1/*
2 * builtin-report.c
3 *
4 * Builtin report command: Analyze the perf.data input file,
5 * look up and read DSOs and symbol information and display
6 * a histogram of results, along various sorting keys.
7 */
8#include "builtin.h"
9
10#include "util/util.h"
11
12#include "util/color.h"
13#include <linux/list.h>
14#include "util/cache.h"
15#include <linux/rbtree.h>
16#include "util/symbol.h"
17#include "util/callchain.h"
18#include "util/strlist.h"
19#include "util/values.h"
20
21#include "perf.h"
22#include "util/debug.h"
23#include "util/header.h"
24#include "util/session.h"
25
26#include "util/parse-options.h"
27#include "util/parse-events.h"
28
29#include "util/thread.h"
30#include "util/sort.h"
31#include "util/hist.h"
32
33static char		const *input_name = "perf.data";
34
35static bool		force;
36static bool		hide_unresolved;
37static bool		dont_use_callchains;
38
39static bool		show_threads;
40static struct perf_read_values	show_threads_values;
41
42static const char	default_pretty_printing_style[] = "normal";
43static const char	*pretty_printing_style = default_pretty_printing_style;
44
45static char		callchain_default_opt[] = "fractal,0.5";
46
47static struct hists *perf_session__hists_findnew(struct perf_session *self,
48						 u64 event_stream, u32 type,
49						 u64 config)
50{
51	struct rb_node **p = &self->hists_tree.rb_node;
52	struct rb_node *parent = NULL;
53	struct hists *iter, *new;
54
55	while (*p != NULL) {
56		parent = *p;
57		iter = rb_entry(parent, struct hists, rb_node);
58		if (iter->config == config)
59			return iter;
60
61
62		if (config > iter->config)
63			p = &(*p)->rb_right;
64		else
65			p = &(*p)->rb_left;
66	}
67
68	new = malloc(sizeof(struct hists));
69	if (new == NULL)
70		return NULL;
71	memset(new, 0, sizeof(struct hists));
72	new->event_stream = event_stream;
73	new->config = config;
74	new->type = type;
75	rb_link_node(&new->rb_node, parent, p);
76	rb_insert_color(&new->rb_node, &self->hists_tree);
77	return new;
78}
79
80static int perf_session__add_hist_entry(struct perf_session *self,
81					struct addr_location *al,
82					struct sample_data *data)
83{
84	struct map_symbol *syms = NULL;
85	struct symbol *parent = NULL;
86	int err = -ENOMEM;
87	struct hist_entry *he;
88	struct hists *hists;
89	struct perf_event_attr *attr;
90
91	if ((sort__has_parent || symbol_conf.use_callchain) && data->callchain) {
92		syms = perf_session__resolve_callchain(self, al->thread,
93						       data->callchain, &parent);
94		if (syms == NULL)
95			return -ENOMEM;
96	}
97
98	attr = perf_header__find_attr(data->id, &self->header);
99	if (attr)
100		hists = perf_session__hists_findnew(self, data->id, attr->type, attr->config);
101	else
102		hists = perf_session__hists_findnew(self, data->id, 0, 0);
103	if (hists == NULL)
104		goto out_free_syms;
105	he = __hists__add_entry(hists, al, parent, data->period);
106	if (he == NULL)
107		goto out_free_syms;
108	err = 0;
109	if (symbol_conf.use_callchain) {
110		err = append_chain(he->callchain, data->callchain, syms, data->period);
111		if (err)
112			goto out_free_syms;
113	}
114	/*
115	 * Only in the newt browser we are doing integrated annotation,
116	 * so we don't allocated the extra space needed because the stdio
117	 * code will not use it.
