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";
32 static int show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
34 static int full_paths;
36 static int print_line;
38 static unsigned long page_size;
39 static unsigned long mmap_window = 32;
41 static struct rb_root threads;
42 static struct thread *last_match;
53 * collect histogram counts
55 static void hist_hit(struct hist_entry *he, u64 ip)
57 unsigned int sym_size, offset;
58 struct symbol *sym = he->sym;
62 if (!sym || !sym->hist)
65 sym_size = sym->end - sym->start;
66 offset = ip - sym->start;
68 if (offset >= sym_size)
75 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
76 (void *)(unsigned long)he->sym->start,
78 (void *)(unsigned long)ip, ip - he->sym->start,
83 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
84 struct symbol *sym, u64 ip, char level)
86 struct rb_node **p = &hist.rb_node;
87 struct rb_node *parent = NULL;
88 struct hist_entry *he;
89 struct hist_entry entry = {
102 he = rb_entry(parent, struct hist_entry, rb_node);
104 cmp = hist_entry__cmp(&entry, he);
118 he = malloc(sizeof(*he));
122 rb_link_node(&he->rb_node, parent, p);
123 rb_insert_color(&he->rb_node, &hist);
129 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
133 struct dso *dso = NULL;
134 struct thread *thread;
135 u64 ip = event->ip.ip;
136 struct map *map = NULL;
138 thread = threads__findnew(event->ip.pid, &threads, &last_match);
140 dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
141 (void *)(offset + head),
142 (void *)(long)(event->header.size),
147 dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
149 if (thread == NULL) {
150 fprintf(stderr, "problem processing %d event, skipping it.\n",
155 if (event->header.misc & PERF_RECORD_MISC_KERNEL) {
161 dump_printf(" ...... dso: %s\n", dso->name);
163 } else if (event->header.misc & PERF_RECORD_MISC_USER) {
168 map = thread__find_map(thread, ip);
170 ip = map->map_ip(map, ip);
174 * If this is outside of all known maps,
175 * and is a negative address, try to look it
176 * up in the kernel dso, as it might be a
177 * vsyscall (which executes in user-mode):
179 if ((long long)ip < 0)
182 dump_printf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
187 dump_printf(" ...... dso: [hypervisor]\n");
190 if (show & show_mask) {
191 struct symbol *sym = NULL;
194 sym = dso->find_symbol(dso, ip);
196 if (hist_entry__add(thread, map, dso, sym, ip, level)) {
198 "problem incrementing symbol count, skipping event\n");
208 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
210 struct thread *thread;
211 struct map *map = map__new(&event->mmap, NULL, 0);
213 thread = threads__findnew(event->mmap.pid, &threads, &last_match);
215 dump_printf("%p [%p]: PERF_RECORD_MMAP %d: [%p(%p) @ %p]: %s\n",
216 (void *)(offset + head),
217 (void *)(long)(event->header.size),
219 (void *)(long)event->mmap.start,
220 (void *)(long)event->mmap.len,
221 (void *)(long)event->mmap.pgoff,
222 event->mmap.filename);
224 if (thread == NULL || map == NULL) {
225 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
229 thread__insert_map(thread, map);
236 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
238 struct thread *thread;
240 thread = threads__findnew(event->comm.pid, &threads, &last_match);
241 dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
242 (void *)(offset + head),
243 (void *)(long)(event->header.size),
244 event->comm.comm, event->comm.pid);
246 if (thread == NULL ||
247 thread__set_comm(thread, event->comm.comm)) {
248 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
257 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
259 struct thread *thread;
260 struct thread *parent;
262 thread = threads__findnew(event->fork.pid, &threads, &last_match);
263 parent = threads__findnew(event->fork.ppid, &threads, &last_match);
264 dump_printf("%p [%p]: PERF_RECORD_FORK: %d:%d\n",
265 (void *)(offset + head),
266 (void *)(long)(event->header.size),
267 event->fork.pid, event->fork.ppid);
270 * A thread clone will have the same PID for both
273 if (thread == parent)
276 if (!thread || !parent || thread__fork(thread, parent)) {
277 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
286 process_event(event_t *event, unsigned long offset, unsigned long head)
288 switch (event->header.type) {
