4 * Builtin annotate 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"
21 #include "util/parse-options.h"
22 #include "util/parse-events.h"
23 #include "util/thread.h"
25 static char const *input_name = "perf.data";
27 static char default_sort_order[] = "comm,symbol";
28 static char *sort_order = default_sort_order;
31 static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
33 static int full_paths;
35 static int print_line;
37 static unsigned long page_size;
38 static unsigned long mmap_window = 32;
40 static struct rb_root threads;
41 static struct thread *last_match;
51 * histogram, sorted on item, collects counts
54 static struct rb_root hist;
57 struct rb_node rb_node;
59 struct thread *thread;
70 * configurable sorting bits
74 struct list_head list;
78 int64_t (*cmp)(struct hist_entry *, struct hist_entry *);
79 int64_t (*collapse)(struct hist_entry *, struct hist_entry *);
80 size_t (*print)(FILE *fp, struct hist_entry *);
86 sort__thread_cmp(struct hist_entry *left, struct hist_entry *right)
88 return right->thread->pid - left->thread->pid;
92 sort__thread_print(FILE *fp, struct hist_entry *self)
94 return fprintf(fp, "%16s:%5d", self->thread->comm ?: "", self->thread->pid);
97 static struct sort_entry sort_thread = {
98 .header = " Command: Pid",
99 .cmp = sort__thread_cmp,
100 .print = sort__thread_print,
106 sort__comm_cmp(struct hist_entry *left, struct hist_entry *right)
108 return right->thread->pid - left->thread->pid;
112 sort__comm_collapse(struct hist_entry *left, struct hist_entry *right)
114 char *comm_l = left->thread->comm;
115 char *comm_r = right->thread->comm;
117 if (!comm_l || !comm_r) {
118 if (!comm_l && !comm_r)
126 return strcmp(comm_l, comm_r);
130 sort__comm_print(FILE *fp, struct hist_entry *self)
132 return fprintf(fp, "%16s", self->thread->comm);
135 static struct sort_entry sort_comm = {
136 .header = " Command",
137 .cmp = sort__comm_cmp,
138 .collapse = sort__comm_collapse,
139 .print = sort__comm_print,
145 sort__dso_cmp(struct hist_entry *left, struct hist_entry *right)
147 struct dso *dso_l = left->dso;
148 struct dso *dso_r = right->dso;
150 if (!dso_l || !dso_r) {
151 if (!dso_l && !dso_r)
159 return strcmp(dso_l->name, dso_r->name);
163 sort__dso_print(FILE *fp, struct hist_entry *self)
166 return fprintf(fp, "%-25s", self->dso->name);
168 return fprintf(fp, "%016llx ", (u64)self->ip);
171 static struct sort_entry sort_dso = {
172 .header = "Shared Object ",
173 .cmp = sort__dso_cmp,
174 .print = sort__dso_print,
180 sort__sym_cmp(struct hist_entry *left, struct hist_entry *right)
184 if (left->sym == right->sym)
187 ip_l = left->sym ? left->sym->start : left->ip;
188 ip_r = right->sym ? right->sym->start : right->ip;
190 return (int64_t)(ip_r - ip_l);
194 sort__sym_print(FILE *fp, struct hist_entry *self)
199 ret += fprintf(fp, "%#018llx ", (u64)self->ip);
202 ret += fprintf(fp, "[%c] %s",
203 self->dso == kernel_dso ? 'k' : '.', self->sym->name);
205 ret += fprintf(fp, "%#016llx", (u64)self->ip);
211 static struct sort_entry sort_sym = {
213 .cmp = sort__sym_cmp,
214 .print = sort__sym_print,
217 static int sort__need_collapse = 0;
219 struct sort_dimension {
221 struct sort_entry *entry;
225 static struct sort_dimension sort_dimensions[] = {
226 { .name = "pid", .entry = &sort_thread, },
227 { .name = "comm", .entry = &sort_comm, },
228 { .name = "dso", .entry = &sort_dso, },
229 { .name = "symbol", .entry = &sort_sym, },
232 static LIST_HEAD(hist_entry__sort_list);
234 static int sort_dimension__add(char *tok)
238 for (i = 0; i < ARRAY_SIZE(sort_dimensions); i++) {
239 struct sort_dimension *sd = &sort_dimensions[i];
244 if (strncasecmp(tok, sd->name, strlen(tok)))
