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__find_symbol(map, ip, symbol_filter);
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));
207 struct thread *thread = threads__findnew(event->mmap.pid);
209 dump_printf("%p [%p]: PERF_RECORD_MMAP %d: [%p(%p) @ %p]: %s\n",
210 (void *)(offset + head),
211 (void *)(long)(event->header.size),
213 (void *)(long)event->mmap.start,
214 (void *)(long)event->mmap.len,
215 (void *)(long)event->mmap.pgoff,
216 event->mmap.filename);
218 if (thread == NULL || map == NULL) {
219 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
223 thread__insert_map(thread, map);
230 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
232 struct thread *thread = threads__findnew(event->comm.pid);
234 dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
235 (void *)(offset + head),
236 (void *)(long)(event->header.size),
237 event->comm.comm, event->comm.pid);
239 if (thread == NULL ||
240 thread__set_comm(thread, event->comm.comm)) {
241 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
250 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
252 struct thread *thread = threads__findnew(event->fork.pid);
253 struct thread *parent = threads__findnew(event->fork.ppid);
255 dump_printf("%p [%p]: PERF_RECORD_FORK: %d:%d\n",
256 (void *)(offset + head),
257 (void *)(long)(event->header.size),
258 event->fork.pid, event->fork.ppid);
261 * A thread clone will have the same PID for both
264 if (thread == parent)
267 if (!thread || !parent || thread__fork(thread, parent)) {
268 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
277 process_event(event_t *event, unsigned long offset, unsigned long head)
279 switch (event->header.type) {
280 case PERF_RECORD_SAMPLE:
281 return process_sample_event(event, offset, head);
283 case PERF_RECORD_MMAP:
284 return process_mmap_event(event, offset, head);
286 case PERF_RECORD_COMM:
287 return process_comm_event(event, offset, head);
289 case PERF_RECORD_FORK:
290 return process_fork_event(event, offset, head);
292 * We dont process them right now but they are fine:
295 case PERF_RECORD_THROTTLE:
296 case PERF_RECORD_UNTHROTTLE:
306 static int parse_line(FILE *file, struct hist_entry *he, u64 len)
308 struct symbol *sym = he->sym;
309 char *line = NULL, *tmp, *tmp2;
310 static const char *prev_line;
311 static const char *prev_color;
319 if (getline(&line, &line_len, file) < 0)
324 c = strchr(line, '\n');
333 * Strip leading spaces:
344 * Parse hexa addresses followed by ':'
346 line_ip = strtoull(tmp, &tmp2, 16);
351 start = he->map->unmap_ip(he->map, sym->start);
354 const char *path = NULL;
355 unsigned int hits = 0;
356 double percent = 0.0;
358 struct sym_priv *priv = dso__sym_priv(he->map->dso, sym);
359 struct sym_ext *sym_ext = priv->ext;
360 struct sym_hist *h = priv->hist;
362 offset = line_ip - start;
364 hits = h->ip[offset];
366 if (offset < len && sym_ext) {
367 path = sym_ext[offset].path;
368 percent = sym_ext[offset].percent;
370 percent = 100.0 * hits / h->sum;
372 color = get_percent_color(percent);
375 * Also color the filename and line if needed, with
376 * the same color than the percentage. Don't print it
377 * twice for close colored ip with the same filename:line
380 if (!prev_line || strcmp(prev_line, path)
381 || color != prev_color) {
382 color_fprintf(stdout, color, " %s", path);
388 color_fprintf(stdout, color, " %7.2f", percent);
390 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
395 printf(" : %s\n", line);
401 static struct rb_root root_sym_ext;
403 static void insert_source_line(struct sym_ext *sym_ext)
405 struct sym_ext *iter;
406 struct rb_node **p = &root_sym_ext.rb_node;
407 struct rb_node *parent = NULL;
411 iter = rb_entry(parent, struct sym_ext, node);
413 if (sym_ext->percent > iter->percent)
419 rb_link_node(&sym_ext->node, parent, p);
420 rb_insert_color(&sym_ext->node, &root_sym_ext);
423 static void free_source_line(struct hist_entry *he, int len)
425 struct sym_priv *priv = dso__sym_priv(he->map->dso, he->sym);
426 struct sym_ext *sym_ext = priv->ext;
432 for (i = 0; i < len; i++)
433 free(sym_ext[i].path);
437 root_sym_ext = RB_ROOT;
440 /* Get the filename:line for the colored entries */
442 get_source_line(struct hist_entry *he, int len, const char *filename)
444 struct symbol *sym = he->sym;
447 char cmd[PATH_MAX * 2];
448 struct sym_ext *sym_ext;
449 struct sym_priv *priv = dso__sym_priv(he->map->dso, sym);
450 struct sym_hist *h = priv->hist;
455 sym_ext = priv->ext = calloc(len, sizeof(struct sym_ext));
459 start = he->map->unmap_ip(he->map, sym->start);
461 for (i = 0; i < len; i++) {
467 sym_ext[i].percent = 100.0 * h->ip[i] / h->sum;
468 if (sym_ext[i].percent <= 0.5)
472 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
473 fp = popen(cmd, "r");
477 if (getline(&path, &line_len, fp) < 0 || !line_len)
480 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
481 if (!sym_ext[i].path)
484 strcpy(sym_ext[i].path, path);
485 insert_source_line(&sym_ext[i]);
492 static void print_summary(const char *filename)
494 struct sym_ext *sym_ext;
495 struct rb_node *node;
497 printf("\nSorted summary for file %s\n", filename);
498 printf("----------------------------------------------\n\n");
500 if (RB_EMPTY_ROOT(&root_sym_ext)) {
501 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
505 node = rb_first(&root_sym_ext);
511 sym_ext = rb_entry(node, struct sym_ext, node);
