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 "perf.h"
19 #include "util/debug.h"
20 
21 #include "util/evlist.h"
22 #include "util/evsel.h"
23 #include "util/annotate.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 
40 #include <dlfcn.h>
41 #include <errno.h>
42 #include <linux/bitmap.h>
43 #include <linux/err.h>
44 
45 struct perf_annotate {
46 	struct perf_tool tool;
47 	struct perf_session *session;
48 	struct annotation_options opts;
49 	bool	   use_tui, use_stdio, use_stdio2, use_gtk;
50 	bool	   skip_missing;
51 	bool	   has_br_stack;
52 	bool	   group_set;
53 	const char *sym_hist_filter;
54 	const char *cpu_list;
55 	DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
56 };
57 
58 /*
59  * Given one basic block:
60  *
61  *	from	to		branch_i
62  *	* ----> *
63  *		|
64  *		| block
65  *		v
66  *		* ----> *
67  *		from	to	branch_i+1
68  *
69  * where the horizontal are the branches and the vertical is the executed
70  * block of instructions.
71  *
72  * We count, for each 'instruction', the number of blocks that covered it as
73  * well as count the ratio each branch is taken.
74  *
75  * We can do this without knowing the actual instruction stream by keeping
76  * track of the address ranges. We break down ranges such that there is no
77  * overlap and iterate from the start until the end.
78  *
79  * @acme: once we parse the objdump output _before_ processing the samples,
80  * we can easily fold the branch.cycles IPC bits in.
81  */
process_basic_block(struct addr_map_symbol * start,struct addr_map_symbol * end,struct branch_flags * flags)82 static void process_basic_block(struct addr_map_symbol *start,
83 				struct addr_map_symbol *end,
84 				struct branch_flags *flags)
85 {
86 	struct symbol *sym = start->ms.sym;
87 	struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
88 	struct block_range_iter iter;
89 	struct block_range *entry;
90 
91 	/*
92 	 * Sanity; NULL isn't executable and the CPU cannot execute backwards
93 	 */
94 	if (!start->addr || start->addr > end->addr)
95 		return;
96 
97 	iter = block_range__create(start->addr, end->addr);
98 	if (!block_range_iter__valid(&iter))
99 		return;
100 
101 	/*
102 	 * First block in range is a branch target.
103 	 */
104 	entry = block_range_iter(&iter);
105 	assert(entry->is_target);
106 	entry->entry++;
107 
108 	do {
109 		entry = block_range_iter(&iter);
110 
111 		entry->coverage++;
112 		entry->sym = sym;
113 
114 		if (notes)
115 			notes->max_coverage = max(notes->max_coverage, entry->coverage);
116 
117 	} while (block_range_iter__next(&iter));
118 
119 	/*
120 	 * Last block in rage is a branch.
121 	 */
122 	entry = block_range_iter(&iter);
123 	assert(entry->is_branch);
124 	entry->taken++;
125 	if (flags->predicted)
126 		entry->pred++;
127 }
128 
process_branch_stack(struct branch_stack * bs,struct addr_location * al,struct perf_sample * sample)129 static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
130 				 struct perf_sample *sample)
131 {
132 	struct addr_map_symbol *prev = NULL;
133 	struct branch_info *bi;
134 	int i;
135 
136 	if (!bs || !bs->nr)
137 		return;
138 
139 	bi = sample__resolve_bstack(sample, al);
140 	if (!bi)
141 		return;
142 
143 	for (i = bs->nr - 1; i >= 0; i--) {
144 		/*
145 		 * XXX filter against symbol
146 		 */
147 		if (prev)
148 			process_basic_block(prev, &bi[i].from, &bi[i].flags);
149 		prev = &bi[i].to;
150 	}
151 
152 	free(bi);
153 }
154 
hist_iter__branch_callback(struct hist_entry_iter * iter,struct addr_location * al __maybe_unused,bool single __maybe_unused,void * arg __maybe_unused)155 static int hist_iter__branch_callback(struct hist_entry_iter *iter,
156 				      struct addr_location *al __maybe_unused,
157 				      bool single __maybe_unused,
158 				      void *arg __maybe_unused)
159 {
160 	struct hist_entry *he = iter->he;
161 	struct branch_info *bi;
162 	struct perf_sample *sample = iter->sample;
163 	struct evsel *evsel = iter->evsel;
164 	int err;
165 
166 	bi = he->branch_info;
