perf tools: Rewrite and improve support for kernel modules
[safe/jmp/linux-2.6] / tools / perf / builtin-annotate.c
1 /*
2  * builtin-annotate.c
3  *
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.
7  */
8 #include "builtin.h"
9
10 #include "util/util.h"
11
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"
18
19 #include "perf.h"
20 #include "util/debug.h"
21
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
28 static char             const *input_name = "perf.data";
29
30 static int              force;
31 static int              input;
32 static int              show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
33
34 static int              full_paths;
35
36 static int              print_line;
37
38 static unsigned long    page_size;
39 static unsigned long    mmap_window = 32;
40
41 static struct rb_root   threads;
42 static struct thread    *last_match;
43
44
45 struct sym_ext {
46         struct rb_node  node;
47         double          percent;
48         char            *path;
49 };
50
51
52 /*
53  * collect histogram counts
54  */
55 static void hist_hit(struct hist_entry *he, u64 ip)
56 {
57         unsigned int sym_size, offset;
58         struct symbol *sym = he->sym;
59
60         he->count++;
61
62         if (!sym || !sym->hist)
63                 return;
64
65         sym_size = sym->end - sym->start;
66         ip = he->map->map_ip(he->map, ip);
67         offset = ip - sym->start;
68
69         if (offset >= sym_size)
70                 return;
71
72         sym->hist_sum++;
73         sym->hist[offset]++;
74
75         if (verbose >= 3)
76                 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
77                         (void *)(unsigned long)he->sym->start,
78                         he->sym->name,
79                         (void *)(unsigned long)ip, ip - he->sym->start,
80                         sym->hist[offset]);
81 }
82
83 static int
84 hist_entry__add(struct thread *thread, struct map *map,
85                 struct symbol *sym, u64 ip, char level)
86 {
87         struct rb_node **p = &hist.rb_node;
88         struct rb_node *parent = NULL;
89         struct hist_entry *he;
90         struct hist_entry entry = {
91                 .thread = thread,
92                 .map    = map,
93                 .sym    = sym,
94                 .ip     = ip,
95                 .level  = level,
96                 .count  = 1,
97         };
98         int cmp;
99
100         while (*p != NULL) {
101                 parent = *p;
102                 he = rb_entry(parent, struct hist_entry, rb_node);
103
104                 cmp = hist_entry__cmp(&entry, he);
105
106                 if (!cmp) {
107                         hist_hit(he, ip);
108
109                         return 0;
110                 }
111
112                 if (cmp < 0)
113                         p = &(*p)->rb_left;
114                 else
115                         p = &(*p)->rb_right;
116         }
117
118         he = malloc(sizeof(*he));
119         if (!he)
120                 return -ENOMEM;
121         *he = entry;
122         rb_link_node(&he->rb_node, parent, p);
123         rb_insert_color(&he->rb_node, &hist);
124
125         return 0;
126 }
127
128 static int
129 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
130 {
131         char level;
132         int show = 0;
133         struct thread *thread;
134         u64 ip = event->ip.ip;
135         struct map *map = NULL;
136         struct symbol *sym = NULL;
137
138         thread = threads__findnew(event->ip.pid, &threads, &last_match);
139
140         dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
141                 (void *)(offset + head),
142                 (void *)(long)(event->header.size),
143                 event->header.misc,
144                 event->ip.pid,
145                 (void *)(long)ip);
146
147         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
148
149         if (thread == NULL) {
150                 fprintf(stderr, "problem processing %d event, skipping it.\n",
151                         event->header.type);
152                 return -1;
153         }
154
155         if (event->header.misc & PERF_RECORD_MISC_KERNEL) {
156                 show = SHOW_KERNEL;
157                 level = 'k';
158                 sym = kernel_maps__find_symbol(ip, &map);
159                 dump_printf(" ...... dso: %s\n",
160                             map ? map->dso->long_name : "<not found>");
161         } else if (event->header.misc & PERF_RECORD_MISC_USER) {
162                 show = SHOW_USER;
163                 level = '.';
164                 map = thread__find_map(thread, ip);
165                 if (map != NULL) {
166 got_map:
167                         ip = map->map_ip(map, ip);
168                         sym = map->dso->find_symbol(map->dso, ip);
169                 } else {
170                         /*
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).
175                          *
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.