512 percent = sym_ext->percent;
513 color = get_percent_color(percent);
514 path = sym_ext->path;
516 color_fprintf(stdout, color, " %7.2f %s", percent, path);
517 node = rb_next(node);
521 static void annotate_sym(struct hist_entry *he)
523 struct map *map = he->map;
524 struct dso *dso = map->dso;
525 struct symbol *sym = he->sym;
526 const char *filename = dso->long_name, *d_filename;
528 char command[PATH_MAX*2];
535 fprintf(stderr, "%s: filename=%s, sym=%s, start=%Lx, end=%Lx\n",
536 __func__, filename, sym->name,
537 map->unmap_ip(map, sym->start),
538 map->unmap_ip(map, sym->end));
541 d_filename = filename;
543 d_filename = basename(filename);
545 len = sym->end - sym->start;
548 get_source_line(he, len, filename);
549 print_summary(filename);
552 printf("\n\n------------------------------------------------\n");
553 printf(" Percent | Source code & Disassembly of %s\n", d_filename);
554 printf("------------------------------------------------\n");
557 printf("annotating [%p] %30s : [%p] %30s\n",
558 dso, dso->long_name, sym, sym->name);
560 sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
561 map->unmap_ip(map, sym->start), map->unmap_ip(map, sym->end),
565 printf("doing: %s\n", command);
567 file = popen(command, "r");
571 while (!feof(file)) {
572 if (parse_line(file, he, len) < 0)
578 free_source_line(he, len);
581 static void find_annotations(void)
585 for (nd = rb_first(&output_hists); nd; nd = rb_next(nd)) {
586 struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
587 struct sym_priv *priv;
592 priv = dso__sym_priv(he->map->dso, he->sym);
593 if (priv->hist == NULL)
598 * Since we have a hist_entry per IP for the same symbol, free
599 * he->sym->hist to signal we already processed this symbol.
606 static int __cmd_annotate(void)
608 int ret, rc = EXIT_FAILURE;
609 unsigned long offset = 0;
610 unsigned long head = 0;
611 struct stat input_stat;
616 register_idle_thread();
618 input = open(input_name, O_RDONLY);
620 perror("failed to open file");
624 ret = fstat(input, &input_stat);
626 perror("failed to stat file");
630 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
631 fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
635 if (!input_stat.st_size) {
636 fprintf(stderr, "zero-sized file, nothing to do!\n");
640 if (load_kernel(sizeof(struct sym_priv), symbol_filter) < 0) {
641 perror("failed to load kernel symbols");
646 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
647 MAP_SHARED, input, offset);
648 if (buf == MAP_FAILED) {
649 perror("failed to mmap file");
654 event = (event_t *)(buf + head);
656 size = event->header.size;
660 if (head + event->header.size >= page_size * mmap_window) {
661 unsigned long shift = page_size * (head / page_size);
664 munmap_ret = munmap(buf, page_size * mmap_window);
665 assert(munmap_ret == 0);
672 size = event->header.size;
674 dump_printf("%p [%p]: event: %d\n",
675 (void *)(offset + head),
676 (void *)(long)event->header.size,
679 if (!size || process_event(event, offset, head) < 0) {
681 dump_printf("%p [%p]: skipping unknown header type: %d\n",
682 (void *)(offset + head),
683 (void *)(long)(event->header.size),
689 * assume we lost track of the stream, check alignment, and
690 * increment a single u64 in the hope to catch on again 'soon'.
693 if (unlikely(head & 7))
701 if (offset + head < (unsigned long)input_stat.st_size)
707 dump_printf(" IP events: %10ld\n", total);
708 dump_printf(" mmap events: %10ld\n", total_mmap);
709 dump_printf(" comm events: %10ld\n", total_comm);
710 dump_printf(" fork events: %10ld\n", total_fork);
711 dump_printf(" unknown events: %10ld\n", total_unknown);
717 threads__fprintf(stdout);
720 dsos__fprintf(stdout);
723 output__resort(total);
730 static const char * const annotate_usage[] = {
731 "perf annotate [<options>] <command>",
735 static const struct option options[] = {
736 OPT_STRING('i', "input", &input_name, "file",
738 OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
739 "symbol to annotate"),
740 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
741 OPT_BOOLEAN('v', "verbose", &verbose,
742 "be more verbose (show symbol address, etc)"),
743 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
744 "dump raw trace in ASCII"),
745 OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
746 OPT_BOOLEAN('m', "modules", &modules,
747 "load module symbols - WARNING: use only with -k and LIVE kernel"),
748 OPT_BOOLEAN('l', "print-line", &print_line,
749 "print matching source lines (may be slow)"),
750 OPT_BOOLEAN('P', "full-paths", &full_paths,
751 "Don't shorten the displayed pathnames"),
755 static void setup_sorting(void)
757 char *tmp, *tok, *str = strdup(sort_order);
759 for (tok = strtok_r(str, ", ", &tmp);
760 tok; tok = strtok_r(NULL, ", ", &tmp)) {
761 if (sort_dimension__add(tok) < 0) {
762 error("Unknown --sort key: `%s'", tok);
763 usage_with_options(annotate_usage, options);
770 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
774 page_size = getpagesize();
776 argc = parse_options(argc, argv, options, annotate_usage, 0);
782 * Special case: if there's an argument left then assume tha
783 * it's a symbol filter:
786 usage_with_options(annotate_usage, options);
788 sym_hist_filter = argv[0];
793 if (field_sep && *field_sep == '.') {
794 fputs("'.' is the only non valid --field-separator argument\n",
799 return __cmd_annotate();