel) {
376 h = mem_type->histograms[pos->core.idx];
377
378 samples = 0;
379 for (i = 0; i < member->size; i++)
380 samples += h->addr[member->offset + i].nr_samples;
381 printf(" %10d", samples);
382 }
383 nr_events = evsel->core.nr_members;
384 }
385
386 printf(" %10d %10d %*s%s\t%s",
387 member->offset, member->size, indent, "", member->type_name,
388 member->var_name ?: "");
389
390 if (!list_empty(&member->children))
391 printf(" {\n");
392
393 list_for_each_entry(child, &member->children, node)
394 print_annotated_data_type(mem_type, child, evsel, indent + 4);
395
396 if (!list_empty(&member->children))
397 printf("%*s}", 11 * nr_events + 24 + indent, "");
398 printf(";\n");
399}
400
401static void print_annotate_data_stat(struct annotated_data_stat *s)
402{
403#define PRINT_STAT(fld) if (s->fld) printf("%10d : %s\n", s->fld, #fld)
404
405 int bad = s->no_sym +
406 s->no_insn +
407 s->no_insn_ops +
408 s->no_mem_ops +
409 s->no_reg +
410 s->no_dbginfo +
411 s->no_cuinfo +
412 s->no_var +
413 s->no_typeinfo +
414 s->invalid_size +
415 s->bad_offset;
416 int ok = s->total - bad;
417
418 printf("Annotate data type stats:\n");
419 printf("total %d, ok %d (%.1f%%), bad %d (%.1f%%)\n",
420 s->total, ok, 100.0 * ok / (s->total ?: 1), bad, 100.0 * bad / (s->total ?: 1));
421 printf("-----------------------------------------------------------\n");
422 PRINT_STAT(no_sym);
423 PRINT_STAT(no_insn);
424 PRINT_STAT(no_insn_ops);
425 PRINT_STAT(no_mem_ops);
426 PRINT_STAT(no_reg);
427 PRINT_STAT(no_dbginfo);
428 PRINT_STAT(no_cuinfo);
429 PRINT_STAT(no_var);
430 PRINT_STAT(no_typeinfo);
431 PRINT_STAT(invalid_size);
432 PRINT_STAT(bad_offset);
433 printf("\n");
434
435#undef PRINT_STAT
436}
437
438static void print_annotate_item_stat(struct list_head *head, const char *title)
439{
440 struct annotated_item_stat *istat, *pos, *iter;
441 int total_good, total_bad, total;
442 int sum1, sum2;
443 LIST_HEAD(tmp);
444
445 /* sort the list by count */
446 list_splice_init(head, &tmp);
447 total_good = total_bad = 0;
448
449 list_for_each_entry_safe(istat, pos, &tmp, list) {
450 total_good += istat->good;
451 total_bad += istat->bad;
452 sum1 = istat->good + istat->bad;
453
454 list_for_each_entry(iter, head, list) {
455 sum2 = iter->good + iter->bad;
456 if (sum1 > sum2)
457 break;
458 }
459 list_move_tail(&istat->list, &iter->list);
460 }
461 total = total_good + total_bad;
462
463 printf("Annotate %s stats\n", title);
464 printf("total %d, ok %d (%.1f%%), bad %d (%.1f%%)\n\n", total,
465 total_good, 100.0 * total_good / (total ?: 1),
466 total_bad, 100.0 * total_bad / (total ?: 1));
467 printf(" %-10s: %5s %5s\n", "Name", "Good", "Bad");
468 printf("-----------------------------------------------------------\n");
469 list_for_each_entry(istat, head, list)
470 printf(" %-10s: %5d %5d\n", istat->name, istat->good, istat->bad);
471 printf("\n");
472}
473
474static void hists__find_annotations(struct hists *hists,
475 struct evsel *evsel,
476 struct perf_annotate *ann)
477{
478 struct rb_node *nd = rb_first_cached(&hists->entries), *next;
479 int key = K_RIGHT;
480
481 if (ann->type_stat)
482 print_annotate_data_stat(&ann_data_stat);
483 if (ann->insn_stat)
484 print_annotate_item_stat(&ann_insn_stat, "Instruction");
485
486 while (nd) {
487 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
488 struct annotation *notes;
489
490 if (he->ms.sym == NULL || map__dso(he->ms.map)->annotate_warned)
491 goto find_next;
492
493 if (ann->sym_hist_filter &&
494 (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
