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;
41 static bool use_callchain;
43 static int show_nr_samples;
45 static int show_threads;
46 static struct perf_read_values show_threads_values;
48 static char default_pretty_printing_style[] = "normal";
49 static char *pretty_printing_style = default_pretty_printing_style;
51 static int exclude_other = 1;
53 static char callchain_default_opt[] = "fractal,0.5";
55 static u64 sample_type;
57 struct symbol_conf symbol_conf;
61 callchain__fprintf_left_margin(FILE *fp, int left_margin)
66 ret = fprintf(fp, " ");
68 for (i = 0; i < left_margin; i++)
69 ret += fprintf(fp, " ");
74 static size_t ipchain__fprintf_graph_line(FILE *fp, int depth, int depth_mask,
80 ret += callchain__fprintf_left_margin(fp, left_margin);
82 for (i = 0; i < depth; i++)
83 if (depth_mask & (1 << i))
84 ret += fprintf(fp, "| ");
86 ret += fprintf(fp, " ");
88 ret += fprintf(fp, "\n");
93 ipchain__fprintf_graph(FILE *fp, struct callchain_list *chain, int depth,
94 int depth_mask, int count, u64 total_samples,
95 int hits, int left_margin)
100 ret += callchain__fprintf_left_margin(fp, left_margin);
101 for (i = 0; i < depth; i++) {
102 if (depth_mask & (1 << i))
103 ret += fprintf(fp, "|");
105 ret += fprintf(fp, " ");
106 if (!count && i == depth - 1) {
109 percent = hits * 100.0 / total_samples;
110 ret += percent_color_fprintf(fp, "--%2.2f%%-- ", percent);
112 ret += fprintf(fp, "%s", " ");
115 ret += fprintf(fp, "%s\n", chain->sym->name);
117 ret += fprintf(fp, "%p\n", (void *)(long)chain->ip);
122 static struct symbol *rem_sq_bracket;
123 static struct callchain_list rem_hits;
125 static void init_rem_hits(void)
127 rem_sq_bracket = malloc(sizeof(*rem_sq_bracket) + 6);
128 if (!rem_sq_bracket) {
129 fprintf(stderr, "Not enough memory to display remaining hits\n");
133 strcpy(rem_sq_bracket->name, "[...]");
134 rem_hits.sym = rem_sq_bracket;
138 __callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
139 u64 total_samples, int depth, int depth_mask,
142 struct rb_node *node, *next;
143 struct callchain_node *child;
144 struct callchain_list *chain;
145 int new_depth_mask = depth_mask;
151 if (callchain_param.mode == CHAIN_GRAPH_REL)
152 new_total = self->children_hit;
154 new_total = total_samples;
156 remaining = new_total;
158 node = rb_first(&self->rb_root);
162 child = rb_entry(node, struct callchain_node, rb_node);
163 cumul = cumul_hits(child);
167 * The depth mask manages the output of pipes that show
168 * the depth. We don't want to keep the pipes of the current
169 * level for the last child of this depth.
170 * Except if we have remaining filtered hits. They will
171 * supersede the last child
173 next = rb_next(node);
174 if (!next && (callchain_param.mode != CHAIN_GRAPH_REL || !remaining))
175 new_depth_mask &= ~(1 << (depth - 1));
178 * But we keep the older depth mask for the line seperator
179 * to keep the level link until we reach the last child
181 ret += ipchain__fprintf_graph_line(fp, depth, depth_mask,
184 list_for_each_entry(chain, &child->val, list) {
185 if (chain->ip >= PERF_CONTEXT_MAX)
187 ret += ipchain__fprintf_graph(fp, chain, depth,
193 ret += __callchain__fprintf_graph(fp, child, new_total,
195 new_depth_mask | (1 << depth),
200 if (callchain_param.mode == CHAIN_GRAPH_REL &&
201 remaining && remaining != new_total) {
206 new_depth_mask &= ~(1 << (depth - 1));
208 ret += ipchain__fprintf_graph(fp, &rem_hits, depth,
209 new_depth_mask, 0, new_total,
210 remaining, left_margin);
218 callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
219 u64 total_samples, int left_margin)
221 struct callchain_list *chain;
222 bool printed = false;
226 list_for_each_entry(chain, &self->val, list) {
227 if (chain->ip >= PERF_CONTEXT_MAX)
230 if (!i++ && sort__first_dimension == SORT_SYM)
234 ret += callchain__fprintf_left_margin(fp, left_margin);
235 ret += fprintf(fp, "|\n");
236 ret += callchain__fprintf_left_margin(fp, left_margin);
237 ret += fprintf(fp, "---");
242 ret += callchain__fprintf_left_margin(fp, left_margin);
245 ret += fprintf(fp, " %s\n", chain->sym->name);
