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.
10 #include "util/util.h"
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/string.h"
18 #include "util/callchain.h"
19 #include "util/strlist.h"
20 #include "util/values.h"
23 #include "util/debug.h"
24 #include "util/header.h"
25 #include "util/session.h"
27 #include "util/parse-options.h"
28 #include "util/parse-events.h"
30 #include "util/thread.h"
31 #include "util/sort.h"
32 #include "util/hist.h"
34 static char const *input_name = "perf.data";
36 static char *dso_list_str, *comm_list_str, *sym_list_str,
38 static struct strlist *dso_list, *comm_list, *sym_list;
42 static int show_nr_samples;
44 static int show_threads;
45 static struct perf_read_values show_threads_values;
47 static char default_pretty_printing_style[] = "normal";
48 static char *pretty_printing_style = default_pretty_printing_style;
50 static int exclude_other = 1;
52 static char callchain_default_opt[] = "fractal,0.5";
54 static u64 sample_type;
56 struct symbol_conf symbol_conf;
60 callchain__fprintf_left_margin(FILE *fp, int left_margin)
65 ret = fprintf(fp, " ");
67 for (i = 0; i < left_margin; i++)
68 ret += fprintf(fp, " ");
73 static size_t ipchain__fprintf_graph_line(FILE *fp, int depth, int depth_mask,
79 ret += callchain__fprintf_left_margin(fp, left_margin);
81 for (i = 0; i < depth; i++)
82 if (depth_mask & (1 << i))
83 ret += fprintf(fp, "| ");
85 ret += fprintf(fp, " ");
87 ret += fprintf(fp, "\n");
92 ipchain__fprintf_graph(FILE *fp, struct callchain_list *chain, int depth,
93 int depth_mask, int count, u64 total_samples,
94 int hits, int left_margin)
99 ret += callchain__fprintf_left_margin(fp, left_margin);
100 for (i = 0; i < depth; i++) {
101 if (depth_mask & (1 << i))
102 ret += fprintf(fp, "|");
104 ret += fprintf(fp, " ");
105 if (!count && i == depth - 1) {
108 percent = hits * 100.0 / total_samples;
109 ret += percent_color_fprintf(fp, "--%2.2f%%-- ", percent);
111 ret += fprintf(fp, "%s", " ");
114 ret += fprintf(fp, "%s\n", chain->sym->name);
116 ret += fprintf(fp, "%p\n", (void *)(long)chain->ip);
121 static struct symbol *rem_sq_bracket;
122 static struct callchain_list rem_hits;
124 static void init_rem_hits(void)
126 rem_sq_bracket = malloc(sizeof(*rem_sq_bracket) + 6);
127 if (!rem_sq_bracket) {
128 fprintf(stderr, "Not enough memory to display remaining hits\n");
132 strcpy(rem_sq_bracket->name, "[...]");
133 rem_hits.sym = rem_sq_bracket;
137 __callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
138 u64 total_samples, int depth, int depth_mask,
141 struct rb_node *node, *next;
142 struct callchain_node *child;
143 struct callchain_list *chain;
144 int new_depth_mask = depth_mask;
150 if (callchain_param.mode == CHAIN_GRAPH_REL)
151 new_total = self->children_hit;
153 new_total = total_samples;
155 remaining = new_total;
157 node = rb_first(&self->rb_root);
161 child = rb_entry(node, struct callchain_node, rb_node);
162 cumul = cumul_hits(child);
166 * The depth mask manages the output of pipes that show
167 * the depth. We don't want to keep the pipes of the current
168 * level for the last child of this depth.