247 if (sd->entry->collapse)
248 sort__need_collapse = 1;
250 list_add_tail(&sd->entry->list, &hist_entry__sort_list);
260 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
262 struct sort_entry *se;
265 list_for_each_entry(se, &hist_entry__sort_list, list) {
266 cmp = se->cmp(left, right);
275 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
277 struct sort_entry *se;
280 list_for_each_entry(se, &hist_entry__sort_list, list) {
281 int64_t (*f)(struct hist_entry *, struct hist_entry *);
283 f = se->collapse ?: se->cmp;
285 cmp = f(left, right);
294 * collect histogram counts
296 static void hist_hit(struct hist_entry *he, u64 ip)
298 unsigned int sym_size, offset;
299 struct symbol *sym = he->sym;
303 if (!sym || !sym->hist)
306 sym_size = sym->end - sym->start;
307 offset = ip - sym->start;
309 if (offset >= sym_size)
316 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
317 (void *)(unsigned long)he->sym->start,
319 (void *)(unsigned long)ip, ip - he->sym->start,
324 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
325 struct symbol *sym, u64 ip, char level)
327 struct rb_node **p = &hist.rb_node;
328 struct rb_node *parent = NULL;
329 struct hist_entry *he;
330 struct hist_entry entry = {
343 he = rb_entry(parent, struct hist_entry, rb_node);
345 cmp = hist_entry__cmp(&entry, he);
359 he = malloc(sizeof(*he));
363 rb_link_node(&he->rb_node, parent, p);
364 rb_insert_color(&he->rb_node, &hist);
369 static void hist_entry__free(struct hist_entry *he)
375 * collapse the histogram
378 static struct rb_root collapse_hists;
380 static void collapse__insert_entry(struct hist_entry *he)
382 struct rb_node **p = &collapse_hists.rb_node;
383 struct rb_node *parent = NULL;
384 struct hist_entry *iter;
389 iter = rb_entry(parent, struct hist_entry, rb_node);
391 cmp = hist_entry__collapse(iter, he);
394 iter->count += he->count;
395 hist_entry__free(he);
405 rb_link_node(&he->rb_node, parent, p);
406 rb_insert_color(&he->rb_node, &collapse_hists);
409 static void collapse__resort(void)
411 struct rb_node *next;
412 struct hist_entry *n;
414 if (!sort__need_collapse)
417 next = rb_first(&hist);
419 n = rb_entry(next, struct hist_entry, rb_node);
420 next = rb_next(&n->rb_node);
422 rb_erase(&n->rb_node, &hist);
423 collapse__insert_entry(n);
428 * reverse the map, sort on count.
431 static struct rb_root output_hists;
433 static void output__insert_entry(struct hist_entry *he)
435 struct rb_node **p = &output_hists.rb_node;
436 struct rb_node *parent = NULL;
437 struct hist_entry *iter;
441 iter = rb_entry(parent, struct hist_entry, rb_node);
443 if (he->count > iter->count)
449 rb_link_node(&he->rb_node, parent, p);
450 rb_insert_color(&he->rb_node, &output_hists);
453 static void output__resort(void)
455 struct rb_node *next;
456 struct hist_entry *n;
457 struct rb_root *tree = &hist;
459 if (sort__need_collapse)
460 tree = &collapse_hists;
462 next = rb_first(tree);
465 n = rb_entry(next, struct hist_entry, rb_node);
466 next = rb_next(&n->rb_node);
468 rb_erase(&n->rb_node, tree);
469 output__insert_entry(n);
473 static void register_idle_thread(void)
475 struct thread *thread = threads__findnew(0, &threads, &last_match);
477 if (thread == NULL ||
478 thread__set_comm(thread, "[idle]")) {
479 fprintf(stderr, "problem inserting idle task.\n");
484 static unsigned long total = 0,
491 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
495 struct dso *dso = NULL;
496 struct thread *thread;
497 u64 ip = event->ip.ip;
498 struct map *map = NULL;
500 thread = threads__findnew(event->ip.pid, &threads, &last_match);
502 dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
503 (void *)(offset + head),
