Linux Audio

Check our new training course

Loading...
v5.4
  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
 45struct 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 */
 82static 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->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
129static 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
155static 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
174out:
175	return err;
176}
177
178static 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);
205
206	ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
207	return ret;
208}
209
210static bool has_annotation(struct perf_annotate *ann)
211{
212	return ui__has_annotation() || ann->use_stdio2;
213}
214
215static int perf_evsel__add_sample(struct evsel *evsel,
216				  struct perf_sample *sample,
217				  struct addr_location *al,
218				  struct perf_annotate *ann,
219				  struct machine *machine)
220{
221	struct hists *hists = evsel__hists(evsel);
222	struct hist_entry *he;
223	int ret;
224
225	if ((!ann->has_br_stack || !has_annotation(ann)) &&
226	    ann->sym_hist_filter != NULL &&
227	    (al->sym == NULL ||
228	     strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
229		/* We're only interested in a symbol named sym_hist_filter */
230		/*
231		 * FIXME: why isn't this done in the symbol_filter when loading
232		 * the DSO?
233		 */
234		if (al->sym != NULL) {
235			rb_erase_cached(&al->sym->rb_node,
236				 &al->map->dso->symbols);
237			symbol__delete(al->sym);
238			dso__reset_find_symbol_cache(al->map->dso);
239		}
240		return 0;
241	}
242
243	/*
244	 * XXX filtered samples can still have branch entires pointing into our
245	 * symbol and are missed.
246	 */
247	process_branch_stack(sample->branch_stack, al, sample);
248
249	if (ann->has_br_stack && has_annotation(ann))
250		return process_branch_callback(evsel, sample, al, ann, machine);
251
252	he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
253	if (he == NULL)
254		return -ENOMEM;
255
256	ret = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
257	hists__inc_nr_samples(hists, true);
258	return ret;
259}
260
261static int process_sample_event(struct perf_tool *tool,
262				union perf_event *event,
263				struct perf_sample *sample,
264				struct evsel *evsel,
265				struct machine *machine)
266{
267	struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
268	struct addr_location al;
269	int ret = 0;
270
271	if (machine__resolve(machine, &al, sample) < 0) {
272		pr_warning("problem processing %d event, skipping it.\n",
273			   event->header.type);
274		return -1;
275	}
276
277	if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
278		goto out_put;
279
280	if (!al.filtered &&
281	    perf_evsel__add_sample(evsel, sample, &al, ann, machine)) {
282		pr_warning("problem incrementing symbol count, "
283			   "skipping event\n");
284		ret = -1;
285	}
286out_put:
287	addr_location__put(&al);
288	return ret;
289}
290
291static int process_feature_event(struct perf_session *session,
292				 union perf_event *event)
293{
294	if (event->feat.feat_id < HEADER_LAST_FEATURE)
295		return perf_event__process_feature(session, event);
296	return 0;
297}
298
299static int hist_entry__tty_annotate(struct hist_entry *he,
300				    struct evsel *evsel,
301				    struct perf_annotate *ann)
302{
303	if (!ann->use_stdio2)
304		return symbol__tty_annotate(he->ms.sym, he->ms.map, evsel, &ann->opts);
305
306	return symbol__tty_annotate2(he->ms.sym, he->ms.map, evsel, &ann->opts);
307}
308
309static void hists__find_annotations(struct hists *hists,
310				    struct evsel *evsel,
311				    struct perf_annotate *ann)
312{
313	struct rb_node *nd = rb_first_cached(&hists->entries), *next;
314	int key = K_RIGHT;
315
316	while (nd) {
317		struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
318		struct annotation *notes;
319
320		if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
321			goto find_next;
322
323		if (ann->sym_hist_filter &&
324		    (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
325			goto find_next;
326
 
 
 
 
 
 
 
 
 
 
 