169 * Except if we have remaining filtered hits. They will
170 * supersede the last child
172 next = rb_next(node);
173 if (!next && (callchain_param.mode != CHAIN_GRAPH_REL || !remaining))
174 new_depth_mask &= ~(1 << (depth - 1));
177 * But we keep the older depth mask for the line seperator
178 * to keep the level link until we reach the last child
180 ret += ipchain__fprintf_graph_line(fp, depth, depth_mask,
183 list_for_each_entry(chain, &child->val, list) {
184 if (chain->ip >= PERF_CONTEXT_MAX)
186 ret += ipchain__fprintf_graph(fp, chain, depth,
192 ret += __callchain__fprintf_graph(fp, child, new_total,
194 new_depth_mask | (1 << depth),
199 if (callchain_param.mode == CHAIN_GRAPH_REL &&
200 remaining && remaining != new_total) {
205 new_depth_mask &= ~(1 << (depth - 1));
207 ret += ipchain__fprintf_graph(fp, &rem_hits, depth,
208 new_depth_mask, 0, new_total,
209 remaining, left_margin);
217 callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
218 u64 total_samples, int left_margin)
220 struct callchain_list *chain;
221 bool printed = false;
225 list_for_each_entry(chain, &self->val, list) {
226 if (chain->ip >= PERF_CONTEXT_MAX)
229 if (!i++ && sort__first_dimension == SORT_SYM)
233 ret += callchain__fprintf_left_margin(fp, left_margin);
234 ret += fprintf(fp, "|\n");
235 ret += callchain__fprintf_left_margin(fp, left_margin);
236 ret += fprintf(fp, "---");
241 ret += callchain__fprintf_left_margin(fp, left_margin);
244 ret += fprintf(fp, " %s\n", chain->sym->name);
246 ret += fprintf(fp, " %p\n", (void *)(long)chain->ip);
249 ret += __callchain__fprintf_graph(fp, self, total_samples, 1, 1, left_margin);
255 callchain__fprintf_flat(FILE *fp, struct callchain_node *self,
258 struct callchain_list *chain;
264 ret += callchain__fprintf_flat(fp, self->parent, total_samples);
267 list_for_each_entry(chain, &self->val, list) {
268 if (chain->ip >= PERF_CONTEXT_MAX)
271 ret += fprintf(fp, " %s\n", chain->sym->name);
273 ret += fprintf(fp, " %p\n",
274 (void *)(long)chain->ip);
281 hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
282 u64 total_samples, int left_margin)
284 struct rb_node *rb_node;
285 struct callchain_node *chain;
288 rb_node = rb_first(&self->sorted_chain);
292 chain = rb_entry(rb_node, struct callchain_node, rb_node);
293 percent = chain->hit * 100.0 / total_samples;
294 switch (callchain_param.mode) {
296 ret += percent_color_fprintf(fp, " %6.2f%%\n",
298 ret += callchain__fprintf_flat(fp, chain, total_samples);
300 case CHAIN_GRAPH_ABS: /* Falldown */
301 case CHAIN_GRAPH_REL:
302 ret += callchain__fprintf_graph(fp, chain, total_samples,
308 ret += fprintf(fp, "\n");
309 rb_node = rb_next(rb_node);
316 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
318 struct sort_entry *se;
321 if (exclude_other && !self->parent)
325 ret = percent_color_fprintf(fp,
326 field_sep ? "%.2f" : " %6.2f%%",
327 (self->count * 100.0) / total_samples);
329 ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count);
331 if (show_nr_samples) {
333 fprintf(fp, "%c%lld", *field_sep, self->count);
335 fprintf(fp, "%11lld", self->count);
338 list_for_each_entry(se, &hist_entry__sort_list, list) {
342 fprintf(fp, "%s", field_sep ?: " ");
343 ret += se->print(fp, self, se->width ? *se->width : 0);
346 ret += fprintf(fp, "\n");
351 if (sort__first_dimension == SORT_COMM) {
352 se = list_first_entry(&hist_entry__sort_list, typeof(*se),
354 left_margin = se->width ? *se->width : 0;
355 left_margin -= thread__comm_len(self->thread);
358 hist_entry_callchain__fprintf(fp, self, total_samples,
369 static void dso__calc_col_width(struct dso *self)
371 if (!col_width_list_str && !field_sep &&
372 (!dso_list || strlist__has_entry(dso_list, self->name))) {
373 unsigned int slen = strlen(self->name);
374 if (slen > dsos__col_width)
375 dsos__col_width = slen;
378 self->slen_calculated = 1;
381 static void thread__comm_adjust(struct thread *self)
383 char *comm = self->comm;
385 if (!col_width_list_str && !field_sep &&
386 (!comm_list || strlist__has_entry(comm_list, comm))) {
387 unsigned int slen = strlen(comm);
389 if (slen > comms__col_width) {
390 comms__col_width = slen;
391 threads__col_width = slen + 6;
396 static int thread__set_comm_adjust(struct thread *self, const char *comm)
398 int ret = thread__set_comm(self, comm);
403 thread__comm_adjust(self);
