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"
27 #include "util/process_events.h"
29 static char const *input_name = "perf.data";
34 static int full_paths;
36 static int print_line;
38 static unsigned long page_size;
39 static unsigned long mmap_window = 32;
53 struct sym_hist *hist;
57 static struct symbol_conf symbol_conf = {
58 .priv_size = sizeof(struct sym_priv),
59 .try_vmlinux_path = true,
62 static const char *sym_hist_filter;
64 static int symbol_filter(struct map *map __used, struct symbol *sym)
66 if (sym_hist_filter == NULL ||
67 strcmp(sym->name, sym_hist_filter) == 0) {
68 struct sym_priv *priv = symbol__priv(sym);
69 const int size = (sizeof(*priv->hist) +
70 (sym->end - sym->start) * sizeof(u64));
72 priv->hist = malloc(size);
74 memset(priv->hist, 0, size);
78 * FIXME: We should really filter it out, as we don't want to go thru symbols
79 * we're not interested, and if a DSO ends up with no symbols, delete it too,
80 * but right now the kernel loading routines in symbol.c bail out if no symbols
81 * are found, fix it later.
87 * collect histogram counts
89 static void hist_hit(struct hist_entry *he, u64 ip)
91 unsigned int sym_size, offset;
92 struct symbol *sym = he->sym;
93 struct sym_priv *priv;
101 priv = symbol__priv(sym);
105 sym_size = sym->end - sym->start;
106 offset = ip - sym->start;
109 fprintf(stderr, "%s: ip=%Lx\n", __func__,
110 he->map->unmap_ip(he->map, ip));
112 if (offset >= sym_size)
120 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
121 (void *)(unsigned long)he->sym->start,
123 (void *)(unsigned long)ip, ip - he->sym->start,
127 static int hist_entry__add(struct thread *thread, struct map *map,
128 struct symbol *sym, u64 ip, u64 count, char level)
131 struct hist_entry *he = __hist_entry__add(thread, map, sym, NULL, ip,
140 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
143 u64 ip = event->ip.ip;
144 struct map *map = NULL;
145 struct symbol *sym = NULL;
146 struct thread *thread = threads__findnew(event->ip.pid);
148 dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
149 (void *)(offset + head),
150 (void *)(long)(event->header.size),
155 if (thread == NULL) {
156 fprintf(stderr, "problem processing %d event, skipping it.\n",
161 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
163 if (event->header.misc & PERF_RECORD_MISC_KERNEL) {
165 sym = kernel_maps__find_function(ip, &map, symbol_filter);
166 dump_printf(" ...... dso: %s\n",
167 map ? map->dso->long_name : "<not found>");
168 } else if (event->header.misc & PERF_RECORD_MISC_USER) {
170 map = thread__find_map(thread, MAP__FUNCTION, ip);
172 ip = map->map_ip(map, ip);
173 sym = map__find_symbol(map, ip, symbol_filter);
176 * If this is outside of all known maps,
177 * and is a negative address, try to look it
178 * up in the kernel dso, as it might be a
179 * vsyscall or vdso (which executes in user-mode).
181 * XXX This is nasty, we should have a symbol list in
182 * the "[vdso]" dso, but for now lets use the old
183 * trick of looking in the whole kernel symbol list.