327		notes = symbol__annotation(he->ms.sym);
328		if (notes->src == NULL) {
329find_next:
330			if (key == K_LEFT)
331				nd = rb_prev(nd);
332			else
333				nd = rb_next(nd);
334			continue;
335		}
336
337		if (use_browser == 2) {
338			int ret;
339			int (*annotate)(struct hist_entry *he,
340					struct evsel *evsel,
341					struct hist_browser_timer *hbt);
342
343			annotate = dlsym(perf_gtk_handle,
344					 "hist_entry__gtk_annotate");
345			if (annotate == NULL) {
346				ui__error("GTK browser not found!\n");
347				return;
348			}
349
350			ret = annotate(he, evsel, NULL);
351			if (!ret || !ann->skip_missing)
352				return;
353
354			/* skip missing symbols */
355			nd = rb_next(nd);
356		} else if (use_browser == 1) {
357			key = hist_entry__tui_annotate(he, evsel, NULL, &ann->opts);
358
359			switch (key) {
360			case -1:
361				if (!ann->skip_missing)
362					return;
363				/* fall through */
364			case K_RIGHT:
365				next = rb_next(nd);
366				break;
367			case K_LEFT:
368				next = rb_prev(nd);
369				break;
370			default:
371				return;
372			}
373
374			if (next != NULL)
375				nd = next;
376		} else {
377			hist_entry__tty_annotate(he, evsel, ann);
378			nd = rb_next(nd);
379			/*
380			 * Since we have a hist_entry per IP for the same
381			 * symbol, free he->ms.sym->src to signal we already
382			 * processed this symbol.
383			 */
384			zfree(&notes->src->cycles_hist);
385			zfree(&notes->src);
386		}
387	}
388}
389
390static int __cmd_annotate(struct perf_annotate *ann)
391{
392	int ret;
393	struct perf_session *session = ann->session;
394	struct evsel *pos;
395	u64 total_nr_samples;
396
397	if (ann->cpu_list) {
398		ret = perf_session__cpu_bitmap(session, ann->cpu_list,
399					       ann->cpu_bitmap);
400		if (ret)
401			goto out;
402	}
403
404	if (!ann->opts.objdump_path) {
405		ret = perf_env__lookup_objdump(&session->header.env,
406					       &ann->opts.objdump_path);
407		if (ret)
408			goto out;
409	}
410
411	ret = perf_session__process_events(session);
412	if (ret)
413		goto out;
414
415	if (dump_trace) {
416		perf_session__fprintf_nr_events(session, stdout);
417		perf_evlist__fprintf_nr_events(session->evlist, stdout);
418		goto out;
419	}
420
421	if (verbose > 3)
422		perf_session__fprintf(session, stdout);
423
424	if (verbose > 2)
425		perf_session__fprintf_dsos(session, stdout);
426
427	total_nr_samples = 0;
428	evlist__for_each_entry(session->evlist, pos) {
429		struct hists *hists = evsel__hists(pos);
430		u32 nr_samples = hists->stats.nr_events[PERF_RECORD_SAMPLE];
431
432		if (nr_samples > 0) {
433			total_nr_samples += nr_samples;
434			hists__collapse_resort(hists, NULL);
435			/* Don't sort callchain */
436			perf_evsel__reset_sample_bit(pos, CALLCHAIN);
437			perf_evsel__output_resort(pos, NULL);
438
439			if (symbol_conf.event_group &&
440			    !perf_evsel__is_group_leader(pos))
441				continue;
442
443			hists__find_annotations(hists, pos, ann);
444		}
445	}
446
447	if (total_nr_samples == 0) {
448		ui__error("The %s data has no samples!\n", session->data->path);
449		goto out;
450	}
451
452	if (use_browser == 2) {
453		void (*show_annotations)(void);
454
455		show_annotations = dlsym(perf_gtk_handle,
456					 "perf_gtk__show_annotations");
457		if (show_annotations == NULL) {
458			ui__error("GTK browser not found!\n");
459			goto out;
460		}
461		show_annotations();
462	}
463
464out:
465	return ret;
466}
467
 
 
 
 
 
 
 
 
 
 
468static const char * const annotate_usage[] = {
469	"perf annotate [<options>]",
470	NULL
471};
472
473int cmd_annotate(int argc, const char **argv)
474{
475	struct perf_annotate annotate = {
476		.tool = {
477			.sample	= process_sample_event,
478			.mmap	= perf_event__process_mmap,
479			.mmap2	= perf_event__process_mmap2,
480			.comm	= perf_event__process_comm,
481			.exit	= perf_event__process_exit,
482			.fork	= perf_event__process_fork,
483			.namespaces = perf_event__process_namespaces,
484			.attr	= perf_event__process_attr,
485			.build_id = perf_event__process_build_id,
 