179                          */
180                         if ((long long)ip < 0) {
181                                 map = kernel_map;
182                                 goto got_map;
183                         }
184                 }
185                 dump_printf(" ...... dso: %s\n",
186                             map ? map->dso->long_name : "<not found>");
187         } else {
188                 show = SHOW_HV;
189                 level = 'H';
190                 dump_printf(" ...... dso: [hypervisor]\n");
191         }
192
193         if (show & show_mask) {
194                 if (hist_entry__add(thread, map, sym, ip, level)) {
195                         fprintf(stderr,
196                 "problem incrementing symbol count, skipping event\n");
197                         return -1;
198                 }
199         }
200         total++;
201
202         return 0;
203 }
204
205 static int
206 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
207 {
208         struct thread *thread;
209         struct map *map = map__new(&event->mmap, NULL, 0);
210
211         thread = threads__findnew(event->mmap.pid, &threads, &last_match);
212
213         dump_printf("%p [%p]: PERF_RECORD_MMAP %d: [%p(%p) @ %p]: %s\n",
214                 (void *)(offset + head),
215                 (void *)(long)(event->header.size),
216                 event->mmap.pid,
217                 (void *)(long)event->mmap.start,
218                 (void *)(long)event->mmap.len,
219                 (void *)(long)event->mmap.pgoff,
220                 event->mmap.filename);
221
222         if (thread == NULL || map == NULL) {
223                 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
224                 return 0;
225         }
226
227         thread__insert_map(thread, map);
228         total_mmap++;
229
230         return 0;
231 }
232
233 static int
234 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
235 {
236         struct thread *thread;
237
238         thread = threads__findnew(event->comm.pid, &threads, &last_match);
239         dump_printf("%p [%p]: PERF_RECORD_COMM: %s:%d\n",
240                 (void *)(offset + head),
241                 (void *)(long)(event->header.size),
242                 event->comm.comm, event->comm.pid);
243
244         if (thread == NULL ||
245             thread__set_comm(thread, event->comm.comm)) {
246                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
247                 return -1;
248         }
249         total_comm++;
250
251         return 0;
252 }
253
254 static int
255 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
256 {
257         struct thread *thread;
258         struct thread *parent;
259
260         thread = threads__findnew(event->fork.pid, &threads, &last_match);
261         parent = threads__findnew(event->fork.ppid, &threads, &last_match);
262         dump_printf("%p [%p]: PERF_RECORD_FORK: %d:%d\n",
263                 (void *)(offset + head),
264                 (void *)(long)(event->header.size),
265                 event->fork.pid, event->fork.ppid);
266
267         /*
268          * A thread clone will have the same PID for both
269          * parent and child.
270          */
271         if (thread == parent)
272                 return 0;
273
274         if (!thread || !parent || thread__fork(thread, parent)) {
275                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
276                 return -1;
277         }
278         total_fork++;
279
280         return 0;
281 }
282
283 static int
284 process_event(event_t *event, unsigned long offset, unsigned long head)
285 {
286         switch (event->header.type) {
287         case PERF_RECORD_SAMPLE:
288                 return process_sample_event(event, offset, head);
289
290         case PERF_RECORD_MMAP:
291                 return process_mmap_event(event, offset, head);
292
293         case PERF_RECORD_COMM:
294                 return process_comm_event(event, offset, head);
295
296         case PERF_RECORD_FORK:
297                 return process_fork_event(event, offset, head);
298         /*
299          * We dont process them right now but they are fine:
300          */
301
302         case PERF_RECORD_THROTTLE:
303         case PERF_RECORD_UNTHROTTLE:
304                 return 0;
305
306         default:
307                 return -1;
308         }
309
310         return 0;
311 }
312
313 static int
314 parse_line(FILE *file, struct symbol *sym, u64 len)
315 {
316         char *line = NULL, *tmp, *tmp2;
317         static const char *prev_line;
318         static const char *prev_color;
319         unsigned int offset;
320         size_t line_len;
321         s64 line_ip;
322         int ret;
323         char *c;
324
325         if (getline(&line, &line_len, file) < 0)
326                 return -1;
327         if (!line)
328                 return -1;
329
330         c = strchr(line, '\n');
331         if (c)
332                 *c = 0;
333
334         line_ip = -1;
335         offset = 0;
336         ret = -2;
337
338         /*
339          * Strip leading spaces:
340          */
341         tmp = line;
342         while (*tmp) {
343                 if (*tmp != ' ')
344                         break;
345                 tmp++;
346         }
347
348         if (*tmp) {
349                 /*
350                  * Parse hexa addresses followed by ':'