504 (void *)(long)(event->header.size),
509 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
511 if (thread == NULL) {
512 fprintf(stderr, "problem processing %d event, skipping it.\n",
517 if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
523 dump_printf(" ...... dso: %s\n", dso->name);
525 } else if (event->header.misc & PERF_EVENT_MISC_USER) {
530 map = thread__find_map(thread, ip);
532 ip = map->map_ip(map, ip);
536 * If this is outside of all known maps,
537 * and is a negative address, try to look it
538 * up in the kernel dso, as it might be a
539 * vsyscall (which executes in user-mode):
541 if ((long long)ip < 0)
544 dump_printf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
549 dump_printf(" ...... dso: [hypervisor]\n");
552 if (show & show_mask) {
553 struct symbol *sym = NULL;
556 sym = dso->find_symbol(dso, ip);
558 if (hist_entry__add(thread, map, dso, sym, ip, level)) {
560 "problem incrementing symbol count, skipping event\n");
570 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
572 struct thread *thread;
573 struct map *map = map__new(&event->mmap, NULL, 0);
575 thread = threads__findnew(event->mmap.pid, &threads, &last_match);
577 dump_printf("%p [%p]: PERF_EVENT_MMAP %d: [%p(%p) @ %p]: %s\n",
578 (void *)(offset + head),
579 (void *)(long)(event->header.size),
581 (void *)(long)event->mmap.start,
582 (void *)(long)event->mmap.len,
583 (void *)(long)event->mmap.pgoff,
584 event->mmap.filename);
586 if (thread == NULL || map == NULL) {
587 dump_printf("problem processing PERF_EVENT_MMAP, skipping event.\n");
591 thread__insert_map(thread, map);
598 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
600 struct thread *thread;
602 thread = threads__findnew(event->comm.pid, &threads, &last_match);
603 dump_printf("%p [%p]: PERF_EVENT_COMM: %s:%d\n",
604 (void *)(offset + head),
605 (void *)(long)(event->header.size),
606 event->comm.comm, event->comm.pid);
608 if (thread == NULL ||
609 thread__set_comm(thread, event->comm.comm)) {
610 dump_printf("problem processing PERF_EVENT_COMM, skipping event.\n");
619 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
621 struct thread *thread;
622 struct thread *parent;
624 thread = threads__findnew(event->fork.pid, &threads, &last_match);
625 parent = threads__findnew(event->fork.ppid, &threads, &last_match);
626 dump_printf("%p [%p]: PERF_EVENT_FORK: %d:%d\n",
627 (void *)(offset + head),
628 (void *)(long)(event->header.size),
629 event->fork.pid, event->fork.ppid);
631 if (!thread || !parent || thread__fork(thread, parent)) {
632 dump_printf("problem processing PERF_EVENT_FORK, skipping event.\n");
641 process_event(event_t *event, unsigned long offset, unsigned long head)
643 switch (event->header.type) {
644 case PERF_EVENT_SAMPLE:
645 return process_sample_event(event, offset, head);
647 case PERF_EVENT_MMAP:
648 return process_mmap_event(event, offset, head);
650 case PERF_EVENT_COMM:
651 return process_comm_event(event, offset, head);
653 case PERF_EVENT_FORK:
654 return process_fork_event(event, offset, head);
656 * We dont process them right now but they are fine:
659 case PERF_EVENT_THROTTLE:
660 case PERF_EVENT_UNTHROTTLE:
671 parse_line(FILE *file, struct symbol *sym, u64 start, u64 len)
673 char *line = NULL, *tmp, *tmp2;
674 static const char *prev_line;
675 static const char *prev_color;
682 if (getline(&line, &line_len, file) < 0)
687 c = strchr(line, '\n');
696 * Strip leading spaces:
707 * Parse hexa addresses followed by ':'
709 line_ip = strtoull(tmp, &tmp2, 16);
715 const char *path = NULL;
716 unsigned int hits = 0;
717 double percent = 0.0;
719 struct sym_ext *sym_ext = sym->priv;
721 offset = line_ip - start;
723 hits = sym->hist[offset];