486			.tracing_data   = perf_event__process_tracing_data,
 
 
 
 
487			.feature	= process_feature_event,
488			.ordered_events = true,
489			.ordering_requires_timestamps = true,
490		},
491		.opts = annotation__default_options,
492	};
493	struct perf_data data = {
494		.mode  = PERF_DATA_MODE_READ,
495	};
 
 
 
496	struct option options[] = {
497	OPT_STRING('i', "input", &input_name, "file",
498		    "input file name"),
499	OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
500		   "only consider symbols in these dsos"),
501	OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
502		    "symbol to annotate"),
503	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
504	OPT_INCR('v', "verbose", &verbose,
505		    "be more verbose (show symbol address, etc)"),
506	OPT_BOOLEAN('q', "quiet", &quiet, "do now show any message"),
507	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
508		    "dump raw trace in ASCII"),
 
509	OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
 
 
510	OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
 
511	OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
512	OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
513	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
514                    "don't load vmlinux even if found"),
515	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
516		   "file", "vmlinux pathname"),
517	OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
518		    "load module symbols - WARNING: use only with -k and LIVE kernel"),
519	OPT_BOOLEAN('l', "print-line", &annotate.opts.print_lines,
520		    "print matching source lines (may be slow)"),
521	OPT_BOOLEAN('P', "full-paths", &annotate.opts.full_path,
522		    "Don't shorten the displayed pathnames"),
523	OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
524		    "Skip symbols that cannot be annotated"),
525	OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
526			&annotate.group_set,
527			"Show event group information together"),
528	OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
529	OPT_CALLBACK(0, "symfs", NULL, "directory",
530		     "Look for files with symbols relative to this directory",
531		     symbol__config_symfs),
532	OPT_BOOLEAN(0, "source", &annotate.opts.annotate_src,
533		    "Interleave source code with assembly code (default)"),
534	OPT_BOOLEAN(0, "asm-raw", &annotate.opts.show_asm_raw,
535		    "Display raw encoding of assembly instructions (default)"),
536	OPT_STRING('M', "disassembler-style", &annotate.opts.disassembler_style, "disassembler style",
537		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
 
 
 
 
538	OPT_STRING(0, "objdump", &annotate.opts.objdump_path, "path",
539		   "objdump binary to use for disassembly and annotations"),
 
 
 
 
540	OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
541		    "Show event group information together"),
542	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
543		    "Show a column with the sum of periods"),
544	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
545		    "Show a column with the number of samples"),
546	OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
547			     "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
548			     stdio__config_color, "always"),
549	OPT_CALLBACK(0, "percent-type", &annotate.opts, "local-period",
550		     "Set percent type local/global-period/hits",
551		     annotate_parse_percent_type),
 
 
 
 
 
552
553	OPT_END()
554	};
555	int ret;
556
557	set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
558	set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
559
560
561	ret = hists__init();
562	if (ret < 0)
563		return ret;
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 (symbol_conf.show_nr_samples && annotate.use_gtk) {
578		pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
579		return ret;
580	}
 
 
 
 
 
581
582	if (quiet)
583		perf_quiet_option();
584
585	data.path = input_name;
586
587	annotate.session = perf_session__new(&data, false, &annotate.tool);
588	if (IS_ERR(annotate.session))
589		return PTR_ERR(annotate.session);
590
 
 
591	annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
592						      HEADER_BRANCH_STACK);
593
594	if (annotate.group_set)
595		perf_evlist__force_leader(annotate.session->evlist);
596
597	ret = symbol__annotation_init();
598	if (ret < 0)
599		goto out_delete;
600
601	annotation_config__init();
602
603	symbol_conf.try_vmlinux_path = true;
604
605	ret = symbol__init(&annotate.session->header.env);
606	if (ret < 0)
607		goto out_delete;
608
609	if (annotate.use_stdio || annotate.use_stdio2)
610		use_browser = 0;
 
611	else if (annotate.use_tui)
612		use_browser = 1;
 
 
613	else if (annotate.use_gtk)
614		use_browser = 2;
 
615
616	setup_browser(true);
617
618	if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack) {
 
 
 
 
 
 
 
 
 
 
 