247 ret += fprintf(fp, " %p\n", (void *)(long)chain->ip);
250 ret += __callchain__fprintf_graph(fp, self, total_samples, 1, 1, left_margin);
256 callchain__fprintf_flat(FILE *fp, struct callchain_node *self,
259 struct callchain_list *chain;
265 ret += callchain__fprintf_flat(fp, self->parent, total_samples);
268 list_for_each_entry(chain, &self->val, list) {
269 if (chain->ip >= PERF_CONTEXT_MAX)
272 ret += fprintf(fp, " %s\n", chain->sym->name);
274 ret += fprintf(fp, " %p\n",
275 (void *)(long)chain->ip);
282 hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
283 u64 total_samples, int left_margin)
285 struct rb_node *rb_node;
286 struct callchain_node *chain;
289 rb_node = rb_first(&self->sorted_chain);
293 chain = rb_entry(rb_node, struct callchain_node, rb_node);
294 percent = chain->hit * 100.0 / total_samples;
295 switch (callchain_param.mode) {
297 ret += percent_color_fprintf(fp, " %6.2f%%\n",
299 ret += callchain__fprintf_flat(fp, chain, total_samples);
301 case CHAIN_GRAPH_ABS: /* Falldown */
302 case CHAIN_GRAPH_REL:
303 ret += callchain__fprintf_graph(fp, chain, total_samples,
309 ret += fprintf(fp, "\n");
310 rb_node = rb_next(rb_node);
316 static size_t hist_entry__fprintf(FILE *fp, struct hist_entry *self,
317 struct perf_session *session,
320 struct sort_entry *se;
323 if (exclude_other && !self->parent)
327 ret = percent_color_fprintf(fp,
328 field_sep ? "%.2f" : " %6.2f%%",
329 (self->count * 100.0) / total_samples);
331 ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count);
333 if (show_nr_samples) {
335 fprintf(fp, "%c%lld", *field_sep, self->count);
337 fprintf(fp, "%11lld", self->count);
340 list_for_each_entry(se, &hist_entry__sort_list, list) {
344 fprintf(fp, "%s", field_sep ?: " ");
345 ret += se->print(fp, self, se->width ? *se->width : 0);
348 ret += fprintf(fp, "\n");
350 if (session->use_callchain) {
353 if (sort__first_dimension == SORT_COMM) {
354 se = list_first_entry(&hist_entry__sort_list, typeof(*se),
356 left_margin = se->width ? *se->width : 0;
357 left_margin -= thread__comm_len(self->thread);
360 hist_entry_callchain__fprintf(fp, self, total_samples,
371 static void dso__calc_col_width(struct dso *self)
373 if (!col_width_list_str && !field_sep &&
374 (!dso_list || strlist__has_entry(dso_list, self->name))) {
375 unsigned int slen = strlen(self->name);
376 if (slen > dsos__col_width)
377 dsos__col_width = slen;
380 self->slen_calculated = 1;
383 static void thread__comm_adjust(struct thread *self)
385 char *comm = self->comm;
387 if (!col_width_list_str && !field_sep &&
388 (!comm_list || strlist__has_entry(comm_list, comm))) {
389 unsigned int slen = strlen(comm);
391 if (slen > comms__col_width) {
392 comms__col_width = slen;
393 threads__col_width = slen + 6;
398 static int thread__set_comm_adjust(struct thread *self, const char *comm)
400 int ret = thread__set_comm(self, comm);
405 thread__comm_adjust(self);
411 * collect histogram counts
414 static int perf_session__add_hist_entry(struct perf_session *self,
415 struct addr_location *al,
416 struct ip_callchain *chain, u64 count)
418 struct symbol **syms = NULL, *parent = NULL;
420 struct hist_entry *he;
422 if ((sort__has_parent || self->use_callchain) && chain)
423 syms = perf_session__resolve_callchain(self, al->thread,
425 he = __perf_session__add_hist_entry(self, al, parent, count, &hit);
432 if (self->use_callchain) {
434 callchain_init(&he->callchain);
435 append_chain(&he->callchain, chain, syms);
442 static size_t perf_session__fprintf_hist_entries(struct perf_session *self,
443 u64 total_samples, FILE *fp)
445 struct hist_entry *pos;
446 struct sort_entry *se;
450 char *col_width = col_width_list_str;
451 int raw_printing_style;
453 raw_printing_style = !strcmp(pretty_printing_style, "raw");
457 fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
460 fprintf(fp, "# Overhead");
461 if (show_nr_samples) {
463 fprintf(fp, "%cSamples", *field_sep);
465 fputs(" Samples ", fp);
467 list_for_each_entry(se, &hist_entry__sort_list, list) {
471 fprintf(fp, "%c%s", *field_sep, se->header);
474 width = strlen(se->header);