185 if ((long long)ip < 0)
186 sym = kernel_maps__find_function(ip, &map,
189 dump_printf(" ...... dso: %s\n",
190 map ? map->dso->long_name : "<not found>");
193 dump_printf(" ...... dso: [hypervisor]\n");
196 if (hist_entry__add(thread, map, sym, ip, 1, level)) {
197 fprintf(stderr, "problem incrementing symbol count, "
207 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
209 struct thread *thread = threads__findnew(event->comm.pid);
211 dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
212 (void *)(offset + head),
213 (void *)(long)(event->header.size),
214 event->comm.comm, event->comm.pid);
216 if (thread == NULL ||
217 thread__set_comm(thread, event->comm.comm)) {
218 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
227 process_event(event_t *event, unsigned long offset, unsigned long head)
229 switch (event->header.type) {
230 case PERF_RECORD_SAMPLE:
231 return process_sample_event(event, offset, head);
233 case PERF_RECORD_MMAP:
234 return process_mmap_event(event, offset, head);
236 case PERF_RECORD_COMM:
237 return process_comm_event(event, offset, head);
239 case PERF_RECORD_FORK:
240 return process_task_event(event, offset, head);
242 * We dont process them right now but they are fine:
245 case PERF_RECORD_THROTTLE:
246 case PERF_RECORD_UNTHROTTLE:
256 static int parse_line(FILE *file, struct hist_entry *he, u64 len)
258 struct symbol *sym = he->sym;
259 char *line = NULL, *tmp, *tmp2;
260 static const char *prev_line;
261 static const char *prev_color;
269 if (getline(&line, &line_len, file) < 0)
274 c = strchr(line, '\n');
283 * Strip leading spaces:
294 * Parse hexa addresses followed by ':'
296 line_ip = strtoull(tmp, &tmp2, 16);
301 start = he->map->unmap_ip(he->map, sym->start);
304 const char *path = NULL;
305 unsigned int hits = 0;
306 double percent = 0.0;
308 struct sym_priv *priv = symbol__priv(sym);
309 struct sym_ext *sym_ext = priv->ext;
310 struct sym_hist *h = priv->hist;
312 offset = line_ip - start;
314 hits = h->ip[offset];
316 if (offset < len && sym_ext) {
317 path = sym_ext[offset].path;
318 percent = sym_ext[offset].percent;
320 percent = 100.0 * hits / h->sum;
322 color = get_percent_color(percent);
325 * Also color the filename and line if needed, with
326 * the same color than the percentage. Don't print it
327 * twice for close colored ip with the same filename:line
330 if (!prev_line || strcmp(prev_line, path)
331 || color != prev_color) {
332 color_fprintf(stdout, color, " %s", path);
338 color_fprintf(stdout, color, " %7.2f", percent);
340 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
345 printf(" : %s\n", line);
351 static struct rb_root root_sym_ext;
353 static void insert_source_line(struct sym_ext *sym_ext)
355 struct sym_ext *iter;
356 struct rb_node **p = &root_sym_ext.rb_node;
357 struct rb_node *parent = NULL;
361 iter = rb_entry(parent, struct sym_ext, node);
363 if (sym_ext->percent > iter->percent)
369 rb_link_node(&sym_ext->node, parent, p);
370 rb_insert_color(&sym_ext->node, &root_sym_ext);
373 static void free_source_line(struct hist_entry *he, int len)
375 struct sym_priv *priv = symbol__priv(he->sym);
376 struct sym_ext *sym_ext = priv->ext;
382 for (i = 0; i < len; i++)
383 free(sym_ext[i].path);
387 root_sym_ext = RB_ROOT;
390 /* Get the filename:line for the colored entries */
392 get_source_line(struct hist_entry *he, int len, const char *filename)
394 struct symbol *sym = he->sym;
397 char cmd[PATH_MAX * 2];
398 struct sym_ext *sym_ext;
399 struct sym_priv *priv = symbol__priv(sym);
400 struct sym_hist *h = priv->hist;
405 sym_ext = priv->ext = calloc(len, sizeof(struct sym_ext));
409 start = he->map->unmap_ip(he->map, sym->start);
411 for (i = 0; i < len; i++) {
417 sym_ext[i].percent = 100.0 * h->ip[i] / h->sum;
418 if (sym_ext[i].percent <= 0.5)
422 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
423 fp = popen(cmd, "r");
427 if (getline(&path, &line_len, fp) < 0 || !line_len)
430 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
431 if (!sym_ext[i].path)
434 strcpy(sym_ext[i].path, path);
435 insert_source_line(&sym_ext[i]);
442 static void print_summary(const char *filename)
444 struct sym_ext *sym_ext;
445 struct rb_node *node;
447 printf("\nSorted summary for file %s\n", filename);
448 printf("----------------------------------------------\n\n");
450 if (RB_EMPTY_ROOT(&root_sym_ext)) {
451 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
455 node = rb_first(&root_sym_ext);
461 sym_ext = rb_entry(node, struct sym_ext, node);
462 percent = sym_ext->percent;
463 color = get_percent_color(percent);
464 path = sym_ext->path;
466 color_fprintf(stdout, color, " %7.2f %s", percent, path);
467 node = rb_next(node);
471 static void annotate_sym(struct hist_entry *he)
473 struct map *map = he->map;
474 struct dso *dso = map->dso;
475 struct symbol *sym = he->sym;
476 const char *filename = dso->long_name, *d_filename;
478 char command[PATH_MAX*2];
485 fprintf(stderr, "%s: filename=%s, sym=%s, start=%Lx, end=%Lx\n",
486 __func__, filename, sym->name,
487 map->unmap_ip(map, sym->start),
488 map->unmap_ip(map, sym->end));
491 d_filename = filename;
493 d_filename = basename(filename);
495 len = sym->end - sym->start;
498 get_source_line(he, len, filename);
499 print_summary(filename);
502 printf("\n\n------------------------------------------------\n");
503 printf(" Percent | Source code & Disassembly of %s\n", d_filename);
504 printf("------------------------------------------------\n");
507 printf("annotating [%p] %30s : [%p] %30s\n",
508 dso, dso->long_name, sym, sym->name);
510 sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
511 map->unmap_ip(map, sym->start), map->unmap_ip(map, sym->end),
515 printf("doing: %s\n", command);
517 file = popen(command, "r");
521 while (!feof(file)) {
522 if (parse_line(file, he, len) < 0)
528 free_source_line(he, len);
531 static void find_annotations(void)
535 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
536 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
537 struct sym_priv *priv;
542 priv = symbol__priv(he->sym);
543 if (priv->hist == NULL)
548 * Since we have a hist_entry per IP for the same symbol, free
549 * he->sym->hist to signal we already processed this symbol.