167 	err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
168 
169 	if (err)
170 		goto out;
171 
172 	err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
173 
174 out:
175 	return err;
176 }
177 
process_branch_callback(struct evsel * evsel,struct perf_sample * sample,struct addr_location * al __maybe_unused,struct perf_annotate * ann,struct machine * machine)178 static int process_branch_callback(struct evsel *evsel,
179 				   struct perf_sample *sample,
180 				   struct addr_location *al __maybe_unused,
181 				   struct perf_annotate *ann,
182 				   struct machine *machine)
183 {
184 	struct hist_entry_iter iter = {
185 		.evsel		= evsel,
186 		.sample		= sample,
187 		.add_entry_cb	= hist_iter__branch_callback,
188 		.hide_unresolved	= symbol_conf.hide_unresolved,
189 		.ops		= &hist_iter_branch,
190 	};
191 
192 	struct addr_location a;
193 	int ret;
194 
195 	if (machine__resolve(machine, &a, sample) < 0)
196 		return -1;
197 
198 	if (a.sym == NULL)
199 		return 0;
200 
201 	if (a.map != NULL)
202 		a.map->dso->hit = 1;
203 
204 	hist__account_cycles(sample->branch_stack, al, sample, false, NULL);
205 
206 	ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
207 	return ret;
208 }
209 
has_annotation(struct perf_annotate * ann)210 static bool has_annotation(struct perf_annotate *ann)
211 {
212 	return ui__has_annotation() || ann->use_stdio2;
213 }
214 
evsel__add_sample(struct evsel * evsel,struct perf_sample * sample,struct addr_location * al,struct perf_annotate * ann,struct machine * machine)215 static int evsel__add_sample(struct evsel *evsel, struct perf_sample *sample,
216 			     struct addr_location *al, struct perf_annotate *ann,
217 			     struct machine *machine)
218 {
219 	struct hists *hists = evsel__hists(evsel);
220 	struct hist_entry *he;
221 	int ret;
222 
223 	if ((!ann->has_br_stack || !has_annotation(ann)) &&
224 	    ann->sym_hist_filter != NULL &&
225 	    (al->sym == NULL ||
226 	     strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
227 		/* We're only interested in a symbol named sym_hist_filter */
228 		/*
229 		 * FIXME: why isn't this done in the symbol_filter when loading
230 		 * the DSO?
231 		 */
232 		if (al->sym != NULL) {
233 			rb_erase_cached(&al->sym->rb_node,
234 				 &al->map->dso->symbols);
235 			symbol__delete(al->sym);
236 			dso__reset_find_symbol_cache(al->map->dso);
237 		}
238 		return 0;
239 	}
240 
241 	/*
242 	 * XXX filtered samples can still have branch entries pointing into our
243 	 * symbol and are missed.
244 	 */
245 	process_branch_stack(sample->branch_stack, al, sample);
246 
247 	if (ann->has_br_stack && has_annotation(ann))
248 		return process_branch_callback(evsel, sample, al, ann, machine);
249 
250 	he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
251 	if (he == NULL)
252 		return -ENOMEM;
253 
254 	ret = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
255 	hists__inc_nr_samples(hists, true);
256 	return ret;
257 }
258 
process_sample_event(struct perf_tool * tool,union perf_event * event,struct perf_sample * sample,struct evsel * evsel,struct machine * machine)259 static int process_sample_event(struct perf_tool *tool,
260 				union perf_event *event,
261 				struct perf_sample *sample,
262 				struct evsel *evsel,
263 				struct machine *machine)
264 {
265 	struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
266 	struct addr_location al;
267 	int ret = 0;
268 
269 	if (machine__resolve(machine, &al, sample) < 0) {
270 		pr_warning("problem processing %d event, skipping it.\n",
271 			   event->header.type);
272 		return -1;
273 	}
274 
275 	if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
276 		goto out_put;
277 
278 	if (!al.filtered &&
279 	    evsel__add_sample(evsel, sample, &al, ann, machine)) {
280 		pr_warning("problem incrementing symbol count, "
281 			   "skipping event\n");
282 		ret = -1;
283 	}
284 out_put:
285 	addr_location__put(&al);
286 	return ret;
287 }
288 