476 if (col_width_list_str) {
478 *se->width = atoi(col_width);
479 col_width = strchr(col_width, ',');
484 width = *se->width = max(*se->width, width);
486 fprintf(fp, " %*s", width, se->header);
493 fprintf(fp, "# ........");
495 fprintf(fp, " ..........");
496 list_for_each_entry(se, &hist_entry__sort_list, list) {
506 width = strlen(se->header);
507 for (i = 0; i < width; i++)
515 for (nd = rb_first(&self->hists); nd; nd = rb_next(nd)) {
516 pos = rb_entry(nd, struct hist_entry, rb_node);
517 ret += hist_entry__fprintf(fp, pos, self, total_samples);
520 if (sort_order == default_sort_order &&
521 parent_pattern == default_parent_pattern) {
523 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
528 free(rem_sq_bracket);
531 perf_read_values_display(fp, &show_threads_values,
537 static int validate_chain(struct ip_callchain *chain, event_t *event)
539 unsigned int chain_size;
541 chain_size = event->header.size;
542 chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
544 if (chain->nr*sizeof(u64) > chain_size)
550 static int process_sample_event(event_t *event, struct perf_session *session)
552 struct sample_data data;
554 struct addr_location al;
555 struct thread *thread;
557 memset(&data, 0, sizeof(data));
560 event__parse_sample(event, sample_type, &data);
562 dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
565 (void *)(long)data.ip,
566 (long long)data.period);
568 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
571 dump_printf("... chain: nr:%Lu\n", data.callchain->nr);
573 if (validate_chain(data.callchain, event) < 0) {
574 pr_debug("call-chain problem with event, "
580 for (i = 0; i < data.callchain->nr; i++)
581 dump_printf("..... %2d: %016Lx\n",
582 i, data.callchain->ips[i]);
586 thread = perf_session__findnew(session, data.pid);
587 if (thread == NULL) {
588 pr_debug("problem processing %d event, skipping it.\n",
593 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
595 if (comm_list && !strlist__has_entry(comm_list, thread->comm))
598 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
600 thread__find_addr_location(thread, session, cpumode,
601 MAP__FUNCTION, data.ip, &al, NULL);
603 * We have to do this here as we may have a dso with no symbol hit that
604 * has a name longer than the ones with symbols sampled.
606 if (al.map && !sort_dso.elide && !al.map->dso->slen_calculated)
607 dso__calc_col_width(al.map->dso);
610 (!al.map || !al.map->dso ||
611 !(strlist__has_entry(dso_list, al.map->dso->short_name) ||
612 (al.map->dso->short_name != al.map->dso->long_name &&
613 strlist__has_entry(dso_list, al.map->dso->long_name)))))
616 if (sym_list && al.sym && !strlist__has_entry(sym_list, al.sym->name))
619 if (perf_session__add_hist_entry(session, &al, data.callchain, data.period)) {
620 pr_debug("problem incrementing symbol count, skipping event\n");
624 event__stats.total += data.period;
629 static int process_comm_event(event_t *event, struct perf_session *session)
631 struct thread *thread = perf_session__findnew(session, event->comm.pid);
633 dump_printf(": %s:%d\n", event->comm.comm, event->comm.pid);
635 if (thread == NULL ||
636 thread__set_comm_adjust(thread, event->comm.comm)) {
637 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
644 static int process_read_event(event_t *event, struct perf_session *session __used)
646 struct perf_event_attr *attr;
648 attr = perf_header__find_attr(event->read.id, &session->header);
651 const char *name = attr ? __event_name(attr->type, attr->config)
653 perf_read_values_add_value(&show_threads_values,
654 event->read.pid, event->read.tid,
660 dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
661 attr ? __event_name(attr->type, attr->config) : "FAIL",
667 static int sample_type_check(u64 type, struct perf_session *session)
671 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
672 if (sort__has_parent) {
673 fprintf(stderr, "selected --sort parent, but no"
674 " callchain data. Did you call"
675 " perf record without -g?\n");
678 if (session->use_callchain) {
679 fprintf(stderr, "selected -g but no callchain data."