725 if (offset < len && sym_ext) {
726 path = sym_ext[offset].path;
727 percent = sym_ext[offset].percent;
728 } else if (sym->hist_sum)
729 percent = 100.0 * hits / sym->hist_sum;
731 color = get_percent_color(percent);
734 * Also color the filename and line if needed, with
735 * the same color than the percentage. Don't print it
736 * twice for close colored ip with the same filename:line
739 if (!prev_line || strcmp(prev_line, path)
740 || color != prev_color) {
741 color_fprintf(stdout, color, " %s", path);
747 color_fprintf(stdout, color, " %7.2f", percent);
749 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
754 printf(" : %s\n", line);
760 static struct rb_root root_sym_ext;
762 static void insert_source_line(struct sym_ext *sym_ext)
764 struct sym_ext *iter;
765 struct rb_node **p = &root_sym_ext.rb_node;
766 struct rb_node *parent = NULL;
770 iter = rb_entry(parent, struct sym_ext, node);
772 if (sym_ext->percent > iter->percent)
778 rb_link_node(&sym_ext->node, parent, p);
779 rb_insert_color(&sym_ext->node, &root_sym_ext);
782 static void free_source_line(struct symbol *sym, int len)
784 struct sym_ext *sym_ext = sym->priv;
790 for (i = 0; i < len; i++)
791 free(sym_ext[i].path);
795 root_sym_ext = RB_ROOT;
798 /* Get the filename:line for the colored entries */
800 get_source_line(struct symbol *sym, u64 start, int len, const char *filename)
803 char cmd[PATH_MAX * 2];
804 struct sym_ext *sym_ext;
809 sym->priv = calloc(len, sizeof(struct sym_ext));
815 for (i = 0; i < len; i++) {
821 sym_ext[i].percent = 100.0 * sym->hist[i] / sym->hist_sum;
822 if (sym_ext[i].percent <= 0.5)
826 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
827 fp = popen(cmd, "r");
831 if (getline(&path, &line_len, fp) < 0 || !line_len)
834 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
835 if (!sym_ext[i].path)
838 strcpy(sym_ext[i].path, path);
839 insert_source_line(&sym_ext[i]);
846 static void print_summary(const char *filename)
848 struct sym_ext *sym_ext;
849 struct rb_node *node;
851 printf("\nSorted summary for file %s\n", filename);
852 printf("----------------------------------------------\n\n");
854 if (RB_EMPTY_ROOT(&root_sym_ext)) {
855 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
859 node = rb_first(&root_sym_ext);
865 sym_ext = rb_entry(node, struct sym_ext, node);
866 percent = sym_ext->percent;
867 color = get_percent_color(percent);
868 path = sym_ext->path;
870 color_fprintf(stdout, color, " %7.2f %s", percent, path);
871 node = rb_next(node);
875 static void annotate_sym(struct dso *dso, struct symbol *sym)
877 const char *filename = dso->name, *d_filename;
879 char command[PATH_MAX*2];
885 filename = sym->module->path;
886 else if (dso == kernel_dso)
887 filename = vmlinux_name;
889 start = sym->obj_start;
893 d_filename = filename;
895 d_filename = basename(filename);
897 end = start + sym->end - sym->start + 1;
898 len = sym->end - sym->start;
901 get_source_line(sym, start, len, filename);
902 print_summary(filename);
905 printf("\n\n------------------------------------------------\n");
906 printf(" Percent | Source code & Disassembly of %s\n", d_filename);
907 printf("------------------------------------------------\n");
910 printf("annotating [%p] %30s : [%p] %30s\n", dso, dso->name, sym, sym->name);
912 sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
913 (u64)start, (u64)end, filename, filename);
916 printf("doing: %s\n", command);
918 file = popen(command, "r");
922 while (!feof(file)) {
923 if (parse_line(file, sym, start, len) < 0)
929 free_source_line(sym, len);
932 static void find_annotations(void)
938 list_for_each_entry(dso, &dsos, node) {