408 static int call__match(struct symbol *sym)
410 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
416 static struct symbol **resolve_callchain(struct thread *thread,
417 struct perf_session *session,
418 struct ip_callchain *chain,
419 struct symbol **parent)
421 u8 cpumode = PERF_RECORD_MISC_USER;
422 struct symbol **syms = NULL;
426 syms = calloc(chain->nr, sizeof(*syms));
428 fprintf(stderr, "Can't allocate memory for symbols\n");
433 for (i = 0; i < chain->nr; i++) {
434 u64 ip = chain->ips[i];
435 struct addr_location al;
437 if (ip >= PERF_CONTEXT_MAX) {
439 case PERF_CONTEXT_HV:
440 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
441 case PERF_CONTEXT_KERNEL:
442 cpumode = PERF_RECORD_MISC_KERNEL; break;
443 case PERF_CONTEXT_USER:
444 cpumode = PERF_RECORD_MISC_USER; break;
451 thread__find_addr_location(thread, session, cpumode,
452 MAP__FUNCTION, ip, &al, NULL);
453 if (al.sym != NULL) {
454 if (sort__has_parent && !*parent &&
467 * collect histogram counts
470 static int hist_entry__add(struct addr_location *al,
471 struct perf_session *session,
472 struct ip_callchain *chain, u64 count)
474 struct symbol **syms = NULL, *parent = NULL;
476 struct hist_entry *he;
478 if ((sort__has_parent || callchain) && chain)
479 syms = resolve_callchain(al->thread, session, chain, &parent);
481 he = __hist_entry__add(al, parent, count, &hit);
490 callchain_init(&he->callchain);
491 append_chain(&he->callchain, chain, syms);
498 static size_t output__fprintf(FILE *fp, u64 total_samples)
500 struct hist_entry *pos;
501 struct sort_entry *se;
505 char *col_width = col_width_list_str;
506 int raw_printing_style;
508 raw_printing_style = !strcmp(pretty_printing_style, "raw");
512 fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
515 fprintf(fp, "# Overhead");
516 if (show_nr_samples) {
518 fprintf(fp, "%cSamples", *field_sep);
520 fputs(" Samples ", fp);
522 list_for_each_entry(se, &hist_entry__sort_list, list) {
526 fprintf(fp, "%c%s", *field_sep, se->header);
529 width = strlen(se->header);
531 if (col_width_list_str) {
533 *se->width = atoi(col_width);
534 col_width = strchr(col_width, ',');
539 width = *se->width = max(*se->width, width);
541 fprintf(fp, " %*s", width, se->header);
548 fprintf(fp, "# ........");
550 fprintf(fp, " ..........");
551 list_for_each_entry(se, &hist_entry__sort_list, list) {
561 width = strlen(se->header);
562 for (i = 0; i < width; i++)
570 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
571 pos = rb_entry(nd, struct hist_entry, rb_node);
572 ret += hist_entry__fprintf(fp, pos, total_samples);
575 if (sort_order == default_sort_order &&
576 parent_pattern == default_parent_pattern) {
578 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
583 free(rem_sq_bracket);
586 perf_read_values_display(fp, &show_threads_values,
592 static int validate_chain(struct ip_callchain *chain, event_t *event)
594 unsigned int chain_size;
596 chain_size = event->header.size;
597 chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
599 if (chain->nr*sizeof(u64) > chain_size)
605 static int process_sample_event(event_t *event, struct perf_session *session)
607 struct sample_data data;
609 struct addr_location al;
610 struct thread *thread;
612 memset(&data, 0, sizeof(data));
615 event__parse_sample(event, sample_type, &data);
617 dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
620 (void *)(long)data.ip,
621 (long long)data.period);
623 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
626 dump_printf("... chain: nr:%Lu\n", data.callchain->nr);
628 if (validate_chain(data.callchain, event) < 0) {
629 pr_debug("call-chain problem with event, "
635 for (i = 0; i < data.callchain->nr; i++)
636 dump_printf("..... %2d: %016Lx\n",
637 i, data.callchain->ips[i]);
641 thread = perf_session__findnew(session, data.pid);
642 if (thread == NULL) {
643 pr_debug("problem processing %d event, skipping it.\n",
648 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
650 if (comm_list && !strlist__has_entry(comm_list, thread->comm))
653 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
655 thread__find_addr_location(thread, session, cpumode,
656 MAP__FUNCTION, data.ip, &al, NULL);
658 * We have to do this here as we may have a dso with no symbol hit that
659 * has a name longer than the ones with symbols sampled.