619		sort__mode = SORT_MODE__BRANCH;
620		if (setup_sorting(annotate.session->evlist) < 0)
621			usage_with_options(annotate_usage, options);
622	} else {
623		if (setup_sorting(NULL) < 0)
624			usage_with_options(annotate_usage, options);
625	}
626
627	ret = __cmd_annotate(&annotate);
628
629out_delete:
630	/*
631	 * Speed up the exit process, for large files this can
632	 * take quite a while.
633	 *
634	 * XXX Enable this when using valgrind or if we ever
635	 * librarize this command.
636	 *
637	 * Also experiment with obstacks to see how much speed
638	 * up we'll get here.
639	 *
640	 * perf_session__delete(session);
641	 */
642	return ret;
643}
v6.2
  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
 45struct perf_annotate {
 46	struct perf_tool tool;
 47	struct perf_session *session;
 48	struct annotation_options opts;
 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	float	   min_percent;
 60	const char *sym_hist_filter;
 61	const char *cpu_list;
 62	DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
 63};
 64
 65/*
 66 * Given one basic block:
 67 *
 68 *	from	to		branch_i
 69 *	* ----> *
 70 *		|
 71 *		| block
 72 *		v
 73 *		* ----> *
 74 *		from	to	branch_i+1
 75 *
 76 * where the horizontal are the branches and the vertical is the executed
 77 * block of instructions.
 78 *
 79 * We count, for each 'instruction', the number of blocks that covered it as
 80 * well as count the ratio each branch is taken.
 81 *
 82 * We can do this without knowing the actual instruction stream by keeping
 83 * track of the address ranges. We break down ranges such that there is no
 84 * overlap and iterate from the start until the end.
 85 *
 86 * @acme: once we parse the objdump output _before_ processing the samples,
 87 * we can easily fold the branch.cycles IPC bits in.
 88 */
 89static void process_basic_block(struct addr_map_symbol *start,
 90				struct addr_map_symbol *end,
 91				struct branch_flags *flags)
 92{
 93	struct symbol *sym = start->ms.sym;
 94	struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
 95	struct block_range_iter iter;
 96	struct block_range *entry;
 97
 98	/*
 99	 * Sanity; NULL isn't executable and the CPU cannot execute backwards
100	 */
101	if (!start->addr || start->addr > end->addr)
102		return;
103
104	iter = block_range__create(start->addr, end->addr);
105	if (!block_range_iter__valid(&iter))
106		return;
107
108	/*
109	 * First block in range is a branch target.
110	 */
111	entry = block_range_iter(&iter);
112	assert(entry->is_target);
113	entry->entry++;
114
115	do {
116		entry = block_range_iter(&iter);
117
118		entry->coverage++;
119		entry->sym = sym;
120
121		if (notes)
122			notes->max_coverage = max(notes->max_coverage, entry->coverage);
123
124	} while (block_range_iter__next(&iter));
125
126	/*
127	 * Last block in rage is a branch.
128	 */
129	entry = block_range_iter(&iter);
130	assert(entry->is_branch);
131	entry->taken++;
132	if (flags->predicted)
133		entry->pred++;
134}
135
136static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
137				 struct perf_sample *sample)
138{
139	struct addr_map_symbol *prev = NULL;
140	struct branch_info *bi;
141	int i;
142
143	if (!bs || !bs->nr)
144		return;
145
146	bi = sample__resolve_bstack(sample, al);
147	if (!bi)
148		return;
149
150	for (i = bs->nr - 1; i >= 0; i--) {
151		/*
152		 * XXX filter against symbol
153		 */
154		if (prev)
155			process_basic_block(prev, &bi[i].from, &bi[i].flags);
156		prev = &bi[i].to;
157	}
158
159	free(bi);
160}
161
162static int hist_iter__branch_callback(struct hist_entry_iter *iter,
163				      struct addr_location *al __maybe_unused,
164				      bool single __maybe_unused,
165				      void *arg __maybe_unused)
166{
167	struct hist_entry *he = iter->he;
168	struct branch_info *bi;
169	struct perf_sample *sample = iter->sample;
170	struct evsel *evsel = iter->evsel;
171	int err;
172
173	bi = he->branch_info;
174	err = addr_map_symbol__inc_samples(&bi->from, sample, evsel);
175
176	if (err)
177		goto out;
178
179	err = addr_map_symbol__inc_samples(&bi->to, sample, evsel);
180
181out:
182	return err;
183}
184
185static int process_branch_callback(struct evsel *evsel,
186				   struct perf_sample *sample,
187				   struct addr_location *al __maybe_unused,
188				   struct perf_annotate *ann,
189				   struct machine *machine)
190{
191	struct hist_entry_iter iter = {
192		.evsel		= evsel,
193		.sample		= sample,
194		.add_entry_cb	= hist_iter__branch_callback,
195		.hide_unresolved	= symbol_conf.hide_unresolved,
196		.ops		= &hist_iter_branch,
197	};
198
199	struct addr_location a;
 
