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"
20 #include "util/debug.h"
22 #include "util/parse-options.h"
23 #include "util/parse-events.h"
24 #include "util/thread.h"
25 #include "util/sort.h"
26 #include "util/hist.h"
28 static char const *input_name = "perf.data";
33 static int full_paths;
35 static int print_line;
37 static unsigned long page_size;
38 static unsigned long mmap_window = 32;
52 struct sym_hist *hist;
56 static const char *sym_hist_filter;
58 static int symbol_filter(struct map *map, struct symbol *sym)
60 if (sym_hist_filter == NULL ||
61 strcmp(sym->name, sym_hist_filter) == 0) {
62 struct sym_priv *priv = dso__sym_priv(map->dso, sym);
63 const int size = (sizeof(*priv->hist) +
64 (sym->end - sym->start) * sizeof(u64));
66 priv->hist = malloc(size);
68 memset(priv->hist, 0, size);
72 * FIXME: We should really filter it out, as we don't want to go thru symbols
73 * we're not interested, and if a DSO ends up with no symbols, delete it too,
74 * but right now the kernel loading routines in symbol.c bail out if no symbols
75 * are found, fix it later.
81 * collect histogram counts
83 static void hist_hit(struct hist_entry *he, u64 ip)
85 unsigned int sym_size, offset;
86 struct symbol *sym = he->sym;
87 struct sym_priv *priv;
95 priv = dso__sym_priv(he->map->dso, sym);
99 sym_size = sym->end - sym->start;
100 offset = ip - sym->start;
103 fprintf(stderr, "%s: ip=%Lx\n", __func__,
104 he->map->unmap_ip(he->map, ip));
106 if (offset >= sym_size)
114 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
115 (void *)(unsigned long)he->sym->start,
117 (void *)(unsigned long)ip, ip - he->sym->start,
121 static int hist_entry__add(struct thread *thread, struct map *map,
122 struct symbol *sym, u64 ip, u64 count, char level)
125 struct hist_entry *he = __hist_entry__add(thread, map, sym, NULL, ip,
134 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
137 u64 ip = event->ip.ip;
138 struct map *map = NULL;
139 struct symbol *sym = NULL;
140 struct thread *thread = threads__findnew(event->ip.pid);
142 dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
143 (void *)(offset + head),
144 (void *)(long)(event->header.size),
149 if (thread == NULL) {
150 fprintf(stderr, "problem processing %d event, skipping it.\n",
155 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
157 if (event->header.misc & PERF_RECORD_MISC_KERNEL) {
159 sym = kernel_maps__find_symbol(ip, &map);
160 dump_printf(" ...... dso: %s\n",
161 map ? map->dso->long_name : "<not found>");
162 } else if (event->header.misc & PERF_RECORD_MISC_USER) {
164 map = thread__find_map(thread, ip);
167 ip = map->map_ip(map, ip);
168 sym = map->dso->find_symbol(map->dso, ip);
171 * If this is outside of all known maps,
172 * and is a negative address, try to look it
173 * up in the kernel dso, as it might be a
174 * vsyscall or vdso (which executes in user-mode).
176 * XXX This is nasty, we should have a symbol list in
177 * the "[vdso]" dso, but for now lets use the old
178 * trick of looking in the whole kernel symbol list.
180 if ((long long)ip < 0) {
185 dump_printf(" ...... dso: %s\n",
186 map ? map->dso->long_name : "<not found>");
189 dump_printf(" ...... dso: [hypervisor]\n");
192 if (hist_entry__add(thread, map, sym, ip, 1, level)) {
193 fprintf(stderr, "problem incrementing symbol count, "
203 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
205 struct map *map = map__new(&event->mmap, NULL, 0,
206 sizeof(struct sym_priv), symbol_filter,
208 struct thread *thread = threads__findnew(event->mmap.pid);
210 dump_printf("%p [%p]: PERF_RECORD_MMAP %d: [%p(%p) @ %p]: %s\n",
211 (void *)(offset + head),
212 (void *)(long)(event->header.size),
214 (void *)(long)event->mmap.start,
215 (void *)(long)event->mmap.len,
216 (void *)(long)event->mmap.pgoff,
217 event->mmap.filename);
219 if (thread == NULL || map == NULL) {
220 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
224 thread__insert_map(thread, map);
231 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
233 struct thread *thread = threads__findnew(event->comm.pid);
235 dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
236 (void *)(offset + head),
237 (void *)(long)(event->header.size),
238 event->comm.comm, event->comm.pid);
240 if (thread == NULL ||
241 thread__set_comm(thread, event->comm.comm)) {
242 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
251 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
253 struct thread *thread = threads__findnew(event->fork.pid);
254 struct thread *parent = threads__findnew(event->fork.ppid);
256 dump_printf("%p [%p]: PERF_RECORD_FORK: %d:%d\n",
257 (void *)(offset + head),
258 (void *)(long)(event->header.size),
259 event->fork.pid, event->fork.ppid);
262 * A thread clone will have the same PID for both
265 if (thread == parent)
268 if (!thread || !parent || thread__fork(thread, parent)) {
269 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
278 process_event(event_t *event, unsigned long offset, unsigned long head)
280 switch (event->header.type) {
281 case PERF_RECORD_SAMPLE:
282 return process_sample_event(event, offset, head);
284 case PERF_RECORD_MMAP:
285 return process_mmap_event(event, offset, head);
287 case PERF_RECORD_COMM:
288 return process_comm_event(event, offset, head);
290 case PERF_RECORD_FORK:
291 return process_fork_event(event, offset, head);
293 * We dont process them right now but they are fine:
296 case PERF_RECORD_THROTTLE:
297 case PERF_RECORD_UNTHROTTLE:
307 static int parse_line(FILE *file, struct hist_entry *he, u64 len)
309 struct symbol *sym = he->sym;
310 char *line = NULL, *tmp, *tmp2;
311 static const char *prev_line;
312 static const char *prev_color;
320 if (getline(&line, &line_len, file) < 0)
325 c = strchr(line, '\n');
334 * Strip leading spaces:
345 * Parse hexa addresses followed by ':'
347 line_ip = strtoull(tmp, &tmp2, 16);
352 start = he->map->unmap_ip(he->map, sym->start);
355 const char *path = NULL;
356 unsigned int hits = 0;
357 double percent = 0.0;
359 struct sym_priv *priv = dso__sym_priv(he->map->dso, sym);
360 struct sym_ext *sym_ext = priv->ext;
361 struct sym_hist *h = priv->hist;
363 offset = line_ip - start;
365 hits = h->ip[offset];
367 if (offset < len && sym_ext) {
368 path = sym_ext[offset].path;
369 percent = sym_ext[offset].percent;
371 percent = 100.0 * hits / h->sum;
373 color = get_percent_color(percent);
376 * Also color the filename and line if needed, with
377 * the same color than the percentage. Don't print it
378 * twice for close colored ip with the same filename:line
381 if (!prev_line || strcmp(prev_line, path)
382 || color != prev_color) {
383 color_fprintf(stdout, color, " %s", path);
389 color_fprintf(stdout, color, " %7.2f", percent);
391 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
396 printf(" : %s\n", line);
402 static struct rb_root root_sym_ext;
404 static void insert_source_line(struct sym_ext *sym_ext)
406 struct sym_ext *iter;
407 struct rb_node **p = &root_sym_ext.rb_node;
408 struct rb_node *parent = NULL;
412 iter = rb_entry(parent, struct sym_ext, node);
414 if (sym_ext->percent > iter->percent)
420 rb_link_node(&sym_ext->node, parent, p);
421 rb_insert_color(&sym_ext->node, &root_sym_ext);
424 static void free_source_line(struct hist_entry *he, int len)
426 struct sym_priv *priv = dso__sym_priv(he->map->dso, he->sym);
427 struct sym_ext *sym_ext = priv->ext;
433 for (i = 0; i < len; i++)
434 free(sym_ext[i].path);
438 root_sym_ext = RB_ROOT;
441 /* Get the filename:line for the colored entries */
443 get_source_line(struct hist_entry *he, int len, const char *filename)
445 struct symbol *sym = he->sym;
448 char cmd[PATH_MAX * 2];
449 struct sym_ext *sym_ext;
450 struct sym_priv *priv = dso__sym_priv(he->map->dso, sym);
451 struct sym_hist *h = priv->hist;
456 sym_ext = priv->ext = calloc(len, sizeof(struct sym_ext));
460 start = he->map->unmap_ip(he->map, sym->start);
462 for (i = 0; i < len; i++) {
468 sym_ext[i].percent = 100.0 * h->ip[i] / h->sum;
469 if (sym_ext[i].percent <= 0.5)
473 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
474 fp = popen(cmd, "r");
478 if (getline(&path, &line_len, fp) < 0 || !line_len)
481 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
482 if (!sym_ext[i].path)
485 strcpy(sym_ext[i].path, path);
486 insert_source_line(&sym_ext[i]);
493 static void print_summary(const char *filename)
495 struct sym_ext *sym_ext;
496 struct rb_node *node;
498 printf("\nSorted summary for file %s\n", filename);
499 printf("----------------------------------------------\n\n");
501 if (RB_EMPTY_ROOT(&root_sym_ext)) {
502 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
506 node = rb_first(&root_sym_ext);
512 sym_ext = rb_entry(node, struct sym_ext, node);
513 percent = sym_ext->percent;
514 color = get_percent_color(percent);
515 path = sym_ext->path;
517 color_fprintf(stdout, color, " %7.2f %s", percent, path);
518 node = rb_next(node);
522 static void annotate_sym(struct hist_entry *he)
524 struct map *map = he->map;
525 struct dso *dso = map->dso;
526 struct symbol *sym = he->sym;
527 const char *filename = dso->long_name, *d_filename;
529 char command[PATH_MAX*2];
536 fprintf(stderr, "%s: filename=%s, sym=%s, start=%Lx, end=%Lx\n",
537 __func__, filename, sym->name,
538 map->unmap_ip(map, sym->start),
539 map->unmap_ip(map, sym->end));
542 d_filename = filename;
544 d_filename = basename(filename);
546 len = sym->end - sym->start;
549 get_source_line(he, len, filename);
550 print_summary(filename);
553 printf("\n\n------------------------------------------------\n");
554 printf(" Percent | Source code & Disassembly of %s\n", d_filename);
555 printf("------------------------------------------------\n");
558 printf("annotating [%p] %30s : [%p] %30s\n",
559 dso, dso->long_name, sym, sym->name);
561 sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
562 map->unmap_ip(map, sym->start), map->unmap_ip(map, sym->end),
566 printf("doing: %s\n", command);
568 file = popen(command, "r");
572 while (!feof(file)) {
573 if (parse_line(file, he, len) < 0)
579 free_source_line(he, len);
582 static void find_annotations(void)
586 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
587 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
588 struct sym_priv *priv;
593 priv = dso__sym_priv(he->map->dso, he->sym);
594 if (priv->hist == NULL)
599 * Since we have a hist_entry per IP for the same symbol, free
600 * he->sym->hist to signal we already processed this symbol.
