99bac6aa72c495837d455f2ce05d58991aa461d1
[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
33 static int              full_paths;
34
35 static int              print_line;
36
37 static unsigned long    page_size;
38 static unsigned long    mmap_window = 32;
39
40 struct sym_hist {
41         u64             sum;
42         u64             ip[0];
43 };
44
45 struct sym_ext {
46         struct rb_node  node;
47         double          percent;
48         char            *path;
49 };
50
51 struct sym_priv {
52         struct sym_hist *hist;
53         struct sym_ext  *ext;
54 };
55
56 static const char *sym_hist_filter;
57
58 static int symbol_filter(struct map *map, struct symbol *sym)
59 {
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));
65
66                 priv->hist = malloc(size);
67                 if (priv->hist)
68                         memset(priv->hist, 0, size);
69                 return 0;
70         }
71         /*
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.
76          */
77         return 0;
78 }
79
80 /*
81  * collect histogram counts
82  */
83 static void hist_hit(struct hist_entry *he, u64 ip)
84 {
85         unsigned int sym_size, offset;
86         struct symbol *sym = he->sym;
87         struct sym_priv *priv;
88         struct sym_hist *h;
89
90         he->count++;
91
92         if (!sym || !he->map)
93                 return;
94
95         priv = dso__sym_priv(he->map->dso, sym);
96         if (!priv->hist)
97                 return;
98
99         sym_size = sym->end - sym->start;
100         offset = ip - sym->start;
101
102         if (verbose)
103                 fprintf(stderr, "%s: ip=%Lx\n", __func__,
104                         he->map->unmap_ip(he->map, ip));
105
106         if (offset >= sym_size)
107                 return;
108
109         h = priv->hist;
110         h->sum++;
111         h->ip[offset]++;
112
113         if (verbose >= 3)
114                 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
115                         (void *)(unsigned long)he->sym->start,
116                         he->sym->name,
117                         (void *)(unsigned long)ip, ip - he->sym->start,
118                         h->ip[offset]);
119 }
120
121 static int hist_entry__add(struct thread *thread, struct map *map,
122                            struct symbol *sym, u64 ip, u64 count, char level)
123 {
124         bool hit;
125         struct hist_entry *he = __hist_entry__add(thread, map, sym, NULL, ip,
126                                                   count, level, &hit);
127         if (he == NULL)
128                 return -ENOMEM;
129         hist_hit(he, ip);
130         return 0;
131 }
132
133 static int
134 process_sample_event(event_t *event, unsigned long offset, unsigned long head)
135 {
136         char level;
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);
141
142         dump_printf("%p [%p]: PERF_EVENT (IP, %d): %d: %p\n",
143                 (void *)(offset + head),
144                 (void *)(long)(event->header.size),
145                 event->header.misc,
146                 event->ip.pid,
147                 (void *)(long)ip);
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         dump_printf(" ... thread: %s:%d\n", thread->comm, thread->pid);
156
157         if (event->header.misc & PERF_RECORD_MISC_KERNEL) {
158                 level = 'k';
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) {
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                 level = 'H';
189                 dump_printf(" ...... dso: [hypervisor]\n");
190         }
191
192         if (hist_entry__add(thread, map, sym, ip, 1, level)) {
193                 fprintf(stderr, "problem incrementing symbol count, "
194                                 "skipping event\n");
195                 return -1;
196         }
197         total++;
198
199         return 0;
200 }
201
202 static int
203 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
204 {
205         struct map *map = map__new(&event->mmap, NULL, 0,
206                                    sizeof(struct sym_priv), symbol_filter,
207                                    verbose);
208         struct thread *thread = threads__findnew(event->mmap.pid);
209
210         dump_printf("%p [%p]: PERF_RECORD_MMAP %d: [%p(%p) @ %p]: %s\n",
211                 (void *)(offset + head),
212                 (void *)(long)(event->header.size),
213                 event->mmap.pid,
214                 (void *)(long)event->mmap.start,
215                 (void *)(long)event->mmap.len,
216                 (void *)(long)event->mmap.pgoff,
217                 event->mmap.filename);
218
219         if (thread == NULL || map == NULL) {
220                 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
221                 return 0;
222         }
223
224         thread__insert_map(thread, map);
225         total_mmap++;
226
227         return 0;
228 }
229
230 static int
231 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
232 {
233         struct thread *thread = threads__findnew(event->comm.pid);
234
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);
239
240         if (thread == NULL ||
241             thread__set_comm(thread, event->comm.comm)) {
242                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
243                 return -1;
244         }
245         total_comm++;
246
247         return 0;
248 }
249
250 static int
251 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
252 {
253         struct thread *thread = threads__findnew(event->fork.pid);
254         struct thread *parent = threads__findnew(event->fork.ppid);
255
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);
260
261         /*
262          * A thread clone will have the same PID for both
263          * parent and child.