351                  */
352                 line_ip = strtoull(tmp, &tmp2, 16);
353                 if (*tmp2 != ':')
354                         line_ip = -1;
355         }
356
357         if (line_ip != -1) {
358                 const char *path = NULL;
359                 unsigned int hits = 0;
360                 double percent = 0.0;
361                 const char *color;
362                 struct sym_ext *sym_ext = sym->priv;
363
364                 offset = line_ip - sym->start;
365                 if (offset < len)
366                         hits = sym->hist[offset];
367
368                 if (offset < len && sym_ext) {
369                         path = sym_ext[offset].path;
370                         percent = sym_ext[offset].percent;
371                 } else if (sym->hist_sum)
372                         percent = 100.0 * hits / sym->hist_sum;
373
374                 color = get_percent_color(percent);
375
376                 /*
377                  * Also color the filename and line if needed, with
378                  * the same color than the percentage. Don't print it
379                  * twice for close colored ip with the same filename:line
380                  */
381                 if (path) {
382                         if (!prev_line || strcmp(prev_line, path)
383                                        || color != prev_color) {
384                                 color_fprintf(stdout, color, " %s", path);
385                                 prev_line = path;
386                                 prev_color = color;
387                         }
388                 }
389
390                 color_fprintf(stdout, color, " %7.2f", percent);
391                 printf(" :      ");
392                 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
393         } else {
394                 if (!*line)
395                         printf("         :\n");
396                 else
397                         printf("         :      %s\n", line);
398         }
399
400         return 0;
401 }
402
403 static struct rb_root root_sym_ext;
404
405 static void insert_source_line(struct sym_ext *sym_ext)
406 {
407         struct sym_ext *iter;
408         struct rb_node **p = &root_sym_ext.rb_node;
409         struct rb_node *parent = NULL;
410
411         while (*p != NULL) {
412                 parent = *p;
413                 iter = rb_entry(parent, struct sym_ext, node);
414
415                 if (sym_ext->percent > iter->percent)
416                         p = &(*p)->rb_left;
417                 else
418                         p = &(*p)->rb_right;
419         }
420
421         rb_link_node(&sym_ext->node, parent, p);
422         rb_insert_color(&sym_ext->node, &root_sym_ext);
423 }
424
425 static void free_source_line(struct symbol *sym, int len)
426 {
427         struct sym_ext *sym_ext = sym->priv;
428         int i;
429
430         if (!sym_ext)
431                 return;
432
433         for (i = 0; i < len; i++)
434                 free(sym_ext[i].path);
435         free(sym_ext);
436
437         sym->priv = NULL;
438         root_sym_ext = RB_ROOT;
439 }
440
441 /* Get the filename:line for the colored entries */
442 static void
443 get_source_line(struct symbol *sym, int len, const char *filename)
444 {
445         int i;
446         char cmd[PATH_MAX * 2];
447         struct sym_ext *sym_ext;
448
449         if (!sym->hist_sum)
450                 return;
451
452         sym->priv = calloc(len, sizeof(struct sym_ext));
453         if (!sym->priv)
454                 return;
455
456         sym_ext = sym->priv;
457
458         for (i = 0; i < len; i++) {
459                 char *path = NULL;
460                 size_t line_len;
461                 u64 offset;
462                 FILE *fp;
463
464                 sym_ext[i].percent = 100.0 * sym->hist[i] / sym->hist_sum;
465                 if (sym_ext[i].percent <= 0.5)
466                         continue;
467
468                 offset = sym->start + i;
469                 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
470                 fp = popen(cmd, "r");
471                 if (!fp)
472                         continue;
473
474                 if (getline(&path, &line_len, fp) < 0 || !line_len)
475                         goto next;
476
477                 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
478                 if (!sym_ext[i].path)
479                         goto next;
480
481                 strcpy(sym_ext[i].path, path);
482                 insert_source_line(&sym_ext[i]);
483
484         next:
485                 pclose(fp);
486         }
487 }
488
489 static void print_summary(const char *filename)
490 {
491         struct sym_ext *sym_ext;
492         struct rb_node *node;
493
494         printf("\nSorted summary for file %s\n", filename);
495         printf("----------------------------------------------\n\n");
496
497         if (RB_EMPTY_ROOT(&root_sym_ext)) {
498                 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
499                 return;
500         }
501
502         node = rb_first(&root_sym_ext);
503         while (node) {
504                 double percent;
505                 const char *color;
506                 char *path;
507
508                 sym_ext = rb_entry(node, struct sym_ext, node);
509                 percent = sym_ext->percent;
510                 color = get_percent_color(percent);