289 case PERF_RECORD_SAMPLE:
290 return process_sample_event(event, offset, head);
292 case PERF_RECORD_MMAP:
293 return process_mmap_event(event, offset, head);
295 case PERF_RECORD_COMM:
296 return process_comm_event(event, offset, head);
298 case PERF_RECORD_FORK:
299 return process_fork_event(event, offset, head);
301 * We dont process them right now but they are fine:
304 case PERF_RECORD_THROTTLE:
305 case PERF_RECORD_UNTHROTTLE:
316 parse_line(FILE *file, struct symbol *sym, u64 start, u64 len)
318 char *line = NULL, *tmp, *tmp2;
319 static const char *prev_line;
320 static const char *prev_color;
327 if (getline(&line, &line_len, file) < 0)
332 c = strchr(line, '\n');
341 * Strip leading spaces:
352 * Parse hexa addresses followed by ':'
354 line_ip = strtoull(tmp, &tmp2, 16);
360 const char *path = NULL;
361 unsigned int hits = 0;
362 double percent = 0.0;
364 struct sym_ext *sym_ext = sym->priv;
366 offset = line_ip - start;
368 hits = sym->hist[offset];
370 if (offset < len && sym_ext) {
371 path = sym_ext[offset].path;
372 percent = sym_ext[offset].percent;
373 } else if (sym->hist_sum)
374 percent = 100.0 * hits / sym->hist_sum;
376 color = get_percent_color(percent);
379 * Also color the filename and line if needed, with
380 * the same color than the percentage. Don't print it
381 * twice for close colored ip with the same filename:line
384 if (!prev_line || strcmp(prev_line, path)
385 || color != prev_color) {
386 color_fprintf(stdout, color, " %s", path);
392 color_fprintf(stdout, color, " %7.2f", percent);
394 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
399 printf(" : %s\n", line);
405 static struct rb_root root_sym_ext;
407 static void insert_source_line(struct sym_ext *sym_ext)
409 struct sym_ext *iter;
410 struct rb_node **p = &root_sym_ext.rb_node;
411 struct rb_node *parent = NULL;
415 iter = rb_entry(parent, struct sym_ext, node);
417 if (sym_ext->percent > iter->percent)
423 rb_link_node(&sym_ext->node, parent, p);
424 rb_insert_color(&sym_ext->node, &root_sym_ext);
427 static void free_source_line(struct symbol *sym, int len)
429 struct sym_ext *sym_ext = sym->priv;
435 for (i = 0; i < len; i++)
436 free(sym_ext[i].path);
440 root_sym_ext = RB_ROOT;
443 /* Get the filename:line for the colored entries */
445 get_source_line(struct symbol *sym, u64 start, int len, const char *filename)
448 char cmd[PATH_MAX * 2];
449 struct sym_ext *sym_ext;
454 sym->priv = calloc(len, sizeof(struct sym_ext));
460 for (i = 0; i < len; i++) {
466 sym_ext[i].percent = 100.0 * sym->hist[i] / sym->hist_sum;
467 if (sym_ext[i].percent <= 0.5)
471 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
472 fp = popen(cmd, "r");
476 if (getline(&path, &line_len, fp) < 0 || !line_len)
479 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
480 if (!sym_ext[i].path)
483 strcpy(sym_ext[i].path, path);
484 insert_source_line(&sym_ext[i]);
491 static void print_summary(const char *filename)
493 struct sym_ext *sym_ext;
494 struct rb_node *node;
496 printf("\nSorted summary for file %s\n", filename);
497 printf("----------------------------------------------\n\n");
499 if (RB_EMPTY_ROOT(&root_sym_ext)) {
500 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
504 node = rb_first(&root_sym_ext);
510 sym_ext = rb_entry(node, struct sym_ext, node);
511 percent = sym_ext->percent;
512 color = get_percent_color(percent);
513 path = sym_ext->path;
515 color_fprintf(stdout, color, " %7.2f %s", percent, path);
516 node = rb_next(node);
520 static void annotate_sym(struct dso *dso, struct symbol *sym)
522 const char *filename = dso->name, *d_filename;
524 char command[PATH_MAX*2];
530 filename = sym->module->path;
531 else if (dso == kernel_dso)
532 filename = vmlinux_name;
534 start = sym->obj_start;
538 d_filename = filename;
540 d_filename = basename(filename);
542 end = start + sym->end - sym->start + 1;
543 len = sym->end - sym->start;
546 get_source_line(sym, start, len, filename);
547 print_summary(filename);
550 printf("\n\n------------------------------------------------\n");
551 printf(" Percent | Source code & Disassembly of %s\n", d_filename);
552 printf("------------------------------------------------\n");
555 printf("annotating [%p] %30s : [%p] %30s\n", dso, dso->name, sym, sym->name);