661 if (al.map && !sort_dso.elide && !al.map->dso->slen_calculated)
662 dso__calc_col_width(al.map->dso);
665 (!al.map || !al.map->dso ||
666 !(strlist__has_entry(dso_list, al.map->dso->short_name) ||
667 (al.map->dso->short_name != al.map->dso->long_name &&
668 strlist__has_entry(dso_list, al.map->dso->long_name)))))
671 if (sym_list && al.sym && !strlist__has_entry(sym_list, al.sym->name))
674 if (hist_entry__add(&al, session, data.callchain, data.period)) {
675 pr_debug("problem incrementing symbol count, skipping event\n");
679 event__stats.total += data.period;
684 static int process_comm_event(event_t *event, struct perf_session *session)
686 struct thread *thread = perf_session__findnew(session, event->comm.pid);
688 dump_printf(": %s:%d\n", event->comm.comm, event->comm.pid);
690 if (thread == NULL ||
691 thread__set_comm_adjust(thread, event->comm.comm)) {
692 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
699 static int process_read_event(event_t *event, struct perf_session *session __used)
701 struct perf_event_attr *attr;
703 attr = perf_header__find_attr(event->read.id, &session->header);
706 const char *name = attr ? __event_name(attr->type, attr->config)
708 perf_read_values_add_value(&show_threads_values,
709 event->read.pid, event->read.tid,
715 dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
716 attr ? __event_name(attr->type, attr->config) : "FAIL",
722 static int sample_type_check(u64 type)
726 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
727 if (sort__has_parent) {
728 fprintf(stderr, "selected --sort parent, but no"
729 " callchain data. Did you call"
730 " perf record without -g?\n");
734 fprintf(stderr, "selected -g but no callchain data."
735 " Did you call perf record without"
739 } else if (callchain_param.mode != CHAIN_NONE && !callchain) {
741 if (register_callchain_param(&callchain_param) < 0) {
742 fprintf(stderr, "Can't register callchain"
751 static struct perf_event_ops event_ops = {
752 .process_sample_event = process_sample_event,
753 .process_mmap_event = event__process_mmap,
754 .process_comm_event = process_comm_event,
755 .process_exit_event = event__process_task,
756 .process_fork_event = event__process_task,
757 .process_lost_event = event__process_lost,
758 .process_read_event = process_read_event,
759 .sample_type_check = sample_type_check,
763 static int __cmd_report(void)
766 struct perf_session *session;
768 session = perf_session__new(input_name, O_RDONLY, force, &symbol_conf);
773 perf_read_values_init(&show_threads_values);
775 ret = perf_session__process_events(session, &event_ops);
780 event__print_totals();
785 perf_session__fprintf(session, stdout);
788 dsos__fprintf(stdout);
791 output__resort(event__stats.total);
792 output__fprintf(stdout, event__stats.total);
795 perf_read_values_destroy(&show_threads_values);
797 perf_session__delete(session);
802 parse_callchain_opt(const struct option *opt __used, const char *arg,
813 tok = strtok((char *)arg, ",");
817 /* get the output mode */
818 if (!strncmp(tok, "graph", strlen(arg)))
819 callchain_param.mode = CHAIN_GRAPH_ABS;
821 else if (!strncmp(tok, "flat", strlen(arg)))
822 callchain_param.mode = CHAIN_FLAT;
824 else if (!strncmp(tok, "fractal", strlen(arg)))
825 callchain_param.mode = CHAIN_GRAPH_REL;
827 else if (!strncmp(tok, "none", strlen(arg))) {
828 callchain_param.mode = CHAIN_NONE;
837 /* get the min percentage */
838 tok = strtok(NULL, ",");
842 callchain_param.min_percent = strtod(tok, &endptr);
847 if (register_callchain_param(&callchain_param) < 0) {