118	 */
119	if (use_browser > 0)
120		err = hist_entry__inc_addr_samples(he, al->addr);
121out_free_syms:
122	free(syms);
123	return err;
124}
125
126static int add_event_total(struct perf_session *session,
127			   struct sample_data *data,
128			   struct perf_event_attr *attr)
129{
130	struct hists *hists;
131
132	if (attr)
133		hists = perf_session__hists_findnew(session, data->id,
134						    attr->type, attr->config);
135	else
136		hists = perf_session__hists_findnew(session, data->id, 0, 0);
137
138	if (!hists)
139		return -ENOMEM;
140
141	hists->stats.total_period += data->period;
142	hists__inc_nr_events(hists, PERF_RECORD_SAMPLE);
143	session->hists.stats.total_period += data->period;
144	return 0;
145}
146
147static int process_sample_event(event_t *event, struct perf_session *session)
148{
149	struct sample_data data = { .period = 1, };
150	struct addr_location al;
151	struct perf_event_attr *attr;
152
153	if (event__preprocess_sample(event, session, &al, &data, NULL) < 0) {
154		fprintf(stderr, "problem processing %d event, skipping it.\n",
155			event->header.type);
156		return -1;
157	}
158
159	if (al.filtered || (hide_unresolved && al.sym == NULL))
160		return 0;
161
162	if (perf_session__add_hist_entry(session, &al, &data)) {
163		pr_debug("problem incrementing symbol period, skipping event\n");
164		return -1;
165	}
166
167	attr = perf_header__find_attr(data.id, &session->header);
168
169	if (add_event_total(session, &data, attr)) {
170		pr_debug("problem adding event period\n");
171		return -1;
172	}
173
174	return 0;
175}
176
177static int process_read_event(event_t *event, struct perf_session *session __used)
178{
179	struct perf_event_attr *attr;
180
181	attr = perf_header__find_attr(event->read.id, &session->header);
182
183	if (show_threads) {
184		const char *name = attr ? __event_name(attr->type, attr->config)
185				   : "unknown";
186		perf_read_values_add_value(&show_threads_values,
187					   event->read.pid, event->read.tid,
188					   event->read.id,
189					   name,
190					   event->read.value);
191	}
192
193	dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
194		    attr ? __event_name(attr->type, attr->config) : "FAIL",
195		    event->read.value);
196
197	return 0;
198}
199
200static int perf_session__setup_sample_type(struct perf_session *self)
201{
202	if (!(self->sample_type & PERF_SAMPLE_CALLCHAIN)) {
203		if (sort__has_parent) {
204			fprintf(stderr, "selected --sort parent, but no"
205					" callchain data. Did you call"
206					" perf record without -g?\n");
207			return -EINVAL;
208		}
209		if (symbol_conf.use_callchain) {
210			fprintf(stderr, "selected -g but no callchain data."
211					" Did you call perf record without"
212					" -g?\n");
213			return -1;
214		}
215	} else if (!dont_use_callchains && callchain_param.mode != CHAIN_NONE &&
216		   !symbol_conf.use_callchain) {
217			symbol_conf.use_callchain = true;
218			if (register_callchain_param(&callchain_param) < 0) {
219				fprintf(stderr, "Can't register callchain"
220						" params\n");
221				return -EINVAL;
222			}
223	}
224
225	return 0;
226}
227
228static struct perf_event_ops event_ops = {
229	.sample	= process_sample_event,
230	.mmap	= event__process_mmap,
231	.comm	= event__process_comm,
232	.exit	= event__process_task,
233	.fork	= event__process_task,
234	.lost	= event__process_lost,
235	.read	= process_read_event,
236	.attr	= event__process_attr,
237	.event_type = event__process_event_type,
238	.tracing_data = event__process_tracing_data,
239	.build_id = event__process_build_id,
240};
241
242extern volatile int session_done;
243
244static void sig_handler(int sig __used)
245{
246	session_done = 1;
247}
248
249static size_t hists__fprintf_nr_sample_events(struct hists *self,
250					      const char *evname, FILE *fp)
251{
252	size_t ret;
253	char unit;
254	unsigned long nr_events = self->stats.nr_events[PERF_RECORD_SAMPLE];
255