557 sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
558 (u64)start, (u64)end, filename, filename);
561 printf("doing: %s\n", command);
563 file = popen(command, "r");
567 while (!feof(file)) {
568 if (parse_line(file, sym, start, len) < 0)
574 free_source_line(sym, len);
577 static void find_annotations(void)
583 list_for_each_entry(dso, &dsos, node) {
585 for (nd = rb_first(&dso->syms); nd; nd = rb_next(nd)) {
586 struct symbol *sym = rb_entry(nd, struct symbol, rb_node);
589 annotate_sym(dso, sym);
596 printf(" Error: symbol '%s' not present amongst the samples.\n", sym_hist_filter);
599 static int __cmd_annotate(void)
601 int ret, rc = EXIT_FAILURE;
602 unsigned long offset = 0;
603 unsigned long head = 0;
604 struct stat input_stat;
609 register_idle_thread(&threads, &last_match);
611 input = open(input_name, O_RDONLY);
613 perror("failed to open file");
617 ret = fstat(input, &input_stat);
619 perror("failed to stat file");
623 if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
624 fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
628 if (!input_stat.st_size) {
629 fprintf(stderr, "zero-sized file, nothing to do!\n");
633 if (load_kernel() < 0) {
634 perror("failed to load kernel symbols");
639 buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
640 MAP_SHARED, input, offset);
641 if (buf == MAP_FAILED) {
642 perror("failed to mmap file");
647 event = (event_t *)(buf + head);
649 size = event->header.size;
653 if (head + event->header.size >= page_size * mmap_window) {
654 unsigned long shift = page_size * (head / page_size);
657 munmap_ret = munmap(buf, page_size * mmap_window);
658 assert(munmap_ret == 0);
665 size = event->header.size;
667 dump_printf("%p [%p]: event: %d\n",
668 (void *)(offset + head),
669 (void *)(long)event->header.size,
672 if (!size || process_event(event, offset, head) < 0) {
674 dump_printf("%p [%p]: skipping unknown header type: %d\n",
675 (void *)(offset + head),
676 (void *)(long)(event->header.size),
682 * assume we lost track of the stream, check alignment, and
683 * increment a single u64 in the hope to catch on again 'soon'.
686 if (unlikely(head & 7))
694 if (offset + head < (unsigned long)input_stat.st_size)
700 dump_printf(" IP events: %10ld\n", total);
701 dump_printf(" mmap events: %10ld\n", total_mmap);
702 dump_printf(" comm events: %10ld\n", total_comm);
703 dump_printf(" fork events: %10ld\n", total_fork);
704 dump_printf(" unknown events: %10ld\n", total_unknown);
710 threads__fprintf(stdout, &threads);
713 dsos__fprintf(stdout);
716 output__resort(total);
723 static const char * const annotate_usage[] = {
724 "perf annotate [<options>] <command>",
728 static const struct option options[] = {
729 OPT_STRING('i', "input", &input_name, "file",
731 OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
732 "symbol to annotate"),
733 OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
734 OPT_BOOLEAN('v', "verbose", &verbose,
735 "be more verbose (show symbol address, etc)"),
736 OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
737 "dump raw trace in ASCII"),
738 OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
739 OPT_BOOLEAN('m', "modules", &modules,
740 "load module symbols - WARNING: use only with -k and LIVE kernel"),
741 OPT_BOOLEAN('l', "print-line", &print_line,
742 "print matching source lines (may be slow)"),
743 OPT_BOOLEAN('P', "full-paths", &full_paths,
744 "Don't shorten the displayed pathnames"),
748 static void setup_sorting(void)
750 char *tmp, *tok, *str = strdup(sort_order);
752 for (tok = strtok_r(str, ", ", &tmp);
753 tok; tok = strtok_r(NULL, ", ", &tmp)) {
754 if (sort_dimension__add(tok) < 0) {
755 error("Unknown --sort key: `%s'", tok);
756 usage_with_options(annotate_usage, options);
763 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
767 page_size = getpagesize();
769 argc = parse_options(argc, argv, options, annotate_usage, 0);
775 * Special case: if there's an argument left then assume tha
776 * it's a symbol filter:
779 usage_with_options(annotate_usage, options);
781 sym_hist_filter = argv[0];
784 if (!sym_hist_filter)
785 usage_with_options(annotate_usage, options);
789 if (field_sep && *field_sep == '.') {
790 fputs("'.' is the only non valid --field-separator argument\n",
795 return __cmd_annotate();