495 goto find_next;
496
497 if (ann->min_percent) {
498 float percent = 0;
499 u64 total = hists__total_period(hists);
500
501 if (total)
502 percent = 100.0 * he->stat.period / total;
503
504 if (percent < ann->min_percent)
505 goto find_next;
506 }
507
508 notes = symbol__annotation(he->ms.sym);
509 if (notes->src == NULL) {
510find_next:
511 if (key == K_LEFT || key == '<')
512 nd = rb_prev(nd);
513 else
514 nd = rb_next(nd);
515 continue;
516 }
517
518 if (ann->data_type) {
519 /* skip unknown type */
520 if (he->mem_type->histograms == NULL)
521 goto find_next;
522
523 if (ann->target_data_type) {
524 const char *type_name = he->mem_type->self.type_name;
525
526 /* skip 'struct ' prefix in the type name */
527 if (strncmp(ann->target_data_type, "struct ", 7) &&
528 !strncmp(type_name, "struct ", 7))
529 type_name += 7;
530
531 /* skip 'union ' prefix in the type name */
532 if (strncmp(ann->target_data_type, "union ", 6) &&
533 !strncmp(type_name, "union ", 6))
534 type_name += 6;
535
536 if (strcmp(ann->target_data_type, type_name))
537 goto find_next;
538 }
539
540 print_annotated_data_header(he, evsel);
541 print_annotated_data_type(he->mem_type, &he->mem_type->self, evsel, 0);
542 printf("\n");
543 goto find_next;
544 }
545
546 if (use_browser == 2) {
547 int ret;
548 int (*annotate)(struct hist_entry *he,
549 struct evsel *evsel,
550 struct hist_browser_timer *hbt);
551
552 annotate = dlsym(perf_gtk_handle,
553 "hist_entry__gtk_annotate");
554 if (annotate == NULL) {
555 ui__error("GTK browser not found!\n");
556 return;
557 }
558
559 ret = annotate(he, evsel, NULL);
560 if (!ret || !ann->skip_missing)
561 return;
562
563 /* skip missing symbols */
564 nd = rb_next(nd);
565 } else if (use_browser == 1) {
566 key = hist_entry__tui_annotate(he, evsel, NULL);
567
568 switch (key) {
569 case -1:
570 if (!ann->skip_missing)
571 return;
572 /* fall through */
573 case K_RIGHT:
574 case '>':
575 next = rb_next(nd);
576 break;
577 case K_LEFT:
578 case '<':
579 next = rb_prev(nd);
580 break;
581 default:
582 return;
583 }
584
585 if (next != NULL)
586 nd = next;
587 } else {
588 hist_entry__tty_annotate(he, evsel, ann);
589 nd = rb_next(nd);
590 }
591 }
592}
593
594static int __cmd_annotate(struct perf_annotate *ann)
595{
596 int ret;
597 struct perf_session *session = ann->session;
598 struct evsel *pos;
599 u64 total_nr_samples;
600
601 if (ann->cpu_list) {
602 ret = perf_session__cpu_bitmap(session, ann->cpu_list,
603 ann->cpu_bitmap);
604 if (ret)
605 goto out;
606 }
607
608 if (!annotate_opts.objdump_path) {
609 ret = perf_env__lookup_objdump(&session->header.env,
610 &annotate_opts.objdump_path);
611 if (ret)
612 goto out;
613 }
614
615 ret = perf_session__process_events(session);
616 if (ret)
617 goto out;
618
619 if (dump_trace) {
620 perf_session__fprintf_nr_events(session, stdout, false);
621 evlist__fprintf_nr_events(session->evlist, stdout, false);
622 goto out;
623 }
624
625 if (verbose > 3)
626 perf_session__fprintf(session, stdout);
627
628 if (verbose > 2)
629 perf_session__fprintf_dsos(session, stdout);
630
631 total_nr_samples = 0;
632 evlist__for_each_entry(session->evlist, pos) {
633 struct hists *hists = evsel__hists(pos);
634 u32 nr_samples = hists->stats.nr_samples;
635
636 if (nr_samples > 0) {
637 total_nr_samples += nr_samples;
638 hists__collapse_resort(hists, NULL);
639 /* Don't sort callchain */
640 evsel__reset_sample_bit(pos, CALLCHAIN);
641 evsel__output_resort(pos, NULL);
642
643 /*
644 * An event group needs to display other events too.