556 static int __cmd_annotate(void)
558 int ret, rc = EXIT_FAILURE;
559 unsigned long offset = 0;
560 unsigned long head = 0;
561 struct stat input_stat;
566 register_idle_thread();
568 input = open(input_name, O_RDONLY);
570 perror("failed to open file");
574 ret = fstat(input, &input_stat);
576 perror("failed to stat file");
580 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
581 fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
585 if (!input_stat.st_size) {
586 fprintf(stderr, "zero-sized file, nothing to do!\n");
591 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
592 MAP_SHARED, input, offset);
593 if (buf == MAP_FAILED) {
594 perror("failed to mmap file");
599 event = (event_t *)(buf + head);
601 size = event->header.size;
605 if (head + event->header.size >= page_size * mmap_window) {
606 unsigned long shift = page_size * (head / page_size);
609 munmap_ret = munmap(buf, page_size * mmap_window);
610 assert(munmap_ret == 0);
617 size = event->header.size;
619 dump_printf("%p [%p]: event: %d\n",
620 (void *)(offset + head),
621 (void *)(long)event->header.size,
624 if (!size || process_event(event, offset, head) < 0) {
626 dump_printf("%p [%p]: skipping unknown header type: %d\n",
627 (void *)(offset + head),
628 (void *)(long)(event->header.size),
634 * assume we lost track of the stream, check alignment, and
635 * increment a single u64 in the hope to catch on again 'soon'.
638 if (unlikely(head & 7))
646 if (offset + head < (unsigned long)input_stat.st_size)
652 dump_printf(" IP events: %10ld\n", total);
653 dump_printf(" mmap events: %10ld\n", total_mmap);
654 dump_printf(" comm events: %10ld\n", total_comm);
655 dump_printf(" fork events: %10ld\n", total_fork);
656 dump_printf(" unknown events: %10ld\n", total_unknown);
662 threads__fprintf(stdout);
665 dsos__fprintf(stdout);
668 output__resort(total);
675 static const char * const annotate_usage[] = {
676 "perf annotate [<options>] <command>",
680 static const struct option options[] = {
681 OPT_STRING('i', "input", &input_name, "file",
683 OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
684 "symbol to annotate"),
685 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
686 OPT_BOOLEAN('v', "verbose", &verbose,
687 "be more verbose (show symbol address, etc)"),
688 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
689 "dump raw trace in ASCII"),
690 OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
691 "file", "vmlinux pathname"),
692 OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
693 "load module symbols - WARNING: use only with -k and LIVE kernel"),
694 OPT_BOOLEAN('l', "print-line", &print_line,
695 "print matching source lines (may be slow)"),
696 OPT_BOOLEAN('P', "full-paths", &full_paths,
697 "Don't shorten the displayed pathnames"),
701 static void setup_sorting(void)
703 char *tmp, *tok, *str = strdup(sort_order);
705 for (tok = strtok_r(str, ", ", &tmp);
706 tok; tok = strtok_r(NULL, ", ", &tmp)) {
707 if (sort_dimension__add(tok) < 0) {
708 error("Unknown --sort key: `%s'", tok);
709 usage_with_options(annotate_usage, options);
716 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
718 if (symbol__init(&symbol_conf) < 0)
721 page_size = getpagesize();
723 argc = parse_options(argc, argv, options, annotate_usage, 0);
729 * Special case: if there's an argument left then assume tha
730 * it's a symbol filter:
733 usage_with_options(annotate_usage, options);
735 sym_hist_filter = argv[0];
740 if (field_sep && *field_sep == '.') {
741 fputs("'.' is the only non valid --field-separator argument\n",
746 return __cmd_annotate();