200
201	if (machine__resolve(machine, &a, sample) < 0)
202		return -1;
203
204	if (a.sym == NULL)
205		return 0;
206
207	if (a.map != NULL)
208		a.map->dso->hit = 1;
209
210	hist__account_cycles(sample->branch_stack, al, sample, false, NULL);
211
212	return hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
 
213}
214
215static bool has_annotation(struct perf_annotate *ann)
216{
217	return ui__has_annotation() || ann->use_stdio2;
218}
219
220static int evsel__add_sample(struct evsel *evsel, struct perf_sample *sample,
221			     struct addr_location *al, struct perf_annotate *ann,
222			     struct machine *machine)
 
 
223{
224	struct hists *hists = evsel__hists(evsel);
225	struct hist_entry *he;
226	int ret;
227
228	if ((!ann->has_br_stack || !has_annotation(ann)) &&
229	    ann->sym_hist_filter != NULL &&
230	    (al->sym == NULL ||
231	     strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
232		/* We're only interested in a symbol named sym_hist_filter */
233		/*
234		 * FIXME: why isn't this done in the symbol_filter when loading
235		 * the DSO?
236		 */
237		if (al->sym != NULL) {
238			rb_erase_cached(&al->sym->rb_node,
239				 &al->map->dso->symbols);
240			symbol__delete(al->sym);
241			dso__reset_find_symbol_cache(al->map->dso);
242		}
243		return 0;
244	}
245
246	/*
247	 * XXX filtered samples can still have branch entries pointing into our
248	 * symbol and are missed.
249	 */
250	process_branch_stack(sample->branch_stack, al, sample);
251
252	if (ann->has_br_stack && has_annotation(ann))
253		return process_branch_callback(evsel, sample, al, ann, machine);
254
255	he = hists__add_entry(hists, al, NULL, NULL, NULL, sample, true);
256	if (he == NULL)
257		return -ENOMEM;
258
259	ret = hist_entry__inc_addr_samples(he, sample, evsel, al->addr);
260	hists__inc_nr_samples(hists, true);
261	return ret;
262}
263
264static int process_sample_event(struct perf_tool *tool,
265				union perf_event *event,
266				struct perf_sample *sample,
267				struct evsel *evsel,
268				struct machine *machine)
269{
270	struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
271	struct addr_location al;
272	int ret = 0;
273
274	if (machine__resolve(machine, &al, sample) < 0) {
275		pr_warning("problem processing %d event, skipping it.\n",
276			   event->header.type);
277		return -1;
278	}
279
280	if (ann->cpu_list && !test_bit(sample->cpu, ann->cpu_bitmap))
281		goto out_put;
282
283	if (!al.filtered &&
284	    evsel__add_sample(evsel, sample, &al, ann, machine)) {
285		pr_warning("problem incrementing symbol count, "
286			   "skipping event\n");
287		ret = -1;
288	}
289out_put:
290	addr_location__put(&al);
291	return ret;
292}
293
294static int process_feature_event(struct perf_session *session,
295				 union perf_event *event)
296{
297	if (event->feat.feat_id < HEADER_LAST_FEATURE)
298		return perf_event__process_feature(session, event);
299	return 0;
300}
301
302static int hist_entry__tty_annotate(struct hist_entry *he,
303				    struct evsel *evsel,
304				    struct perf_annotate *ann)
305{
306	if (!ann->use_stdio2)
307		return symbol__tty_annotate(&he->ms, evsel, &ann->opts);
308
309	return symbol__tty_annotate2(&he->ms, evsel, &ann->opts);
310}
311
312static void hists__find_annotations(struct hists *hists,
313				    struct evsel *evsel,
314				    struct perf_annotate *ann)
315{
316	struct rb_node *nd = rb_first_cached(&hists->entries), *next;
317	int key = K_RIGHT;
318
319	while (nd) {
320		struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
321		struct annotation *notes;
322
323		if (he->ms.sym == NULL || he->ms.map->dso->annotate_warned)
324			goto find_next;
325
326		if (ann->sym_hist_filter &&
327		    (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
328			goto find_next;
329
330		if (ann->min_percent) {
331			float percent = 0;
332			u64 total = hists__total_period(hists);
333
334			if (total)
335				percent = 100.0 * he->stat.period / total;
336
337			if (percent < ann->min_percent)
338				goto find_next;
339		}
340
341		notes = symbol__annotation(he->ms.sym);
342		if (notes->src == NULL) {
343find_next:
344			if (key == K_LEFT)
345				nd = rb_prev(nd);
346			else
347				nd = rb_next(nd);
348			continue;
349		}
350
351		if (use_browser == 2) {
352			int ret;
353			int (*annotate)(struct hist_entry *he,
354					struct evsel *evsel,
355					struct hist_browser_timer *hbt);
356
357			annotate = dlsym(perf_gtk_handle,
358					 "hist_entry__gtk_annotate");
359			if (annotate == NULL) {
360				ui__error("GTK browser not found!\n");
361				return;
362			}
363
364			ret = annotate(he, evsel, NULL);
365			if (!ret || !ann->skip_missing)
366				return;
367
368			/* skip missing symbols */
369			nd = rb_next(nd);
370		} else if (use_browser == 1) {
371			key = hist_entry__tui_annotate(he, evsel, NULL, &ann->opts);
372
373			switch (key) {
374			case -1:
375				if (!ann->skip_missing)
376					return;
377				/* fall through */
378			case K_RIGHT:
379				next = rb_next(nd);
380				break;
381			case K_LEFT:
382				next = rb_prev(nd);
383				break;
384			default:
385				return;
386			}
387
388			if (next != NULL)
389				nd = next;
390		} else {
391			hist_entry__tty_annotate(he, evsel, ann);
392			nd = rb_next(nd);
 