645 * Let's delay printing until other events are processed.
646 */
647 if (symbol_conf.event_group) {
648 if (!evsel__is_group_leader(pos)) {
649 struct hists *leader_hists;
650
651 leader_hists = evsel__hists(evsel__leader(pos));
652 hists__match(leader_hists, hists);
653 hists__link(leader_hists, hists);
654 }
655 continue;
656 }
657
658 hists__find_annotations(hists, pos, ann);
659 }
660 }
661
662 if (total_nr_samples == 0) {
663 ui__error("The %s data has no samples!\n", session->data->path);
664 goto out;
665 }
666
667 /* Display group events together */
668 evlist__for_each_entry(session->evlist, pos) {
669 struct hists *hists = evsel__hists(pos);
670 u32 nr_samples = hists->stats.nr_samples;
671
672 if (nr_samples == 0)
673 continue;
674
675 if (!symbol_conf.event_group || !evsel__is_group_leader(pos))
676 continue;
677
678 hists__find_annotations(hists, pos, ann);
679 }
680
681 if (use_browser == 2) {
682 void (*show_annotations)(void);
683
684 show_annotations = dlsym(perf_gtk_handle,
685 "perf_gtk__show_annotations");
686 if (show_annotations == NULL) {
687 ui__error("GTK browser not found!\n");
688 goto out;
689 }
690 show_annotations();
691 }
692
693out:
694 return ret;
695}
696
697static int parse_percent_limit(const struct option *opt, const char *str,
698 int unset __maybe_unused)
699{
700 struct perf_annotate *ann = opt->value;
701 double pcnt = strtof(str, NULL);
702
703 ann->min_percent = pcnt;
704 return 0;
705}
706
707static int parse_data_type(const struct option *opt, const char *str, int unset)
708{
709 struct perf_annotate *ann = opt->value;
710
711 ann->data_type = !unset;
712 if (str)
713 ann->target_data_type = strdup(str);
714
715 return 0;
716}
717
718static const char * const annotate_usage[] = {
719 "perf annotate [<options>]",
720 NULL
721};
722
723int cmd_annotate(int argc, const char **argv)
724{
725 struct perf_annotate annotate = {
726 .tool = {
727 .sample = process_sample_event,
728 .mmap = perf_event__process_mmap,
729 .mmap2 = perf_event__process_mmap2,
730 .comm = perf_event__process_comm,
731 .exit = perf_event__process_exit,
732 .fork = perf_event__process_fork,
733 .namespaces = perf_event__process_namespaces,
734 .attr = perf_event__process_attr,
735 .build_id = perf_event__process_build_id,
736#ifdef HAVE_LIBTRACEEVENT
737 .tracing_data = perf_event__process_tracing_data,
738#endif
739 .id_index = perf_event__process_id_index,
740 .auxtrace_info = perf_event__process_auxtrace_info,
741 .auxtrace = perf_event__process_auxtrace,
742 .feature = process_feature_event,
743 .ordered_events = true,
744 .ordering_requires_timestamps = true,
745 },
746 };
747 struct perf_data data = {
748 .mode = PERF_DATA_MODE_READ,
749 };
750 struct itrace_synth_opts itrace_synth_opts = {
751 .set = 0,
752 };
753 const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
754 struct option options[] = {
755 OPT_STRING('i', "input", &input_name, "file",
756 "input file name"),
757 OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
758 "only consider symbols in these dsos"),
759 OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
760 "symbol to annotate"),
761 OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
762 OPT_INCR('v', "verbose", &verbose,
763 "be more verbose (show symbol address, etc)"),
764 OPT_BOOLEAN('q', "quiet", &quiet, "do now show any warnings or messages"),
765 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
766 "dump raw trace in ASCII"),
767#ifdef HAVE_GTK2_SUPPORT
768 OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
