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"
26 #include "util/parse-options.h"
27 #include "util/parse-events.h"
29 #include "util/data_map.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 full_paths;
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";
58 static struct perf_header *header;
60 static u64 sample_type;
62 static size_t ipchain__fprintf_graph_line(FILE *fp, int depth, int depth_mask)
67 ret += fprintf(fp, "%s", " ");
69 for (i = 0; i < depth; i++)
70 if (depth_mask & (1 << i))
71 ret += fprintf(fp, "| ");
73 ret += fprintf(fp, " ");
75 ret += fprintf(fp, "\n");
80 ipchain__fprintf_graph(FILE *fp, struct callchain_list *chain, int depth,
81 int depth_mask, int count, u64 total_samples,
87 ret += fprintf(fp, "%s", " ");
88 for (i = 0; i < depth; i++) {
89 if (depth_mask & (1 << i))
90 ret += fprintf(fp, "|");
92 ret += fprintf(fp, " ");
93 if (!count && i == depth - 1) {
96 percent = hits * 100.0 / total_samples;
97 ret += percent_color_fprintf(fp, "--%2.2f%%-- ", percent);
99 ret += fprintf(fp, "%s", " ");
102 ret += fprintf(fp, "%s\n", chain->sym->name);
104 ret += fprintf(fp, "%p\n", (void *)(long)chain->ip);
109 static struct symbol *rem_sq_bracket;
110 static struct callchain_list rem_hits;
112 static void init_rem_hits(void)
114 rem_sq_bracket = malloc(sizeof(*rem_sq_bracket) + 6);
115 if (!rem_sq_bracket) {
116 fprintf(stderr, "Not enough memory to display remaining hits\n");
120 strcpy(rem_sq_bracket->name, "[...]");
121 rem_hits.sym = rem_sq_bracket;
125 callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
126 u64 total_samples, int depth, int depth_mask)
128 struct rb_node *node, *next;
129 struct callchain_node *child;
130 struct callchain_list *chain;
131 int new_depth_mask = depth_mask;
137 if (callchain_param.mode == CHAIN_GRAPH_REL)
138 new_total = self->children_hit;
140 new_total = total_samples;
142 remaining = new_total;
144 node = rb_first(&self->rb_root);
148 child = rb_entry(node, struct callchain_node, rb_node);
149 cumul = cumul_hits(child);
153 * The depth mask manages the output of pipes that show
154 * the depth. We don't want to keep the pipes of the current
155 * level for the last child of this depth.
156 * Except if we have remaining filtered hits. They will
157 * supersede the last child
159 next = rb_next(node);
160 if (!next && (callchain_param.mode != CHAIN_GRAPH_REL || !remaining))
161 new_depth_mask &= ~(1 << (depth - 1));
164 * But we keep the older depth mask for the line seperator
165 * to keep the level link until we reach the last child
167 ret += ipchain__fprintf_graph_line(fp, depth, depth_mask);
169 list_for_each_entry(chain, &child->val, list) {
170 if (chain->ip >= PERF_CONTEXT_MAX)
172 ret += ipchain__fprintf_graph(fp, chain, depth,
177 ret += callchain__fprintf_graph(fp, child, new_total,
179 new_depth_mask | (1 << depth));
183 if (callchain_param.mode == CHAIN_GRAPH_REL &&
184 remaining && remaining != new_total) {
189 new_depth_mask &= ~(1 << (depth - 1));
191 ret += ipchain__fprintf_graph(fp, &rem_hits, depth,
192 new_depth_mask, 0, new_total,
200 callchain__fprintf_flat(FILE *fp, struct callchain_node *self,
203 struct callchain_list *chain;
209 ret += callchain__fprintf_flat(fp, self->parent, total_samples);
212 list_for_each_entry(chain, &self->val, list) {
213 if (chain->ip >= PERF_CONTEXT_MAX)
216 ret += fprintf(fp, " %s\n", chain->sym->name);
218 ret += fprintf(fp, " %p\n",
219 (void *)(long)chain->ip);
226 hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
229 struct rb_node *rb_node;
230 struct callchain_node *chain;
233 rb_node = rb_first(&self->sorted_chain);
237 chain = rb_entry(rb_node, struct callchain_node, rb_node);
238 percent = chain->hit * 100.0 / total_samples;
239 switch (callchain_param.mode) {
241 ret += percent_color_fprintf(fp, " %6.2f%%\n",
243 ret += callchain__fprintf_flat(fp, chain, total_samples);
245 case CHAIN_GRAPH_ABS: /* Falldown */
246 case CHAIN_GRAPH_REL:
247 ret += callchain__fprintf_graph(fp, chain,