 
 
 
 
 
 
393		}
394	}
395}
396
397static int __cmd_annotate(struct perf_annotate *ann)
398{
399	int ret;
400	struct perf_session *session = ann->session;
401	struct evsel *pos;
402	u64 total_nr_samples;
403
404	if (ann->cpu_list) {
405		ret = perf_session__cpu_bitmap(session, ann->cpu_list,
406					       ann->cpu_bitmap);
407		if (ret)
408			goto out;
409	}
410
411	if (!ann->opts.objdump_path) {
412		ret = perf_env__lookup_objdump(&session->header.env,
413					       &ann->opts.objdump_path);
414		if (ret)
415			goto out;
416	}
417
418	ret = perf_session__process_events(session);
419	if (ret)
420		goto out;
421
422	if (dump_trace) {
423		perf_session__fprintf_nr_events(session, stdout, false);
424		evlist__fprintf_nr_events(session->evlist, stdout, false);
425		goto out;
426	}
427
428	if (verbose > 3)
429		perf_session__fprintf(session, stdout);
430
431	if (verbose > 2)
432		perf_session__fprintf_dsos(session, stdout);
433
434	total_nr_samples = 0;
435	evlist__for_each_entry(session->evlist, pos) {
436		struct hists *hists = evsel__hists(pos);
437		u32 nr_samples = hists->stats.nr_samples;
438
439		if (nr_samples > 0) {
440			total_nr_samples += nr_samples;
441			hists__collapse_resort(hists, NULL);
442			/* Don't sort callchain */
443			evsel__reset_sample_bit(pos, CALLCHAIN);
444			evsel__output_resort(pos, NULL);
445
446			if (symbol_conf.event_group && !evsel__is_group_leader(pos))
 