769#endif
770#ifdef HAVE_SLANG_SUPPORT
771 OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
772#endif
773 OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
774 OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
775 OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
776 "don't load vmlinux even if found"),
777 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
778 "file", "vmlinux pathname"),
779 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
780 "load module symbols - WARNING: use only with -k and LIVE kernel"),
781 OPT_BOOLEAN('l', "print-line", &annotate_opts.print_lines,
782 "print matching source lines (may be slow)"),
783 OPT_BOOLEAN('P', "full-paths", &annotate_opts.full_path,
784 "Don't shorten the displayed pathnames"),
785 OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
786 "Skip symbols that cannot be annotated"),
787 OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
788 &annotate.group_set,
789 "Show event group information together"),
790 OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
791 OPT_CALLBACK(0, "symfs", NULL, "directory",
792 "Look for files with symbols relative to this directory",
793 symbol__config_symfs),
794 OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
795 "Interleave source code with assembly code (default)"),
796 OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
797 "Display raw encoding of assembly instructions (default)"),
798 OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
799 "Specify disassembler style (e.g. -M intel for intel syntax)"),
800 OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
801 "Add prefix to source file path names in programs (with --prefix-strip)"),
802 OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
803 "Strip first N entries of source file path name in programs (with --prefix)"),
804 OPT_STRING(0, "objdump", &objdump_path, "path",
805 "objdump binary to use for disassembly and annotations"),
806 OPT_STRING(0, "addr2line", &addr2line_path, "path",
807 "addr2line binary to use for line numbers"),
808 OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
809 "Enable symbol demangling"),
810 OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
811 "Enable kernel symbol demangling"),
812 OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
813 "Show a column with the sum of periods"),
814 OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
815 "Show a column with the number of samples"),
816 OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
817 "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
818 stdio__config_color, "always"),
819 OPT_CALLBACK(0, "percent-type", &annotate_opts, "local-period",
820 "Set percent type local/global-period/hits",
821 annotate_parse_percent_type),
822 OPT_CALLBACK(0, "percent-limit", &annotate, "percent",
823 "Don't show entries under that percent", parse_percent_limit),
824 OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
825 "Instruction Tracing options\n" ITRACE_HELP,
826 itrace_parse_synth_opts),
827 OPT_CALLBACK_OPTARG(0, "data-type", &annotate, NULL, "name",
828 "Show data type annotate for the memory accesses",
829 parse_data_type),
830 OPT_BOOLEAN(0, "type-stat", &annotate.type_stat,
831 "Show stats for the data type annotation"),
832 OPT_BOOLEAN(0, "insn-stat", &annotate.insn_stat,
833 "Show instruction stats for the data type annotation"),
834 OPT_END()
835 };
836 int ret;
837
838 set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
839 set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
