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 ip_callchain *chain,
418 struct symbol **parent)
420 u8 cpumode = PERF_RECORD_MISC_USER;
421 struct symbol **syms = NULL;
425 syms = calloc(chain->nr, sizeof(*syms));
427 fprintf(stderr, "Can't allocate memory for symbols\n");
432 for (i = 0; i < chain->nr; i++) {
433 u64 ip = chain->ips[i];
434 struct addr_location al;
436 if (ip >= PERF_CONTEXT_MAX) {
438 case PERF_CONTEXT_HV:
439 cpumode = PERF_RECORD_MISC_HYPERVISOR; break;
440 case PERF_CONTEXT_KERNEL:
441 cpumode = PERF_RECORD_MISC_KERNEL; break;
442 case PERF_CONTEXT_USER:
443 cpumode = PERF_RECORD_MISC_USER; break;
450 thread__find_addr_location(thread, cpumode, MAP__FUNCTION,
452 if (al.sym != NULL) {
453 if (sort__has_parent && !*parent &&
466 * collect histogram counts
469 static int hist_entry__add(struct addr_location *al,
470 struct ip_callchain *chain, u64 count)
472 struct symbol **syms = NULL, *parent = NULL;
474 struct hist_entry *he;
476 if ((sort__has_parent || callchain) && chain)
477 syms = resolve_callchain(al->thread, chain, &parent);
479 he = __hist_entry__add(al, parent, count, &hit);
488 callchain_init(&he->callchain);
489 append_chain(&he->callchain, chain, syms);
496 static size_t output__fprintf(FILE *fp, u64 total_samples)
498 struct hist_entry *pos;
499 struct sort_entry *se;
503 char *col_width = col_width_list_str;
504 int raw_printing_style;
506 raw_printing_style = !strcmp(pretty_printing_style, "raw");
510 fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
513 fprintf(fp, "# Overhead");
514 if (show_nr_samples) {
516 fprintf(fp, "%cSamples", *field_sep);
518 fputs(" Samples ", fp);
520 list_for_each_entry(se, &hist_entry__sort_list, list) {
524 fprintf(fp, "%c%s", *field_sep, se->header);
527 width = strlen(se->header);
529 if (col_width_list_str) {
531 *se->width = atoi(col_width);
532 col_width = strchr(col_width, ',');
537 width = *se->width = max(*se->width, width);
539 fprintf(fp, " %*s", width, se->header);
546 fprintf(fp, "# ........");
548 fprintf(fp, " ..........");
549 list_for_each_entry(se, &hist_entry__sort_list, list) {
559 width = strlen(se->header);
560 for (i = 0; i < width; i++)
568 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
569 pos = rb_entry(nd, struct hist_entry, rb_node);
570 ret += hist_entry__fprintf(fp, pos, total_samples);
573 if (sort_order == default_sort_order &&
574 parent_pattern == default_parent_pattern) {
576 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
581 free(rem_sq_bracket);
584 perf_read_values_display(fp, &show_threads_values,
590 static int validate_chain(struct ip_callchain *chain, event_t *event)
592 unsigned int chain_size;
594 chain_size = event->header.size;
595 chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
597 if (chain->nr*sizeof(u64) > chain_size)
603 static int process_sample_event(event_t *event, struct perf_session *session __used)
605 struct sample_data data;
607 struct addr_location al;
608 struct thread *thread;
610 memset(&data, 0, sizeof(data));
613 event__parse_sample(event, sample_type, &data);
615 dump_printf("(IP, %d): %d/%d: %p period: %Ld\n",
618 (void *)(long)data.ip,
619 (long long)data.period);
621 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
624 dump_printf("... chain: nr:%Lu\n", data.callchain->nr);
626 if (validate_chain(data.callchain, event) < 0) {
627 pr_debug("call-chain problem with event, "
633 for (i = 0; i < data.callchain->nr; i++)
634 dump_printf("..... %2d: %016Lx\n",
635 i, data.callchain->ips[i]);
639 thread = threads__findnew(data.pid);
640 if (thread == NULL) {
641 pr_debug("problem processing %d event, skipping it.\n",
646 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
648 if (comm_list && !strlist__has_entry(comm_list, thread->comm))
651 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
653 thread__find_addr_location(thread, cpumode,
654 MAP__FUNCTION, data.ip, &al, NULL);
656 * We have to do this here as we may have a dso with no symbol hit that
657 * has a name longer than the ones with symbols sampled.
659 if (al.map && !sort_dso.elide && !al.map->dso->slen_calculated)
660 dso__calc_col_width(al.map->dso);
663 (!al.map || !al.map->dso ||
664 !(strlist__has_entry(dso_list, al.map->dso->short_name) ||
665 (al.map->dso->short_name != al.map->dso->long_name &&
666 strlist__has_entry(dso_list, al.map->dso->long_name)))))
669 if (sym_list && al.sym && !strlist__has_entry(sym_list, al.sym->name))
672 if (hist_entry__add(&al, data.callchain, data.period)) {
673 pr_debug("problem incrementing symbol count, skipping event\n");
677 event__stats.total += data.period;
682 static int process_comm_event(event_t *event, struct perf_session *session __used)
684 struct thread *thread = threads__findnew(event->comm.pid);
686 dump_printf(": %s:%d\n", event->comm.comm, event->comm.pid);
688 if (thread == NULL ||
689 thread__set_comm_adjust(thread, event->comm.comm)) {
690 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
697 static int process_read_event(event_t *event, struct perf_session *session __used)
699 struct perf_event_attr *attr;
701 attr = perf_header__find_attr(event->read.id, &session->header);
704 const char *name = attr ? __event_name(attr->type, attr->config)
706 perf_read_values_add_value(&show_threads_values,
707 event->read.pid, event->read.tid,
713 dump_printf(": %d %d %s %Lu\n", event->read.pid, event->read.tid,
714 attr ? __event_name(attr->type, attr->config) : "FAIL",
720 static int sample_type_check(u64 type)
724 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
725 if (sort__has_parent) {
726 fprintf(stderr, "selected --sort parent, but no"
727 " callchain data. Did you call"
728 " perf record without -g?\n");
732 fprintf(stderr, "selected -g but no callchain data."