680 " Did you call perf record without"
684 } else if (callchain_param.mode != CHAIN_NONE && !session->use_callchain) {
685 session->use_callchain = true;
686 if (register_callchain_param(&callchain_param) < 0) {
687 fprintf(stderr, "Can't register callchain"
696 static struct perf_event_ops event_ops = {
697 .process_sample_event = process_sample_event,
698 .process_mmap_event = event__process_mmap,
699 .process_comm_event = process_comm_event,
700 .process_exit_event = event__process_task,
701 .process_fork_event = event__process_task,
702 .process_lost_event = event__process_lost,
703 .process_read_event = process_read_event,
704 .sample_type_check = sample_type_check,
708 static int __cmd_report(void)
711 struct perf_session *session;
713 session = perf_session__new(input_name, O_RDONLY, force, &symbol_conf);
717 session->use_callchain = use_callchain;
720 perf_read_values_init(&show_threads_values);
722 ret = perf_session__process_events(session, &event_ops);
727 event__print_totals();
732 perf_session__fprintf(session, stdout);
735 dsos__fprintf(stdout);
737 perf_session__collapse_resort(session);
738 perf_session__output_resort(session, event__stats.total);
739 perf_session__fprintf_hist_entries(session, event__stats.total, stdout);
742 perf_read_values_destroy(&show_threads_values);
744 perf_session__delete(session);
749 parse_callchain_opt(const struct option *opt __used, const char *arg,
755 use_callchain = true;
760 tok = strtok((char *)arg, ",");
764 /* get the output mode */
765 if (!strncmp(tok, "graph", strlen(arg)))
766 callchain_param.mode = CHAIN_GRAPH_ABS;
768 else if (!strncmp(tok, "flat", strlen(arg)))
769 callchain_param.mode = CHAIN_FLAT;
771 else if (!strncmp(tok, "fractal", strlen(arg)))
772 callchain_param.mode = CHAIN_GRAPH_REL;
774 else if (!strncmp(tok, "none", strlen(arg))) {
775 callchain_param.mode = CHAIN_NONE;
776 use_callchain = true;
784 /* get the min percentage */
785 tok = strtok(NULL, ",");
789 callchain_param.min_percent = strtod(tok, &endptr);
794 if (register_callchain_param(&callchain_param) < 0) {
795 fprintf(stderr, "Can't register callchain params\n");
801 //static const char * const report_usage[] = {
802 const char * const report_usage[] = {
803 "perf report [<options>] <command>",
807 static const struct option options[] = {
808 OPT_STRING('i', "input", &input_name, "file",
810 OPT_BOOLEAN('v', "verbose", &verbose,
811 "be more verbose (show symbol address, etc)"),
812 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
813 "dump raw trace in ASCII"),
814 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
815 "file", "vmlinux pathname"),
816 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
817 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
818 "load module symbols - WARNING: use only with -k and LIVE kernel"),
819 OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
820 "Show a column with the number of samples"),
821 OPT_BOOLEAN('T', "threads", &show_threads,
822 "Show per-thread event counters"),
823 OPT_STRING(0, "pretty", &pretty_printing_style, "key",
824 "pretty printing style key: normal raw"),
825 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
826 "sort by key(s): pid, comm, dso, symbol, parent"),
827 OPT_BOOLEAN('P', "full-paths", &event_ops.full_paths,
828 "Don't shorten the pathnames taking into account the cwd"),
829 OPT_STRING('p', "parent", &parent_pattern, "regex",
830 "regex filter to identify parent, see: '--sort parent'"),
831 OPT_BOOLEAN('x', "exclude-other", &exclude_other,
832 "Only display entries with parent-match"),
833 OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
834 "Display callchains using output_type and min percent threshold. "
835 "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
836 OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
837 "only consider symbols in these dsos"),
838 OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
839 "only consider symbols in these comms"),
840 OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
841 "only consider these symbols"),
842 OPT_STRING('w', "column-widths", &col_width_list_str,
844 "don't try to adjust column width, use these fixed values"),
845 OPT_STRING('t', "field-separator", &field_sep, "separator",
846 "separator for columns, no spaces will be added between "
847 "columns '.' is reserved."),
851 static void setup_sorting(void)
853 char *tmp, *tok, *str = strdup(sort_order);
855 for (tok = strtok_r(str, ", ", &tmp);
856 tok; tok = strtok_r(NULL, ", ", &tmp)) {
857 if (sort_dimension__add(tok) < 0) {
858 error("Unknown --sort key: `%s'", tok);
859 usage_with_options(report_usage, options);
866 static void setup_list(struct strlist **list, const char *list_str,
867 struct sort_entry *se, const char *list_name,
871 *list = strlist__new(true, list_str);
873 fprintf(stderr, "problems parsing %s list\n",
877 if (strlist__nr_entries(*list) == 1) {
878 fprintf(fp, "# %s: %s\n", list_name,
879 strlist__entry(*list, 0)->s);
885 int cmd_report(int argc, const char **argv, const char *prefix __used)
887 if (symbol__init(&symbol_conf) < 0)
890 argc = parse_options(argc, argv, options, report_usage, 0);
894 if (parent_pattern != default_parent_pattern) {
895 sort_dimension__add("parent");
896 sort_parent.elide = 1;
901 * Any (unrecognized) arguments left?
904 usage_with_options(report_usage, options);
908 setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
909 setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
910 setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
912 if (field_sep && *field_sep == '.') {
913 fputs("'.' is the only non valid --field-separator argument\n",
918 return __cmd_report();