840
841 annotation_options__init();
842
843 ret = hists__init();
844 if (ret < 0)
845 return ret;
846
847 annotation_config__init();
848
849 argc = parse_options(argc, argv, options, annotate_usage, 0);
850 if (argc) {
851 /*
852 * Special case: if there's an argument left then assume that
853 * it's a symbol filter:
854 */
855 if (argc > 1)
856 usage_with_options(annotate_usage, options);
857
858 annotate.sym_hist_filter = argv[0];
859 }
860
861 if (disassembler_style) {
862 annotate_opts.disassembler_style = strdup(disassembler_style);
863 if (!annotate_opts.disassembler_style)
864 return -ENOMEM;
865 }
866 if (objdump_path) {
867 annotate_opts.objdump_path = strdup(objdump_path);
868 if (!annotate_opts.objdump_path)
869 return -ENOMEM;
870 }
871 if (addr2line_path) {
872 symbol_conf.addr2line_path = strdup(addr2line_path);
873 if (!symbol_conf.addr2line_path)
874 return -ENOMEM;
875 }
876
877 if (annotate_check_args() < 0)
878 return -EINVAL;
879
880#ifdef HAVE_GTK2_SUPPORT
881 if (symbol_conf.show_nr_samples && annotate.use_gtk) {
882 pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
883 return ret;
884 }
885#endif
886
887#ifndef HAVE_DWARF_GETLOCATIONS_SUPPORT
888 if (annotate.data_type) {
889 pr_err("Error: Data type profiling is disabled due to missing DWARF support\n");
890 return -ENOTSUP;
891 }
892#endif
893
894 ret = symbol__validate_sym_arguments();
895 if (ret)
896 return ret;
897
898 if (quiet)
899 perf_quiet_option();
900
901 data.path = input_name;
902
903 annotate.session = perf_session__new(&data, &annotate.tool);
904 if (IS_ERR(annotate.session))
905 return PTR_ERR(annotate.session);
906
907 annotate.session->itrace_synth_opts = &itrace_synth_opts;
908
909 annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
910 HEADER_BRANCH_STACK);
911
912 if (annotate.group_set)
913 evlist__force_leader(annotate.session->evlist);
914
915 ret = symbol__annotation_init();
916 if (ret < 0)
917 goto out_delete;
918
919 symbol_conf.try_vmlinux_path = true;
920
921 ret = symbol__init(&annotate.session->header.env);
922 if (ret < 0)
923 goto out_delete;
924
925 if (annotate.use_stdio || annotate.use_stdio2)
926 use_browser = 0;
927#ifdef HAVE_SLANG_SUPPORT
928 else if (annotate.use_tui)
929 use_browser = 1;
930#endif
931#ifdef HAVE_GTK2_SUPPORT
932 else if (annotate.use_gtk)
933 use_browser = 2;
934#endif
935
936 /* FIXME: only support stdio for now */
937 if (annotate.data_type) {
938 use_browser = 0;
939 annotate_opts.annotate_src = false;
940 symbol_conf.annotate_data_member = true;
941 symbol_conf.annotate_data_sample = true;
942 }
943
944 setup_browser(true);
945
946 /*
947 * Events of different processes may correspond to the same
948 * symbol, we do not care about the processes in annotate,
949 * set sort order to avoid repeated output.
950 */
951 if (annotate.data_type)
952 sort_order = "dso,type";
953 else
954 sort_order = "dso,symbol";
955
956 /*
957 * Set SORT_MODE__BRANCH so that annotate display IPC/Cycle
958 * if branch info is in perf data in TUI mode.
959 */
960 if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack)
961 sort__mode = SORT_MODE__BRANCH;
962
963 if (setup_sorting(NULL) < 0)
964 usage_with_options(annotate_usage, options);
965
966 ret = __cmd_annotate(&annotate);
967
968out_delete:
969 /*
970 * Speed up the exit process by only deleting for debug builds. For
971 * large files this can save time.
972 */
973#ifndef NDEBUG
974 perf_session__delete(annotate.session);
975#endif
976 annotation_options__exit();
977
978 return ret;
979}