447				continue;
448
449			hists__find_annotations(hists, pos, ann);
450		}
451	}
452
453	if (total_nr_samples == 0) {
454		ui__error("The %s data has no samples!\n", session->data->path);
455		goto out;
456	}
457
458	if (use_browser == 2) {
459		void (*show_annotations)(void);
460
461		show_annotations = dlsym(perf_gtk_handle,
462					 "perf_gtk__show_annotations");
463		if (show_annotations == NULL) {
464			ui__error("GTK browser not found!\n");
465			goto out;
466		}
467		show_annotations();
468	}
469
470out:
471	return ret;
472}
473
474static int parse_percent_limit(const struct option *opt, const char *str,
475			       int unset __maybe_unused)
476{
477	struct perf_annotate *ann = opt->value;
478	double pcnt = strtof(str, NULL);
479
480	ann->min_percent = pcnt;
481	return 0;
482}
483
484static const char * const annotate_usage[] = {
485	"perf annotate [<options>]",
486	NULL
487};
488
489int cmd_annotate(int argc, const char **argv)
490{
491	struct perf_annotate annotate = {
492		.tool = {
493			.sample	= process_sample_event,
494			.mmap	= perf_event__process_mmap,
495			.mmap2	= perf_event__process_mmap2,
496			.comm	= perf_event__process_comm,
497			.exit	= perf_event__process_exit,
498			.fork	= perf_event__process_fork,
499			.namespaces = perf_event__process_namespaces,
500			.attr	= perf_event__process_attr,
501			.build_id = perf_event__process_build_id,
502#ifdef HAVE_LIBTRACEEVENT
503			.tracing_data   = perf_event__process_tracing_data,
504#endif
505			.id_index	= perf_event__process_id_index,
506			.auxtrace_info	= perf_event__process_auxtrace_info,
507			.auxtrace	= perf_event__process_auxtrace,
508			.feature	= process_feature_event,
509			.ordered_events = true,
510			.ordering_requires_timestamps = true,
511		},
512		.opts = annotation__default_options,
513	};
514	struct perf_data data = {
515		.mode  = PERF_DATA_MODE_READ,
516	};
517	struct itrace_synth_opts itrace_synth_opts = {
518		.set = 0,
519	};
520	struct option options[] = {
521	OPT_STRING('i', "input", &input_name, "file",
522		    "input file name"),
523	OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
524		   "only consider symbols in these dsos"),
525	OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
526		    "symbol to annotate"),
527	OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
528	OPT_INCR('v', "verbose", &verbose,
529		    "be more verbose (show symbol address, etc)"),
530	OPT_BOOLEAN('q', "quiet", &quiet, "do now show any warnings or messages"),
531	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
532		    "dump raw trace in ASCII"),
533#ifdef HAVE_GTK2_SUPPORT
534	OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
535#endif
536#ifdef HAVE_SLANG_SUPPORT
537	OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
538#endif
539	OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
540	OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
541	OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
542                    "don't load vmlinux even if found"),
543	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
544		   "file", "vmlinux pathname"),
545	OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
546		    "load module symbols - WARNING: use only with -k and LIVE kernel"),
547	OPT_BOOLEAN('l', "print-line", &annotate.opts.print_lines,
548		    "print matching source lines (may be slow)"),
549	OPT_BOOLEAN('P', "full-paths", &annotate.opts.full_path,
550		    "Don't shorten the displayed pathnames"),
551	OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
552		    "Skip symbols that cannot be annotated"),
553	OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
554			&annotate.group_set,
555			"Show event group information together"),
556	OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
557	OPT_CALLBACK(0, "symfs", NULL, "directory",
558		     "Look for files with symbols relative to this directory",
559		     symbol__config_symfs),
560	OPT_BOOLEAN(0, "source", &annotate.opts.annotate_src,
561		    "Interleave source code with assembly code (default)"),
562	OPT_BOOLEAN(0, "asm-raw", &annotate.opts.show_asm_raw,
563		    "Display raw encoding of assembly instructions (default)"),
564	OPT_STRING('M', "disassembler-style", &annotate.opts.disassembler_style, "disassembler style",
565		   "Specify disassembler style (e.g. -M intel for intel syntax)"),
566	OPT_STRING(0, "prefix", &annotate.opts.prefix, "prefix",
567		    "Add prefix to source file path names in programs (with --prefix-strip)"),
568	OPT_STRING(0, "prefix-strip", &annotate.opts.prefix_strip, "N",
569		    "Strip first N entries of source file path name in programs (with --prefix)"),
570	OPT_STRING(0, "objdump", &annotate.opts.objdump_path, "path",
571		   "objdump binary to use for disassembly and annotations"),
572	OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
573		    "Enable symbol demangling"),
574	OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
575		    "Enable kernel symbol demangling"),
576	OPT_BOOLEAN(0, "group", &symbol_conf.event_group,
577		    "Show event group information together"),
578	OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
579		    "Show a column with the sum of periods"),
580	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
581		    "Show a column with the number of samples"),
582	OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
583			     "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
584			     stdio__config_color, "always"),
585	OPT_CALLBACK(0, "percent-type", &annotate.opts, "local-period",
586		     "Set percent type local/global-period/hits",
587		     annotate_parse_percent_type),
588	OPT_CALLBACK(0, "percent-limit", &annotate, "percent",
589		     "Don't show entries under that percent", parse_percent_limit),
590	OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
591			    "Instruction Tracing options\n" ITRACE_HELP,
592			    itrace_parse_synth_opts),
593
594	OPT_END()
595	};
596	int ret;
597
598	set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
599	set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
600
601
602	ret = hists__init();
603	if (ret < 0)
604		return ret;
605
606	annotation_config__init(&annotate.opts);
607
608	argc = parse_options(argc, argv, options, annotate_usage, 0);
609	if (argc) {
610		/*
611		 * Special case: if there's an argument left then assume that
612		 * it's a symbol filter:
613		 */
614		if (argc > 1)
615			usage_with_options(annotate_usage, options);
616
617		annotate.sym_hist_filter = argv[0];
618	}
619
620	if (annotate_check_args(&annotate.opts) < 0)
621		return -EINVAL;
622
623#ifdef HAVE_GTK2_SUPPORT
624	if (symbol_conf.show_nr_samples && annotate.use_gtk) {
625		pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
626		return ret;
627	}
628#endif
629
630	ret = symbol__validate_sym_arguments();
631	if (ret)
632		return ret;
633
634	if (quiet)
635		perf_quiet_option();
636
637	data.path = input_name;
638
639	annotate.session = perf_session__new(&data, &annotate.tool);
640	if (IS_ERR(annotate.session))
641		return PTR_ERR(annotate.session);
642
643	annotate.session->itrace_synth_opts = &itrace_synth_opts;
644
645	annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
646						      HEADER_BRANCH_STACK);
647
648	if (annotate.group_set)
649		evlist__force_leader(annotate.session->evlist);
650
651	ret = symbol__annotation_init();
652	if (ret < 0)
653		goto out_delete;
654
 