256	nr_events = convert_unit(nr_events, &unit);
257	ret = fprintf(fp, "# Events: %lu%c", nr_events, unit);
258	if (evname != NULL)
259		ret += fprintf(fp, " %s", evname);
260	return ret + fprintf(fp, "\n#\n");
261}
262
263static int hists__tty_browse_tree(struct rb_root *tree, const char *help)
264{
265	struct rb_node *next = rb_first(tree);
266
267	while (next) {
268		struct hists *hists = rb_entry(next, struct hists, rb_node);
269		const char *evname = NULL;
270
271		if (rb_first(&hists->entries) != rb_last(&hists->entries))
272			evname = __event_name(hists->type, hists->config);
273
274		hists__fprintf_nr_sample_events(hists, evname, stdout);
275		hists__fprintf(hists, NULL, false, stdout);
276		fprintf(stdout, "\n\n");
277		next = rb_next(&hists->rb_node);
278	}
279
280	if (sort_order == default_sort_order &&
281	    parent_pattern == default_parent_pattern) {
282		fprintf(stdout, "#\n# (%s)\n#\n", help);
283
284		if (show_threads) {
285			bool style = !strcmp(pretty_printing_style, "raw");
286			perf_read_values_display(stdout, &show_threads_values,
287						 style);
288			perf_read_values_destroy(&show_threads_values);
289		}
290	}
291
292	return 0;
293}
294
295static int __cmd_report(void)
296{
297	int ret = -EINVAL;
298	struct perf_session *session;
299	struct rb_node *next;
300	const char *help = "For a higher level overview, try: perf report --sort comm,dso";
301
302	signal(SIGINT, sig_handler);
303
304	session = perf_session__new(input_name, O_RDONLY, force, false);
305	if (session == NULL)
306		return -ENOMEM;
307
308	if (show_threads)
309		perf_read_values_init(&show_threads_values);
310
311	ret = perf_session__setup_sample_type(session);
312	if (ret)
313		goto out_delete;
314
315	ret = perf_session__process_events(session, &event_ops);
316	if (ret)
317		goto out_delete;
318
319	if (dump_trace) {
320		perf_session__fprintf_nr_events(session, stdout);
321		goto out_delete;
322	}
323
324	if (verbose > 3)
325		perf_session__fprintf(session, stdout);
326
327	if (verbose > 2)
328		perf_session__fprintf_dsos(session, stdout);
329
330	next = rb_first(&session->hists_tree);
331	while (next) {
332		struct hists *hists;
333
334		hists = rb_entry(next, struct hists, rb_node);
335		hists__collapse_resort(hists);
336		hists__output_resort(hists);
337		next = rb_next(&hists->rb_node);
338	}
339
340	if (use_browser > 0)
341		hists__tui_browse_tree(&session->hists_tree, help);
342	else
343		hists__tty_browse_tree(&session->hists_tree, help);
344
345out_delete:
346	return ret;
347}
348
349static int
350parse_callchain_opt(const struct option *opt __used, const char *arg,
351		    int unset)
352{
353	char *tok, *tok2;
354	char *endptr;
355
356	/*
357	 * --no-call-graph
358	 */
359	if (unset) {
360		dont_use_callchains = true;
361		return 0;
362	}
363
364	symbol_conf.use_callchain = true;
365
366	if (!arg)
367		return 0;
368
369	tok = strtok((char *)arg, ",");
370	if (!tok)
371		return -1;
372
373	/* get the output mode */
374	if (!strncmp(tok, "graph", strlen(arg)))
375		callchain_param.mode = CHAIN_GRAPH_ABS;
376
377	else if (!strncmp(tok, "flat", strlen(arg)))
378		callchain_param.mode = CHAIN_FLAT;
379
380	else if (!strncmp(tok, "fractal", strlen(arg)))
381		callchain_param.mode = CHAIN_GRAPH_REL;
382
383	else if (!strncmp(tok, "none", strlen(arg))) {
384		callchain_param.mode = CHAIN_NONE;
385		symbol_conf.use_callchain = false;
386
387		return 0;
388	}
389
390	else
391		return -1;
392
393	/* get the min percentage */
394	tok = strtok(NULL, ",");
395	if (!tok)
396		goto setup;
397
398	tok2 = strtok(NULL, ",");
399	callchain_param.min_percent = strtod(tok, &endptr);
400	if (tok == endptr)
401		return -1;
402
403	if (tok2)
404		callchain_param.print_limit = strtod(tok2, &endptr);
405setup:
406	if (register_callchain_param(&callchain_param) < 0) {
407		fprintf(stderr, "Can't register callchain params\n");
408		return -1;
409	}
410	return 0;
411}
412
413static const char * const report_usage[] = {