511                 path = sym_ext->path;
512
513                 color_fprintf(stdout, color, " %7.2f %s", percent, path);
514                 node = rb_next(node);
515         }
516 }
517
518 static void annotate_sym(struct dso *dso, struct symbol *sym)
519 {
520         const char *filename = dso->long_name, *d_filename;
521         u64 len;
522         char command[PATH_MAX*2];
523         FILE *file;
524
525         if (!filename)
526                 return;
527
528         if (full_paths)
529                 d_filename = filename;
530         else
531                 d_filename = basename(filename);
532
533         len = sym->end - sym->start;
534
535         if (print_line) {
536                 get_source_line(sym, len, filename);
537                 print_summary(filename);
538         }
539
540         printf("\n\n------------------------------------------------\n");
541         printf(" Percent |      Source code & Disassembly of %s\n", d_filename);
542         printf("------------------------------------------------\n");
543
544         if (verbose >= 2)
545                 printf("annotating [%p] %30s : [%p] %30s\n",
546                        dso, dso->long_name, sym, sym->name);
547
548         sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
549                 sym->start, sym->end, filename, filename);
550
551         if (verbose >= 3)
552                 printf("doing: %s\n", command);
553
554         file = popen(command, "r");
555         if (!file)
556                 return;
557
558         while (!feof(file)) {
559                 if (parse_line(file, sym, len) < 0)
560                         break;
561         }
562
563         pclose(file);
564         if (print_line)
565                 free_source_line(sym, len);
566 }
567
568 static void find_annotations(void)
569 {
570         struct rb_node *nd;
571         struct dso *dso;
572         int count = 0;
573
574         list_for_each_entry(dso, &dsos, node) {
575
576                 for (nd = rb_first(&dso->syms); nd; nd = rb_next(nd)) {
577                         struct symbol *sym = rb_entry(nd, struct symbol, rb_node);
578
579                         if (sym->hist) {
580                                 annotate_sym(dso, sym);
581                                 count++;
582                         }
583                 }
584         }
585
586         if (!count)
587                 printf(" Error: symbol '%s' not present amongst the samples.\n", sym_hist_filter);
588 }
589
590 static int __cmd_annotate(void)
591 {
592         int ret, rc = EXIT_FAILURE;
593         unsigned long offset = 0;
594         unsigned long head = 0;
595         struct stat input_stat;
596         event_t *event;
597         uint32_t size;
598         char *buf;
599
600         register_idle_thread(&threads, &last_match);
601
602         input = open(input_name, O_RDONLY);
603         if (input < 0) {
604                 perror("failed to open file");
605                 exit(-1);
606         }
607
608         ret = fstat(input, &input_stat);
609         if (ret < 0) {
610                 perror("failed to stat file");
611                 exit(-1);
612         }
613
614         if (!force && input_stat.st_uid && (input_stat.st_uid != geteuid())) {
615                 fprintf(stderr, "file: %s not owned by current user or root\n", input_name);
616                 exit(-1);
617         }
618
619         if (!input_stat.st_size) {
620                 fprintf(stderr, "zero-sized file, nothing to do!\n");
621                 exit(0);
622         }
623
624         if (load_kernel() < 0) {
625                 perror("failed to load kernel symbols");
626                 return EXIT_FAILURE;
627         }
628
629 remap:
630         buf = (char *)mmap(NULL, page_size * mmap_window, PROT_READ,
631                            MAP_SHARED, input, offset);
632         if (buf == MAP_FAILED) {
633                 perror("failed to mmap file");
634                 exit(-1);
635         }
636
637 more:
638         event = (event_t *)(buf + head);
639
640         size = event->header.size;
641         if (!size)
642                 size = 8;
643
644         if (head + event->header.size >= page_size * mmap_window) {
645                 unsigned long shift = page_size * (head / page_size);
646                 int munmap_ret;
647
648                 munmap_ret = munmap(buf, page_size * mmap_window);
649                 assert(munmap_ret == 0);
650
651                 offset += shift;
652                 head -= shift;
653                 goto remap;
654         }
655
656         size = event->header.size;
657
658         dump_printf("%p [%p]: event: %d\n",
659                         (void *)(offset + head),
660                         (void *)(long)event->header.size,
661                         event->header.type);
662
663         if (!size || process_event(event, offset, head) < 0) {
664
665                 dump_printf("%p [%p]: skipping unknown header type: %d\n",
666                         (void *)(offset + head),
667                         (void *)(long)(event->header.size),
668                         event->header.type);
669
670                 total_unknown++;
671
672                 /*
673                  * assume we lost track of the stream, check alignment, and
674                  * increment a single u64 in the hope to catch on again 'soon'.