 
655	symbol_conf.try_vmlinux_path = true;
656
657	ret = symbol__init(&annotate.session->header.env);
658	if (ret < 0)
659		goto out_delete;
660
661	if (annotate.use_stdio || annotate.use_stdio2)
662		use_browser = 0;
663#ifdef HAVE_SLANG_SUPPORT
664	else if (annotate.use_tui)
665		use_browser = 1;
666#endif
667#ifdef HAVE_GTK2_SUPPORT
668	else if (annotate.use_gtk)
669		use_browser = 2;
670#endif
671
672	setup_browser(true);
673
674	/*
675	 * Events of different processes may correspond to the same
676	 * symbol, we do not care about the processes in annotate,
677	 * set sort order to avoid repeated output.
678	 */
679	sort_order = "dso,symbol";
680
681	/*
682	 * Set SORT_MODE__BRANCH so that annotate display IPC/Cycle
683	 * if branch info is in perf data in TUI mode.
684	 */
685	if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack)
686		sort__mode = SORT_MODE__BRANCH;
687
688	if (setup_sorting(NULL) < 0)
689		usage_with_options(annotate_usage, options);
 
 
 
690
691	ret = __cmd_annotate(&annotate);
692
693out_delete:
694	/*
695	 * Speed up the exit process, for large files this can
696	 * take quite a while.
697	 *
698	 * XXX Enable this when using valgrind or if we ever
699	 * librarize this command.
700	 *
701	 * Also experiment with obstacks to see how much speed
702	 * up we'll get here.
703	 *
704	 * perf_session__delete(session);
705	 */
706	return ret;
707}