940 for (nd = rb_first(&dso->syms); nd; nd = rb_next(nd)) {
941 struct symbol *sym = rb_entry(nd, struct symbol, rb_node);
944 annotate_sym(dso, sym);
951 printf(" Error: symbol '%s' not present amongst the samples.\n", sym_hist_filter);
954 static int __cmd_annotate(void)
956 int ret, rc = EXIT_FAILURE;
957 unsigned long offset = 0;
958 unsigned long head = 0;
959 struct stat input_stat;
964 register_idle_thread();
966 input = open(input_name, O_RDONLY);
968 perror("failed to open file");
972 ret = fstat(input, &input_stat);
974 perror("failed to stat file");
978 if (!input_stat.st_size) {
979 fprintf(stderr, "zero-sized file, nothing to do!\n");
983 if (load_kernel() < 0) {
984 perror("failed to load kernel symbols");
989 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
990 MAP_SHARED, input, offset);
991 if (buf == MAP_FAILED) {
992 perror("failed to mmap file");
997 event = (event_t *)(buf + head);
999 size = event->header.size;
1003 if (head + event->header.size >= page_size * mmap_window) {
1004 unsigned long shift = page_size * (head / page_size);
1007 munmap_ret = munmap(buf, page_size * mmap_window);
1008 assert(munmap_ret == 0);
1015 size = event->header.size;
1017 dump_printf("%p [%p]: event: %d\n",
1018 (void *)(offset + head),
1019 (void *)(long)event->header.size,
1020 event->header.type);
1022 if (!size || process_event(event, offset, head) < 0) {
1024 dump_printf("%p [%p]: skipping unknown header type: %d\n",
1025 (void *)(offset + head),
1026 (void *)(long)(event->header.size),
1027 event->header.type);
1032 * assume we lost track of the stream, check alignment, and
1033 * increment a single u64 in the hope to catch on again 'soon'.
1036 if (unlikely(head & 7))
1044 if (offset + head < (unsigned long)input_stat.st_size)
1050 dump_printf(" IP events: %10ld\n", total);
1051 dump_printf(" mmap events: %10ld\n", total_mmap);
1052 dump_printf(" comm events: %10ld\n", total_comm);
1053 dump_printf(" fork events: %10ld\n", total_fork);
1054 dump_printf(" unknown events: %10ld\n", total_unknown);
1060 threads__fprintf(stdout, &threads);
1063 dsos__fprintf(stdout);
1073 static const char * const annotate_usage[] = {
1074 "perf annotate [<options>] <command>",
1078 static const struct option options[] = {
1079 OPT_STRING('i', "input", &input_name, "file",
1081 OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
1082 "symbol to annotate"),
1083 OPT_BOOLEAN('v', "verbose", &verbose,
1084 "be more verbose (show symbol address, etc)"),
1085 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1086 "dump raw trace in ASCII"),
1087 OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
1088 OPT_BOOLEAN('m', "modules", &modules,
1089 "load module symbols - WARNING: use only with -k and LIVE kernel"),
1090 OPT_BOOLEAN('l', "print-line", &print_line,
1091 "print matching source lines (may be slow)"),
1092 OPT_BOOLEAN('P', "full-paths", &full_paths,
1093 "Don't shorten the displayed pathnames"),
1097 static void setup_sorting(void)
1099 char *tmp, *tok, *str = strdup(sort_order);
1101 for (tok = strtok_r(str, ", ", &tmp);
1102 tok; tok = strtok_r(NULL, ", ", &tmp)) {
1103 if (sort_dimension__add(tok) < 0) {
1104 error("Unknown --sort key: `%s'", tok);
1105 usage_with_options(annotate_usage, options);
1112 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
1116 page_size = getpagesize();
1118 argc = parse_options(argc, argv, options, annotate_usage, 0);
1124 * Special case: if there's an argument left then assume tha
1125 * it's a symbol filter:
1128 usage_with_options(annotate_usage, options);
1130 sym_hist_filter = argv[0];
1133 if (!sym_hist_filter)
1134 usage_with_options(annotate_usage, options);
1138 return __cmd_annotate();