248 total_samples, 1, 1);
253 ret += fprintf(fp, "\n");
254 rb_node = rb_next(rb_node);
261 hist_entry__fprintf(FILE *fp, struct hist_entry *self, u64 total_samples)
263 struct sort_entry *se;
266 if (exclude_other && !self->parent)
270 ret = percent_color_fprintf(fp,
271 field_sep ? "%.2f" : " %6.2f%%",
272 (self->count * 100.0) / total_samples);
274 ret = fprintf(fp, field_sep ? "%lld" : "%12lld ", self->count);
276 if (show_nr_samples) {
278 fprintf(fp, "%c%lld", *field_sep, self->count);
280 fprintf(fp, "%11lld", self->count);
283 list_for_each_entry(se, &hist_entry__sort_list, list) {
287 fprintf(fp, "%s", field_sep ?: " ");
288 ret += se->print(fp, self, se->width ? *se->width : 0);
291 ret += fprintf(fp, "\n");
294 hist_entry_callchain__fprintf(fp, self, total_samples);
303 static void dso__calc_col_width(struct dso *self)
305 if (!col_width_list_str && !field_sep &&
306 (!dso_list || strlist__has_entry(dso_list, self->name))) {
307 unsigned int slen = strlen(self->name);
308 if (slen > dsos__col_width)
309 dsos__col_width = slen;
312 self->slen_calculated = 1;
315 static void thread__comm_adjust(struct thread *self)
317 char *comm = self->comm;
319 if (!col_width_list_str && !field_sep &&
320 (!comm_list || strlist__has_entry(comm_list, comm))) {
321 unsigned int slen = strlen(comm);
323 if (slen > comms__col_width) {
324 comms__col_width = slen;
325 threads__col_width = slen + 6;
330 static int thread__set_comm_adjust(struct thread *self, const char *comm)
332 int ret = thread__set_comm(self, comm);
337 thread__comm_adjust(self);
343 static struct symbol *
344 resolve_symbol(struct thread *thread, struct map **mapp, u64 *ipp)
346 struct map *map = mapp ? *mapp : NULL;
355 map = thread__find_map(thread, ip);
358 * We have to do this here as we may have a dso
359 * with no symbol hit that has a name longer than
360 * the ones with symbols sampled.
362 if (!sort_dso.elide && !map->dso->slen_calculated)
363 dso__calc_col_width(map->dso);
368 ip = map->map_ip(map, ip);
371 * If this is outside of all known maps,
372 * and is a negative address, try to look it
373 * up in the kernel dso, as it might be a
374 * vsyscall or vdso (which executes in user-mode).
376 * XXX This is nasty, we should have a symbol list in
377 * the "[vdso]" dso, but for now lets use the old
378 * trick of looking in the whole kernel symbol list.
380 if ((long long)ip < 0)
381 return kernel_maps__find_symbol(ip, mapp);
383 dump_printf(" ...... dso: %s\n",
384 map ? map->dso->long_name : "<not found>");
385 dump_printf(" ...... map: %Lx -> %Lx\n", *ipp, ip);
388 return map ? map->dso->find_symbol(map->dso, ip) : NULL;
391 static int call__match(struct symbol *sym)
393 if (sym->name && !regexec(&parent_regex, sym->name, 0, NULL, 0))
399 static struct symbol **resolve_callchain(struct thread *thread, struct map *map,
400 struct ip_callchain *chain,
401 struct symbol **parent)
403 u64 context = PERF_CONTEXT_MAX;
404 struct symbol **syms = NULL;
408 syms = calloc(chain->nr, sizeof(*syms));
410 fprintf(stderr, "Can't allocate memory for symbols\n");
415 for (i = 0; i < chain->nr; i++) {
416 u64 ip = chain->ips[i];
417 struct symbol *sym = NULL;
419 if (ip >= PERF_CONTEXT_MAX) {
425 case PERF_CONTEXT_HV:
427 case PERF_CONTEXT_KERNEL:
428 sym = kernel_maps__find_symbol(ip, &map);
431 sym = resolve_symbol(thread, &map, &ip);
436 if (sort__has_parent && !*parent && call__match(sym))
448 * collect histogram counts
452 hist_entry__add(struct thread *thread, struct map *map,
453 struct symbol *sym, u64 ip, struct ip_callchain *chain,
454 char level, u64 count)
456 struct symbol **syms = NULL, *parent = NULL;
458 struct hist_entry *he;
460 if ((sort__has_parent || callchain) && chain)
461 syms = resolve_callchain(thread, map, chain, &parent);
463 he = __hist_entry__add(thread, map, sym, parent,
464 ip, count, level, &hit);
473 callchain_init(&he->callchain);
474 append_chain(&he->callchain, chain, syms);
481 static size_t output__fprintf(FILE *fp, u64 total_samples)