process_feature_event(struct perf_session * session,union perf_event * event)289 static int process_feature_event(struct perf_session *session,
290 				 union perf_event *event)
291 {
292 	if (event->feat.feat_id < HEADER_LAST_FEATURE)
293 		return perf_event__process_feature(session, event);
294 	return 0;
295 }
296 
hist_entry__tty_annotate(struct hist_entry * he,struct evsel * evsel,struct perf_annotate * ann)297 static int hist_entry__tty_annotate(struct hist_entry *he,
298 				    struct evsel *evsel,
299 				    struct perf_annotate *ann)
300 {
301 	if (!ann->use_stdio2)
302 		return symbol__tty_annotate(&he->ms, evsel, &ann->opts);
303 
304 	return symbol__tty_annotate2(&he->ms, evsel, &ann->opts);
305 }
306 
hists__find_annotations(struct hists * hists,struct evsel * evsel,struct perf_annotate * ann)307 static void hists__find_annotations(struct hists *hists,
308 				    struct evsel *evsel,
309 				    struct perf_annotate *ann)
310 {
311 	struct rb_node *nd = rb_first_cached(&hists->entries), *next;
312 	int key = K_RIGHT;
313 
314 	while (nd) {
315 		struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
316 		struct annotation *notes;
317 
318 		if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
319 			goto find_next;
320 
321 		if (ann->sym_hist_filter &&
322 		    (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
323 			goto find_next;
324 
325 		notes = symbol__annotation(he->ms.sym);
326 		if (notes->src == NULL) {
327 find_next:
328 			if (key == K_LEFT)
329 				nd = rb_prev(nd);
330 			else
331 				nd = rb_next(nd);
332 			continue;
333 		}
334 
335 		if (use_browser == 2) {
336 			int ret;
337 			int (*annotate)(struct hist_entry *he,
338 					struct evsel *evsel,
339 					struct hist_browser_timer *hbt);
340 
341 			annotate = dlsym(perf_gtk_handle,
342 					 "hist_entry__gtk_annotate");
343 			if (annotate == NULL) {
344 				ui__error("GTK browser not found!\n");
345 				return;
346 			}
347 
348 			ret = annotate(he, evsel, NULL);
349 			if (!ret || !ann->skip_missing)
350 				return;
351 
352 			/* skip missing symbols */
353 			nd = rb_next(nd);
354 		} else if (use_browser == 1) {
355 			key = hist_entry__tui_annotate(he, evsel, NULL, &ann->opts);
356 
357 			switch (key) {
358 			case -1:
359 				if (!ann->skip_missing)
360 					return;
361 				/* fall through */
362 			case K_RIGHT:
363 				next = rb_next(nd);
364 				break;
365 			case K_LEFT:
366 				next = rb_prev(nd);
367 				break;
368 			default:
369 				return;
370 			}
371 
372 			if (next != NULL)
373 				nd = next;
374 		} else {
375 			hist_entry__tty_annotate(he, evsel, ann);
376 			nd = rb_next(nd);
377 		}
378 	}
379 }
380 
__cmd_annotate(struct perf_annotate * ann)381 static int __cmd_annotate(struct perf_annotate *ann)
382 {
383 	int ret;
384 	struct perf_session *session = ann->session;
385 	struct evsel *pos;
386 	u64 total_nr_samples;
387 
388 	if (ann->cpu_list) {
389 		ret = perf_session__cpu_bitmap(session, ann->cpu_list,
390 					       ann->cpu_bitmap);
391 		if (ret)
392 			goto out;
393 	}
394 
395 	if (!ann->opts.objdump_path) {
396 		ret = perf_env__lookup_objdump(&session->header.env,
397 					       &ann->opts.objdump_path);
398 		if (ret)
399 			goto out;
400 	}
401 
402 	ret = perf_session__process_events(session);
403 	if (ret)
404 		goto out;
405 
406 	if (dump_trace) {
407 		perf_session__fprintf_nr_events(session, stdout, false);
408 		evlist__fprintf_nr_events(session->evlist, stdout, false);
409 		goto out;
410 	}
411 
412 	if (verbose > 3)
413 		perf_session__fprintf(session, stdout);
414 
415 	if (verbose > 2)
416 		perf_session__fprintf_dsos(session, stdout);
417 
418 	total_nr_samples = 0;
419 	evlist__for_each_entry(session->evlist, pos) {
420 		struct hists *hists = evsel__hists(pos);
421 		u32 nr_samples = hists->stats.nr_samples;
422 
423 		if (nr_samples > 0) {
424 			total_nr_samples += nr_samples;
425 			hists__collapse_resort(hists, NULL);
426 			/* Don't sort callchain */
427 			evsel__reset_sample_bit(pos, CALLCHAIN);