733 " Did you call perf record without"
737 } else if (callchain_param.mode != CHAIN_NONE && !callchain) {
739 if (register_callchain_param(&callchain_param) < 0) {
740 fprintf(stderr, "Can't register callchain"
749 static struct perf_event_ops event_ops = {
750 .process_sample_event = process_sample_event,
751 .process_mmap_event = event__process_mmap,
752 .process_comm_event = process_comm_event,
753 .process_exit_event = event__process_task,
754 .process_fork_event = event__process_task,
755 .process_lost_event = event__process_lost,
756 .process_read_event = process_read_event,
757 .sample_type_check = sample_type_check,
761 static int __cmd_report(void)
764 struct perf_session *session;
766 session = perf_session__new(input_name, O_RDONLY, force);
771 perf_read_values_init(&show_threads_values);
773 ret = perf_session__process_events(session, &event_ops);
778 event__print_totals();
783 threads__fprintf(stdout);
786 dsos__fprintf(stdout);
789 output__resort(event__stats.total);
790 output__fprintf(stdout, event__stats.total);
793 perf_read_values_destroy(&show_threads_values);
795 perf_session__delete(session);
800 parse_callchain_opt(const struct option *opt __used, const char *arg,
811 tok = strtok((char *)arg, ",");
815 /* get the output mode */
816 if (!strncmp(tok, "graph", strlen(arg)))
817 callchain_param.mode = CHAIN_GRAPH_ABS;
819 else if (!strncmp(tok, "flat", strlen(arg)))
820 callchain_param.mode = CHAIN_FLAT;
822 else if (!strncmp(tok, "fractal", strlen(arg)))
823 callchain_param.mode = CHAIN_GRAPH_REL;
825 else if (!strncmp(tok, "none", strlen(arg))) {
826 callchain_param.mode = CHAIN_NONE;
835 /* get the min percentage */
836 tok = strtok(NULL, ",");
840 callchain_param.min_percent = strtod(tok, &endptr);
845 if (register_callchain_param(&callchain_param) < 0) {
846 fprintf(stderr, "Can't register callchain params\n");
852 //static const char * const report_usage[] = {
853 const char * const report_usage[] = {
854 "perf report [<options>] <command>",
858 static const struct option options[] = {
859 OPT_STRING('i', "input", &input_name, "file",
861 OPT_BOOLEAN('v', "verbose", &verbose,
862 "be more verbose (show symbol address, etc)"),
863 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
864 "dump raw trace in ASCII"),
865 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
866 "file", "vmlinux pathname"),
867 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
868 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
869 "load module symbols - WARNING: use only with -k and LIVE kernel"),
870 OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
871 "Show a column with the number of samples"),
872 OPT_BOOLEAN('T', "threads", &show_threads,
873 "Show per-thread event counters"),
874 OPT_STRING(0, "pretty", &pretty_printing_style, "key",
875 "pretty printing style key: normal raw"),
876 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
877 "sort by key(s): pid, comm, dso, symbol, parent"),
878 OPT_BOOLEAN('P', "full-paths", &event_ops.full_paths,
879 "Don't shorten the pathnames taking into account the cwd"),
880 OPT_STRING('p', "parent", &parent_pattern, "regex",
881 "regex filter to identify parent, see: '--sort parent'"),
882 OPT_BOOLEAN('x', "exclude-other", &exclude_other,
883 "Only display entries with parent-match"),
884 OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
885 "Display callchains using output_type and min percent threshold. "
886 "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
887 OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
888 "only consider symbols in these dsos"),
889 OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
890 "only consider symbols in these comms"),
891 OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
892 "only consider these symbols"),
893 OPT_STRING('w', "column-widths", &col_width_list_str,
895 "don't try to adjust column width, use these fixed values"),
896 OPT_STRING('t', "field-separator", &field_sep, "separator",
897 "separator for columns, no spaces will be added between "
898 "columns '.' is reserved."),
902 static void setup_sorting(void)
904 char *tmp, *tok, *str = strdup(sort_order);
906 for (tok = strtok_r(str, ", ", &tmp);
907 tok; tok = strtok_r(NULL, ", ", &tmp)) {
908 if (sort_dimension__add(tok) < 0) {
909 error("Unknown --sort key: `%s'", tok);
910 usage_with_options(report_usage, options);
917 static void setup_list(struct strlist **list, const char *list_str,
918 struct sort_entry *se, const char *list_name,
922 *list = strlist__new(true, list_str);
924 fprintf(stderr, "problems parsing %s list\n",
928 if (strlist__nr_entries(*list) == 1) {
929 fprintf(fp, "# %s: %s\n", list_name,
930 strlist__entry(*list, 0)->s);
936 int cmd_report(int argc, const char **argv, const char *prefix __used)
938 if (symbol__init(&symbol_conf) < 0)
941 argc = parse_options(argc, argv, options, report_usage, 0);
945 if (parent_pattern != default_parent_pattern) {
946 sort_dimension__add("parent");
947 sort_parent.elide = 1;
952 * Any (unrecognized) arguments left?
955 usage_with_options(report_usage, options);
959 setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
960 setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
961 setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
963 if (field_sep && *field_sep == '.') {
964 fputs("'.' is the only non valid --field-separator argument\n",
969 return __cmd_report();