264          */
265         if (thread == parent)
266                 return 0;
267
268         if (!thread || !parent || thread__fork(thread, parent)) {
269                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
270                 return -1;
271         }
272         total_fork++;
273
274         return 0;
275 }
276
277 static int
278 process_event(event_t *event, unsigned long offset, unsigned long head)
279 {
280         switch (event->header.type) {
281         case PERF_RECORD_SAMPLE:
282                 return process_sample_event(event, offset, head);
283
284         case PERF_RECORD_MMAP:
285                 return process_mmap_event(event, offset, head);
286
287         case PERF_RECORD_COMM:
288                 return process_comm_event(event, offset, head);
289
290         case PERF_RECORD_FORK:
291                 return process_fork_event(event, offset, head);
292         /*
293          * We dont process them right now but they are fine:
294          */
295
296         case PERF_RECORD_THROTTLE:
297         case PERF_RECORD_UNTHROTTLE:
298                 return 0;
299
300         default:
301                 return -1;
302         }
303
304         return 0;
305 }
306
307 static int parse_line(FILE *file, struct hist_entry *he, u64 len)
308 {
309         struct symbol *sym = he->sym;
310         char *line = NULL, *tmp, *tmp2;
311         static const char *prev_line;
312         static const char *prev_color;
313         unsigned int offset;
314         size_t line_len;
315         u64 start;
316         s64 line_ip;
317         int ret;
318         char *c;
319
320         if (getline(&line, &line_len, file) < 0)
321                 return -1;
322         if (!line)
323                 return -1;
324
325         c = strchr(line, '\n');
326         if (c)
327                 *c = 0;
328
329         line_ip = -1;
330         offset = 0;
331         ret = -2;
332
333         /*
334          * Strip leading spaces:
335          */
336         tmp = line;
337         while (*tmp) {
338                 if (*tmp != ' ')
339                         break;
340                 tmp++;
341         }
342
343         if (*tmp) {
344                 /*
345                  * Parse hexa addresses followed by ':'
346                  */
347                 line_ip = strtoull(tmp, &tmp2, 16);
348                 if (*tmp2 != ':')
349                         line_ip = -1;
350         }
351
352         start = he->map->unmap_ip(he->map, sym->start);
353
354         if (line_ip != -1) {
355                 const char *path = NULL;
356                 unsigned int hits = 0;
357                 double percent = 0.0;
358                 const char *color;
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;
362
363                 offset = line_ip - start;
364                 if (offset < len)
365                         hits = h->ip[offset];
366
367                 if (offset < len && sym_ext) {
368                         path = sym_ext[offset].path;
369                         percent = sym_ext[offset].percent;
370                 } else if (h->sum)
371                         percent = 100.0 * hits / h->sum;
372
373                 color = get_percent_color(percent);
374
375                 /*
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
379                  */
380                 if (path) {
381                         if (!prev_line || strcmp(prev_line, path)
382                                        || color != prev_color) {
383                                 color_fprintf(stdout, color, " %s", path);
384                                 prev_line = path;
385                                 prev_color = color;
386                         }
387                 }
388
389                 color_fprintf(stdout, color, " %7.2f", percent);
390                 printf(" :      ");
391                 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
392         } else {
393                 if (!*line)
394                         printf("         :\n");
395                 else
396                         printf("         :      %s\n", line);
397         }
398
399         return 0;
400 }
401
402 static struct rb_root root_sym_ext;
403
404 static void insert_source_line(struct sym_ext *sym_ext)
405 {
406         struct sym_ext *iter;
407         struct rb_node **p = &root_sym_ext.rb_node;
408         struct rb_node *parent = NULL;
409
410         while (*p != NULL) {
411                 parent = *p;