428 			evsel__output_resort(pos, NULL);
429 
430 			if (symbol_conf.event_group && !evsel__is_group_leader(pos))
431 				continue;
432 
433 			hists__find_annotations(hists, pos, ann);
434 		}
435 	}
436 
437 	if (total_nr_samples == 0) {
438 		ui__error("The %s data has no samples!\n", session->data->path);
439 		goto out;
440 	}
441 
442 	if (use_browser == 2) {
443 		void (*show_annotations)(void);
444 
445 		show_annotations = dlsym(perf_gtk_handle,
446 					 "perf_gtk__show_annotations");
447 		if (show_annotations == NULL) {
448 			ui__error("GTK browser not found!\n");
449 			goto out;
450 		}
451 		show_annotations();
452 	}
453 
454 out:
455 	return ret;
456 }
457 
458 static const char * const annotate_usage[] = {
459 	"perf annotate [<options>]",
460 	NULL
461 };
462 
cmd_annotate(int argc,const char ** argv)463 int cmd_annotate(int argc, const char **argv)
464 {
465 	struct perf_annotate annotate = {
466 		.tool = {
467 			.sample	= process_sample_event,
468 			.mmap	= perf_event__process_mmap,
469 			.mmap2	= perf_event__process_mmap2,
470 			.comm	= perf_event__process_comm,
471 			.exit	= perf_event__process_exit,
472 			.fork	= perf_event__process_fork,
473 			.namespaces = perf_event__process_namespaces,
474 			.attr	= perf_event__process_attr,
475 			.build_id = perf_event__process_build_id,
476 			.tracing_data   = perf_event__process_tracing_data,
477 			.feature	= process_feature_event,
478 			.ordered_events = true,
479 			.ordering_requires_timestamps = true,
480 		},
481 		.opts = annotation__default_options,
482 	};
483 	struct perf_data data = {
484 		.mode  = PERF_DATA_MODE_READ,
485 	};
486 	struct option options[] = {
487 	OPT_STRING('i', "input", &input_name, "file",
488 		    "input file name"),
489 	OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
490 		   "only consider symbols in these dsos"),
491 	OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
492 		    "symbol to annotate"),
493 	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
494 	OPT_INCR('v', "verbose", &verbose,
495 		    "be more verbose (show symbol address, etc)"),
496 	OPT_BOOLEAN('q', "quiet", &quiet, "do now show any message"),
497 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
498 		    "dump raw trace in ASCII"),
499 	OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
500 	OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
501 	OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
502 	OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
503 	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
504                     "don't load vmlinux even if found"),
505 	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
506 		   "file", "vmlinux pathname"),
507 	OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
508 		    "load module symbols - WARNING: use only with -k and LIVE kernel"),
509 	OPT_BOOLEAN('l', "print-line", &annotate.opts.print_lines,
510 		    "print matching source lines (may be slow)"),
511 	OPT_BOOLEAN('P', "full-paths", &annotate.opts.full_path,
512 		    "Don't shorten the displayed pathnames"),
513 	OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
514 		    "Skip symbols that cannot be annotated"),
515 	OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
516 			&annotate.group_set,
517 			"Show event group information together"),
518 	OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
519 	OPT_CALLBACK(0, "symfs", NULL, "directory",
520 		     "Look for files with symbols relative to this directory",
521 		     symbol__config_symfs),
522 	OPT_BOOLEAN(0, "source", &annotate.opts.annotate_src,
523 		    "Interleave source code with assembly code (default)"),
524 	OPT_BOOLEAN(0, "asm-raw", &annotate.opts.show_asm_raw,
525 		    "Display raw encoding of assembly instructions (default)"),
526 	OPT_STRING('M', "disassembler-style", &annotate.opts.disassembler_style, "disassembler style",
527 		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
528 	OPT_STRING(0, "prefix", &annotate.opts.prefix, "prefix",