414	"perf report [<options>] <command>",
415	NULL
416};
417
418static const struct option options[] = {
419	OPT_STRING('i', "input", &input_name, "file",
420		    "input file name"),
421	OPT_INCR('v', "verbose", &verbose,
422		    "be more verbose (show symbol address, etc)"),
423	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
424		    "dump raw trace in ASCII"),
425	OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
426		   "file", "vmlinux pathname"),
427	OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
428	OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
429		    "load module symbols - WARNING: use only with -k and LIVE kernel"),
430	OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
431		    "Show a column with the number of samples"),
432	OPT_BOOLEAN('T', "threads", &show_threads,
433		    "Show per-thread event counters"),
434	OPT_STRING(0, "pretty", &pretty_printing_style, "key",
435		   "pretty printing style key: normal raw"),
436	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
437		   "sort by key(s): pid, comm, dso, symbol, parent"),
438	OPT_BOOLEAN(0, "showcpuutilization", &symbol_conf.show_cpu_utilization,
439		    "Show sample percentage for different cpu modes"),
440	OPT_STRING('p', "parent", &parent_pattern, "regex",
441		   "regex filter to identify parent, see: '--sort parent'"),
442	OPT_BOOLEAN('x', "exclude-other", &symbol_conf.exclude_other,
443		    "Only display entries with parent-match"),
444	OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
445		     "Display callchains using output_type (graph, flat, fractal, or none) and min percent threshold. "
446		     "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
447	OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
448		   "only consider symbols in these dsos"),
449	OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
450		   "only consider symbols in these comms"),
451	OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
452		   "only consider these symbols"),
453	OPT_STRING('w', "column-widths", &symbol_conf.col_width_list_str,
454		   "width[,width...]",
455		   "don't try to adjust column width, use these fixed values"),
456	OPT_STRING('t', "field-separator", &symbol_conf.field_sep, "separator",
457		   "separator for columns, no spaces will be added between "
458		   "columns '.' is reserved."),
459	OPT_BOOLEAN('U', "hide-unresolved", &hide_unresolved,
460		    "Only display entries resolved to a symbol"),
461	OPT_END()
462};
463
464int cmd_report(int argc, const char **argv, const char *prefix __used)
465{
466	argc = parse_options(argc, argv, options, report_usage, 0);
467
468	if (strcmp(input_name, "-") != 0)
469		setup_browser();
470	/*
471	 * Only in the newt browser we are doing integrated annotation,
472	 * so don't allocate extra space that won't be used in the stdio
473	 * implementation.
474	 */
475	if (use_browser > 0) {
476		symbol_conf.priv_size = sizeof(struct sym_priv);
477		/*
478 		 * For searching by name on the "Browse map details".
479 		 * providing it only in verbose mode not to bloat too
480 		 * much struct symbol.
481 		 */
482		if (verbose) {
483			symbol_conf.priv_size += sizeof(u32);
484			symbol_conf.sort_by_name = true;
485		}
486	}
487
488	if (symbol__init() < 0)
489		return -1;
490
491	setup_sorting(report_usage, options);
492
493	if (parent_pattern != default_parent_pattern) {
494		if (sort_dimension__add("parent") < 0)
495			return -1;
496		sort_parent.elide = 1;
497	} else
498		symbol_conf.exclude_other = false;
499
500	/*
501	 * Any (unrecognized) arguments left?
502	 */
503	if (argc)
504		usage_with_options(report_usage, options);
505
506	sort_entry__setup_elide(&sort_dso, symbol_conf.dso_list, "dso", stdout);
507	sort_entry__setup_elide(&sort_comm, symbol_conf.comm_list, "comm", stdout);
508	sort_entry__setup_elide(&sort_sym, symbol_conf.sym_list, "symbol", stdout);
509
510	return __cmd_report();
511}
512