483 struct hist_entry *pos;
484 struct sort_entry *se;
488 char *col_width = col_width_list_str;
489 int raw_printing_style;
491 raw_printing_style = !strcmp(pretty_printing_style, "raw");
495 fprintf(fp, "# Samples: %Ld\n", (u64)total_samples);
498 fprintf(fp, "# Overhead");
499 if (show_nr_samples) {
501 fprintf(fp, "%cSamples", *field_sep);
503 fputs(" Samples ", fp);
505 list_for_each_entry(se, &hist_entry__sort_list, list) {
509 fprintf(fp, "%c%s", *field_sep, se->header);
512 width = strlen(se->header);
514 if (col_width_list_str) {
516 *se->width = atoi(col_width);
517 col_width = strchr(col_width, ',');
522 width = *se->width = max(*se->width, width);
524 fprintf(fp, " %*s", width, se->header);
531 fprintf(fp, "# ........");
533 fprintf(fp, " ..........");
534 list_for_each_entry(se, &hist_entry__sort_list, list) {
544 width = strlen(se->header);
545 for (i = 0; i < width; i++)
553 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
554 pos = rb_entry(nd, struct hist_entry, rb_node);
555 ret += hist_entry__fprintf(fp, pos, total_samples);
558 if (sort_order == default_sort_order &&
559 parent_pattern == default_parent_pattern) {
561 fprintf(fp, "# (For a higher level overview, try: perf report --sort comm,dso)\n");
566 free(rem_sq_bracket);
569 perf_read_values_display(fp, &show_threads_values,
575 static int validate_chain(struct ip_callchain *chain, event_t *event)
577 unsigned int chain_size;
579 chain_size = event->header.size;
580 chain_size -= (unsigned long)&event->ip.__more_data - (unsigned long)event;
582 if (chain->nr*sizeof(u64) > chain_size)
589 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
592 struct symbol *sym = NULL;
593 u64 ip = event->ip.ip;
595 struct map *map = NULL;
596 void *more_data = event->ip.__more_data;
597 struct ip_callchain *chain = NULL;
599 struct thread *thread = threads__findnew(event->ip.pid);
601 if (sample_type & PERF_SAMPLE_PERIOD) {
602 period = *(u64 *)more_data;
603 more_data += sizeof(u64);
606 dump_printf("%p [%p]: PERF_RECORD_SAMPLE (IP, %d): %d/%d: %p period: %Ld\n",
607 (void *)(offset + head),
608 (void *)(long)(event->header.size),
610 event->ip.pid, event->ip.tid,
614 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
617 chain = (void *)more_data;
619 dump_printf("... chain: nr:%Lu\n", chain->nr);
621 if (validate_chain(chain, event) < 0) {
622 eprintf("call-chain problem with event, skipping it.\n");
627 for (i = 0; i < chain->nr; i++)
628 dump_printf("..... %2d: %016Lx\n", i, chain->ips[i]);
632 if (thread == NULL) {
633 eprintf("problem processing %d event, skipping it.\n",
638 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
640 if (comm_list && !strlist__has_entry(comm_list, thread->comm))
643 cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
645 if (cpumode == PERF_RECORD_MISC_KERNEL) {
647 sym = kernel_maps__find_symbol(ip, &map);
648 dump_printf(" ...... dso: %s\n",
649 map ? map->dso->long_name : "<not found>");
650 } else if (cpumode == PERF_RECORD_MISC_USER) {
652 sym = resolve_symbol(thread, &map, &ip);
656 dump_printf(" ...... dso: [hypervisor]\n");
660 (!map || !map->dso ||
661 !(strlist__has_entry(dso_list, map->dso->short_name) ||
662 (map->dso->short_name != map->dso->long_name &&
663 strlist__has_entry(dso_list, map->dso->long_name)))))
666 if (sym_list && sym && !strlist__has_entry(sym_list, sym->name))
669 if (hist_entry__add(thread, map, sym, ip,
670 chain, level, period)) {
671 eprintf("problem incrementing symbol count, skipping event\n");
681 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
683 struct map *map = map__new(&event->mmap, cwd, cwdlen);
684 struct thread *thread = threads__findnew(event->mmap.pid);
686 dump_printf("%p [%p]: PERF_RECORD_MMAP %d/%d: [%p(%p) @ %p]: %s\n",
687 (void *)(offset + head),
688 (void *)(long)(event->header.size),
691 (void *)(long)event->mmap.start,
692 (void *)(long)event->mmap.len,
693 (void *)(long)event->mmap.pgoff,
694 event->mmap.filename);
696 if (thread == NULL || map == NULL) {