607 static int __cmd_annotate(void)
609 int ret, rc = EXIT_FAILURE;
610 unsigned long offset = 0;
611 unsigned long head = 0;
612 struct stat input_stat;
617 register_idle_thread();
619 input = open(input_name, O_RDONLY);
621 perror("failed to open file");
625 ret = fstat(input, &input_stat);
627 perror("failed to stat file");
631 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
632 fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
636 if (!input_stat.st_size) {
637 fprintf(stderr, "zero-sized file, nothing to do!\n");
641 if (load_kernel(sizeof(struct sym_priv), symbol_filter) < 0) {
642 perror("failed to load kernel symbols");
647 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
648 MAP_SHARED, input, offset);
649 if (buf == MAP_FAILED) {
650 perror("failed to mmap file");
655 event = (event_t *)(buf + head);
657 size = event->header.size;
661 if (head + event->header.size >= page_size * mmap_window) {
662 unsigned long shift = page_size * (head / page_size);
665 munmap_ret = munmap(buf, page_size * mmap_window);
666 assert(munmap_ret == 0);
673 size = event->header.size;
675 dump_printf("%p [%p]: event: %d\n",
676 (void *)(offset + head),
677 (void *)(long)event->header.size,
680 if (!size || process_event(event, offset, head) < 0) {
682 dump_printf("%p [%p]: skipping unknown header type: %d\n",
683 (void *)(offset + head),
684 (void *)(long)(event->header.size),
690 * assume we lost track of the stream, check alignment, and
691 * increment a single u64 in the hope to catch on again 'soon'.
694 if (unlikely(head & 7))
702 if (offset + head < (unsigned long)input_stat.st_size)
708 dump_printf(" IP events: %10ld\n", total);
709 dump_printf(" mmap events: %10ld\n", total_mmap);
710 dump_printf(" comm events: %10ld\n", total_comm);
711 dump_printf(" fork events: %10ld\n", total_fork);
712 dump_printf(" unknown events: %10ld\n", total_unknown);
718 threads__fprintf(stdout);
721 dsos__fprintf(stdout);
724 output__resort(total);
731 static const char * const annotate_usage[] = {
732 "perf annotate [<options>] <command>",
736 static const struct option options[] = {
737 OPT_STRING('i', "input", &input_name, "file",
739 OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
740 "symbol to annotate"),
741 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
742 OPT_BOOLEAN('v', "verbose", &verbose,
743 "be more verbose (show symbol address, etc)"),
744 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
745 "dump raw trace in ASCII"),
746 OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
747 OPT_BOOLEAN('m', "modules", &modules,
748 "load module symbols - WARNING: use only with -k and LIVE kernel"),
749 OPT_BOOLEAN('l', "print-line", &print_line,
750 "print matching source lines (may be slow)"),
751 OPT_BOOLEAN('P', "full-paths", &full_paths,
752 "Don't shorten the displayed pathnames"),
756 static void setup_sorting(void)
758 char *tmp, *tok, *str = strdup(sort_order);
760 for (tok = strtok_r(str, ", ", &tmp);
761 tok; tok = strtok_r(NULL, ", ", &tmp)) {
762 if (sort_dimension__add(tok) < 0) {
763 error("Unknown --sort key: `%s'", tok);
764 usage_with_options(annotate_usage, options);
771 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
775 page_size = getpagesize();
777 argc = parse_options(argc, argv, options, annotate_usage, 0);
783 * Special case: if there's an argument left then assume tha
784 * it's a symbol filter:
787 usage_with_options(annotate_usage, options);
789 sym_hist_filter = argv[0];
794 if (field_sep && *field_sep == '.') {
795 fputs("'.' is the only non valid --field-separator argument\n",
800 return __cmd_annotate();