675                  */
676
677                 if (unlikely(head & 7))
678                         head &= ~7ULL;
679
680                 size = 8;
681         }
682
683         head += size;
684
685         if (offset + head < (unsigned long)input_stat.st_size)
686                 goto more;
687
688         rc = EXIT_SUCCESS;
689         close(input);
690
691         dump_printf("      IP events: %10ld\n", total);
692         dump_printf("    mmap events: %10ld\n", total_mmap);
693         dump_printf("    comm events: %10ld\n", total_comm);
694         dump_printf("    fork events: %10ld\n", total_fork);
695         dump_printf(" unknown events: %10ld\n", total_unknown);
696
697         if (dump_trace)
698                 return 0;
699
700         if (verbose >= 3)
701                 threads__fprintf(stdout, &threads);
702
703         if (verbose >= 2)
704                 dsos__fprintf(stdout);
705
706         collapse__resort();
707         output__resort(total);
708
709         find_annotations();
710
711         return rc;
712 }
713
714 static const char * const annotate_usage[] = {
715         "perf annotate [<options>] <command>",
716         NULL
717 };
718
719 static const struct option options[] = {
720         OPT_STRING('i', "input", &input_name, "file",
721                     "input file name"),
722         OPT_STRING('s', "symbol", &sym_hist_filter, "symbol",
723                     "symbol to annotate"),
724         OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
725         OPT_BOOLEAN('v', "verbose", &verbose,
726                     "be more verbose (show symbol address, etc)"),
727         OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
728                     "dump raw trace in ASCII"),
729         OPT_STRING('k', "vmlinux", &vmlinux_name, "file", "vmlinux pathname"),
730         OPT_BOOLEAN('m', "modules", &modules,
731                     "load module symbols - WARNING: use only with -k and LIVE kernel"),
732         OPT_BOOLEAN('l', "print-line", &print_line,
733                     "print matching source lines (may be slow)"),
734         OPT_BOOLEAN('P', "full-paths", &full_paths,
735                     "Don't shorten the displayed pathnames"),
736         OPT_END()
737 };
738
739 static void setup_sorting(void)
740 {
741         char *tmp, *tok, *str = strdup(sort_order);
742
743         for (tok = strtok_r(str, ", ", &tmp);
744                         tok; tok = strtok_r(NULL, ", ", &tmp)) {
745                 if (sort_dimension__add(tok) < 0) {
746                         error("Unknown --sort key: `%s'", tok);
747                         usage_with_options(annotate_usage, options);
748                 }
749         }
750
751         free(str);
752 }
753
754 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
755 {
756         symbol__init();
757
758         page_size = getpagesize();
759
760         argc = parse_options(argc, argv, options, annotate_usage, 0);
761
762         setup_sorting();
763
764         if (argc) {
765                 /*
766                  * Special case: if there's an argument left then assume tha
767                  * it's a symbol filter:
768                  */
769                 if (argc > 1)
770                         usage_with_options(annotate_usage, options);
771
772                 sym_hist_filter = argv[0];
773         }
774
775         if (!sym_hist_filter)
776                 usage_with_options(annotate_usage, options);
777
778         setup_pager();
779
780         if (field_sep && *field_sep == '.') {
781                 fputs("'.' is the only non valid --field-separator argument\n",
782                                 stderr);
783                 exit(129);
784         }
785
786         return __cmd_annotate();
787 }