697 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
701 thread__insert_map(thread, map);
708 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
710 struct thread *thread = threads__findnew(event->comm.pid);
712 dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
713 (void *)(offset + head),
714 (void *)(long)(event->header.size),
715 event->comm.comm, event->comm.pid);
717 if (thread == NULL ||
718 thread__set_comm_adjust(thread, event->comm.comm)) {
719 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
728 process_task_event(event_t *event, unsigned long offset, unsigned long head)
730 struct thread *thread = threads__findnew(event->fork.pid);
731 struct thread *parent = threads__findnew(event->fork.ppid);
733 dump_printf("%p [%p]: PERF_RECORD_%s: (%d:%d):(%d:%d)\n",
734 (void *)(offset + head),
735 (void *)(long)(event->header.size),
736 event->header.type == PERF_RECORD_FORK ? "FORK" : "EXIT",
737 event->fork.pid, event->fork.tid,
738 event->fork.ppid, event->fork.ptid);
741 * A thread clone will have the same PID for both
744 if (thread == parent)
747 if (event->header.type == PERF_RECORD_EXIT)
750 if (!thread || !parent || thread__fork(thread, parent)) {
751 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
760 process_lost_event(event_t *event, unsigned long offset, unsigned long head)
762 dump_printf("%p [%p]: PERF_RECORD_LOST: id:%Ld: lost:%Ld\n",
763 (void *)(offset + head),
764 (void *)(long)(event->header.size),
768 total_lost += event->lost.lost;
774 process_read_event(event_t *event, unsigned long offset, unsigned long head)
776 struct perf_event_attr *attr;
778 attr = perf_header__find_attr(event->read.id, header);
781 const char *name = attr ? __event_name(attr->type, attr->config)
783 perf_read_values_add_value(&show_threads_values,
784 event->read.pid, event->read.tid,
790 dump_printf("%p [%p]: PERF_RECORD_READ: %d %d %s %Lu\n",
791 (void *)(offset + head),
792 (void *)(long)(event->header.size),
795 attr ? __event_name(attr->type, attr->config)
802 static int sample_type_check(u64 type)
806 if (!(sample_type & PERF_SAMPLE_CALLCHAIN)) {
807 if (sort__has_parent) {
808 fprintf(stderr, "selected --sort parent, but no"
809 " callchain data. Did you call"
810 " perf record without -g?\n");
814 fprintf(stderr, "selected -g but no callchain data."
815 " Did you call perf record without"
819 } else if (callchain_param.mode != CHAIN_NONE && !callchain) {
821 if (register_callchain_param(&callchain_param) < 0) {
822 fprintf(stderr, "Can't register callchain"
831 static struct perf_file_handler file_handler = {
832 .process_sample_event = process_sample_event,
833 .process_mmap_event = process_mmap_event,
834 .process_comm_event = process_comm_event,
835 .process_exit_event = process_task_event,
836 .process_fork_event = process_task_event,
837 .process_lost_event = process_lost_event,
838 .process_read_event = process_read_event,
839 .sample_type_check = sample_type_check,
843 static int __cmd_report(void)
848 idle = register_idle_thread();
849 thread__comm_adjust(idle);
852 perf_read_values_init(&show_threads_values);
854 register_perf_file_handler(&file_handler);
856 ret = mmap_dispatch_perf_file(&header, input_name, force, full_paths,
861 dump_printf(" IP events: %10ld\n", total);
862 dump_printf(" mmap events: %10ld\n", total_mmap);
863 dump_printf(" comm events: %10ld\n", total_comm);
864 dump_printf(" fork events: %10ld\n", total_fork);
865 dump_printf(" lost events: %10ld\n", total_lost);
866 dump_printf(" unknown events: %10ld\n", file_handler.total_unknown);
872 threads__fprintf(stdout);
875 dsos__fprintf(stdout);
878 output__resort(total);
879 output__fprintf(stdout, total);
882 perf_read_values_destroy(&show_threads_values);
888 parse_callchain_opt(const struct option *opt __used, const char *arg,
899 tok = strtok((char *)arg, ",");
903 /* get the output mode */
904 if (!strncmp(tok, "graph", strlen(arg)))
905 callchain_param.mode = CHAIN_GRAPH_ABS;
907 else if (!strncmp(tok, "flat", strlen(arg)))