529 		    "Add prefix to source file path names in programs (with --prefix-strip)"),
530 	OPT_STRING(0, "prefix-strip", &annotate.opts.prefix_strip, "N",
531 		    "Strip first N entries of source file path name in programs (with --prefix)"),
532 	OPT_STRING(0, "objdump", &annotate.opts.objdump_path, "path",
533 		   "objdump binary to use for disassembly and annotations"),
534 	OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
535 		    "Enable symbol demangling"),
536 	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
537 		    "Enable kernel symbol demangling"),
538 	OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
539 		    "Show event group information together"),
540 	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
541 		    "Show a column with the sum of periods"),
542 	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
543 		    "Show a column with the number of samples"),
544 	OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
545 			     "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
546 			     stdio__config_color, "always"),
547 	OPT_CALLBACK(0, "percent-type", &annotate.opts, "local-period",
548 		     "Set percent type local/global-period/hits",
549 		     annotate_parse_percent_type),
550 
551 	OPT_END()
552 	};
553 	int ret;
554 
555 	set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
556 	set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
557 
558 
559 	ret = hists__init();
560 	if (ret < 0)
561 		return ret;
562 
563 	annotation_config__init(&annotate.opts);
564 
565 	argc = parse_options(argc, argv, options, annotate_usage, 0);
566 	if (argc) {
567 		/*
568 		 * Special case: if there's an argument left then assume that
569 		 * it's a symbol filter:
570 		 */
571 		if (argc > 1)
572 			usage_with_options(annotate_usage, options);
573 
574 		annotate.sym_hist_filter = argv[0];
575 	}
576 
577 	if (annotate_check_args(&annotate.opts) < 0)
578 		return -EINVAL;
579 
580 	if (symbol_conf.show_nr_samples && annotate.use_gtk) {
581 		pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
582 		return ret;
583 	}
584 
585 	if (quiet)
586 		perf_quiet_option();
587 
588 	data.path = input_name;
589 
590 	annotate.session = perf_session__new(&data, false, &annotate.tool);
591 	if (IS_ERR(annotate.session))
592 		return PTR_ERR(annotate.session);
593 
594 	annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
595 						      HEADER_BRANCH_STACK);
596 
597 	if (annotate.group_set)
598 		evlist__force_leader(annotate.session->evlist);
599 
600 	ret = symbol__annotation_init();
601 	if (ret < 0)
602 		goto out_delete;
603 
604 	symbol_conf.try_vmlinux_path = true;
605 
606 	ret = symbol__init(&annotate.session->header.env);
607 	if (ret < 0)
608 		goto out_delete;
609 
610 	if (annotate.use_stdio || annotate.use_stdio2)
611 		use_browser = 0;
612 	else if (annotate.use_tui)
613 		use_browser = 1;
614 	else if (annotate.use_gtk)
615 		use_browser = 2;
616 
617 	setup_browser(true);
618 
619 	/*
620 	 * Events of different processes may correspond to the same
621 	 * symbol, we do not care about the processes in annotate,
622 	 * set sort order to avoid repeated output.
623 	 */
624 	sort_order = "dso,symbol";
625 
626 	/*
627 	 * Set SORT_MODE__BRANCH so that annotate display IPC/Cycle
628 	 * if branch info is in perf data in TUI mode.
629 	 */
630 	if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack)
631 		sort__mode = SORT_MODE__BRANCH;
632 
633 	if (setup_sorting(NULL) < 0)
634 		usage_with_options(annotate_usage, options);
635 
636 	ret = __cmd_annotate(&annotate);
637 
638 out_delete:
639 	/*
640 	 * Speed up the exit process, for large files this can
641 	 * take quite a while.
642 	 *
643 	 * XXX Enable this when using valgrind or if we ever
644 	 * librarize this command.
645 	 *
646 	 * Also experiment with obstacks to see how much speed
647 	 * up we'll get here.
648 	 *
649 	 * perf_session__delete(session);
650 	 */
651 	return ret;
652 }
653