848 fprintf(stderr, "Can't register callchain params\n");
854 //static const char * const report_usage[] = {
855 const char * const report_usage[] = {
856 "perf report [<options>] <command>",
860 static const struct option options[] = {
861 OPT_STRING('i', "input", &input_name, "file",
863 OPT_BOOLEAN('v', "verbose", &verbose,
864 "be more verbose (show symbol address, etc)"),
865 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
866 "dump raw trace in ASCII"),
867 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
868 "file", "vmlinux pathname"),
869 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
870 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
871 "load module symbols - WARNING: use only with -k and LIVE kernel"),
872 OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
873 "Show a column with the number of samples"),
874 OPT_BOOLEAN('T', "threads", &show_threads,
875 "Show per-thread event counters"),
876 OPT_STRING(0, "pretty", &pretty_printing_style, "key",
877 "pretty printing style key: normal raw"),
878 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
879 "sort by key(s): pid, comm, dso, symbol, parent"),
880 OPT_BOOLEAN('P', "full-paths", &event_ops.full_paths,
881 "Don't shorten the pathnames taking into account the cwd"),
882 OPT_STRING('p', "parent", &parent_pattern, "regex",
883 "regex filter to identify parent, see: '--sort parent'"),
884 OPT_BOOLEAN('x', "exclude-other", &exclude_other,
885 "Only display entries with parent-match"),
886 OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
887 "Display callchains using output_type and min percent threshold. "
888 "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
889 OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
890 "only consider symbols in these dsos"),
891 OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
892 "only consider symbols in these comms"),
893 OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
894 "only consider these symbols"),
895 OPT_STRING('w', "column-widths", &col_width_list_str,
897 "don't try to adjust column width, use these fixed values"),
898 OPT_STRING('t', "field-separator", &field_sep, "separator",
899 "separator for columns, no spaces will be added between "
900 "columns '.' is reserved."),
904 static void setup_sorting(void)
906 char *tmp, *tok, *str = strdup(sort_order);
908 for (tok = strtok_r(str, ", ", &tmp);
909 tok; tok = strtok_r(NULL, ", ", &tmp)) {
910 if (sort_dimension__add(tok) < 0) {
911 error("Unknown --sort key: `%s'", tok);
912 usage_with_options(report_usage, options);
919 static void setup_list(struct strlist **list, const char *list_str,
920 struct sort_entry *se, const char *list_name,
924 *list = strlist__new(true, list_str);
926 fprintf(stderr, "problems parsing %s list\n",
930 if (strlist__nr_entries(*list) == 1) {
931 fprintf(fp, "# %s: %s\n", list_name,
932 strlist__entry(*list, 0)->s);
938 int cmd_report(int argc, const char **argv, const char *prefix __used)
940 if (symbol__init(&symbol_conf) < 0)
943 argc = parse_options(argc, argv, options, report_usage, 0);
947 if (parent_pattern != default_parent_pattern) {
948 sort_dimension__add("parent");
949 sort_parent.elide = 1;
954 * Any (unrecognized) arguments left?
957 usage_with_options(report_usage, options);
961 setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
962 setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
963 setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
965 if (field_sep && *field_sep == '.') {
966 fputs("'.' is the only non valid --field-separator argument\n",
971 return __cmd_report();