908 callchain_param.mode = CHAIN_FLAT;
910 else if (!strncmp(tok, "fractal", strlen(arg)))
911 callchain_param.mode = CHAIN_GRAPH_REL;
913 else if (!strncmp(tok, "none", strlen(arg))) {
914 callchain_param.mode = CHAIN_NONE;
923 /* get the min percentage */
924 tok = strtok(NULL, ",");
928 callchain_param.min_percent = strtod(tok, &endptr);
933 if (register_callchain_param(&callchain_param) < 0) {
934 fprintf(stderr, "Can't register callchain params\n");
940 //static const char * const report_usage[] = {
941 const char * const report_usage[] = {
942 "perf report [<options>] <command>",
946 static const struct option options[] = {
947 OPT_STRING('i', "input", &input_name, "file",
949 OPT_BOOLEAN('v', "verbose", &verbose,
950 "be more verbose (show symbol address, etc)"),
951 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
952 "dump raw trace in ASCII"),
953 OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
954 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
955 OPT_BOOLEAN('m', "modules", &modules,
956 "load module symbols - WARNING: use only with -k and LIVE kernel"),
957 OPT_BOOLEAN('n', "show-nr-samples", &show_nr_samples,
958 "Show a column with the number of samples"),
959 OPT_BOOLEAN('T', "threads", &show_threads,
960 "Show per-thread event counters"),
961 OPT_STRING(0, "pretty", &pretty_printing_style, "key",
962 "pretty printing style key: normal raw"),
963 OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
964 "sort by key(s): pid, comm, dso, symbol, parent"),
965 OPT_BOOLEAN('P', "full-paths", &full_paths,
966 "Don't shorten the pathnames taking into account the cwd"),
967 OPT_STRING('p', "parent", &parent_pattern, "regex",
968 "regex filter to identify parent, see: '--sort parent'"),
969 OPT_BOOLEAN('x', "exclude-other", &exclude_other,
970 "Only display entries with parent-match"),
971 OPT_CALLBACK_DEFAULT('g', "call-graph", NULL, "output_type,min_percent",
972 "Display callchains using output_type and min percent threshold. "
973 "Default: fractal,0.5", &parse_callchain_opt, callchain_default_opt),
974 OPT_STRING('d', "dsos", &dso_list_str, "dso[,dso...]",
975 "only consider symbols in these dsos"),
976 OPT_STRING('C', "comms", &comm_list_str, "comm[,comm...]",
977 "only consider symbols in these comms"),
978 OPT_STRING('S', "symbols", &sym_list_str, "symbol[,symbol...]",
979 "only consider these symbols"),
980 OPT_STRING('w', "column-widths", &col_width_list_str,
982 "don't try to adjust column width, use these fixed values"),
983 OPT_STRING('t', "field-separator", &field_sep, "separator",
984 "separator for columns, no spaces will be added between "
985 "columns '.' is reserved."),
989 static void setup_sorting(void)
991 char *tmp, *tok, *str = strdup(sort_order);
993 for (tok = strtok_r(str, ", ", &tmp);
994 tok; tok = strtok_r(NULL, ", ", &tmp)) {
995 if (sort_dimension__add(tok) < 0) {
996 error("Unknown --sort key: `%s'", tok);
997 usage_with_options(report_usage, options);
1004 static void setup_list(struct strlist **list, const char *list_str,
1005 struct sort_entry *se, const char *list_name,
1009 *list = strlist__new(true, list_str);
1011 fprintf(stderr, "problems parsing %s list\n",
1015 if (strlist__nr_entries(*list) == 1) {
1016 fprintf(fp, "# %s: %s\n", list_name,
1017 strlist__entry(*list, 0)->s);
1023 int cmd_report(int argc, const char **argv, const char *prefix __used)
1027 argc = parse_options(argc, argv, options, report_usage, 0);
1031 if (parent_pattern != default_parent_pattern) {
1032 sort_dimension__add("parent");
1033 sort_parent.elide = 1;
1038 * Any (unrecognized) arguments left?
1041 usage_with_options(report_usage, options);
1045 setup_list(&dso_list, dso_list_str, &sort_dso, "dso", stdout);
1046 setup_list(&comm_list, comm_list_str, &sort_comm, "comm", stdout);
1047 setup_list(&sym_list, sym_list_str, &sort_sym, "symbol", stdout);
1049 if (field_sep && *field_sep == '.') {
1050 fputs("'.' is the only non valid --field-separator argument\n",
1055 return __cmd_report();