412                 iter = rb_entry(parent, struct sym_ext, node);
413
414                 if (sym_ext->percent > iter->percent)
415                         p = &(*p)->rb_left;
416                 else
417                         p = &(*p)->rb_right;
418         }
419
420         rb_link_node(&sym_ext->node, parent, p);
421         rb_insert_color(&sym_ext->node, &root_sym_ext);
422 }
423
424 static void free_source_line(struct hist_entry *he, int len)
425 {
426         struct sym_priv *priv = dso__sym_priv(he->map->dso, he->sym);
427         struct sym_ext *sym_ext = priv->ext;
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         priv->ext = 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 hist_entry *he, int len, const char *filename)
444 {
445         struct symbol *sym = he->sym;
446         u64 start;
447         int i;
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;
452
453         if (!h->sum)
454                 return;
455
456         sym_ext = priv->ext = calloc(len, sizeof(struct sym_ext));
457         if (!priv->ext)
458                 return;
459
460         start = he->map->unmap_ip(he->map, sym->start);
461
462         for (i = 0; i < len; i++) {
463                 char *path = NULL;
464                 size_t line_len;
465                 u64 offset;
466                 FILE *fp;
467
468                 sym_ext[i].percent = 100.0 * h->ip[i] / h->sum;
469                 if (sym_ext[i].percent <= 0.5)
470                         continue;
471
472                 offset = start + i;
473                 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
474                 fp = popen(cmd, "r");
475                 if (!fp)
476                         continue;
477
478                 if (getline(&path, &line_len, fp) < 0 || !line_len)
479                         goto next;
480
481                 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
482                 if (!sym_ext[i].path)
483                         goto next;
484
485                 strcpy(sym_ext[i].path, path);
486                 insert_source_line(&sym_ext[i]);
487
488         next:
489                 pclose(fp);
490         }
491 }
492
493 static void print_summary(const char *filename)
494 {
495         struct sym_ext *sym_ext;
496         struct rb_node *node;
497
498         printf("\nSorted summary for file %s\n", filename);
499         printf("----------------------------------------------\n\n");
500
501         if (RB_EMPTY_ROOT(&root_sym_ext)) {
502                 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
503                 return;
504         }
505
506         node = rb_first(&root_sym_ext);
507         while (node) {
508                 double percent;
509                 const char *color;
510                 char *path;
511
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;
516
517                 color_fprintf(stdout, color, " %7.2f %s", percent, path);
518                 node = rb_next(node);
519         }
520 }
521
522 static void annotate_sym(struct hist_entry *he)
523 {
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;
528         u64 len;
529         char command[PATH_MAX*2];
530         FILE *file;
531
532         if (!filename)
533                 return;
534
535         if (verbose)
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));
540
541         if (full_paths)
542                 d_filename = filename;
543         else
544                 d_filename = basename(filename);
545
546         len = sym->end - sym->start;
547
548         if (print_line) {
549                 get_source_line(he, len, filename);
550                 print_summary(filename);
551         }
552
553         printf("\n\n------------------------------------------------\n");
554         printf(" Percent |      Source code & Disassembly of %s\n", d_filename);
555         printf("------------------------------------------------\n");
556
557         if (verbose >= 2)
558                 printf("annotating [%p] %30s : [%p] %30s\n",
559                        dso, dso->long_name, sym, sym->name);
560
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),
563                 filename, filename);
564
565         if (verbose >= 3)
566                 printf("doing: %s\n", command);
567
568         file = popen(command, "r");
569         if (!file)
570                 return;
571
572         while (!feof(file)) {
573                 if (parse_line(file, he, len) < 0)
574                         break;
575         }
576
577         pclose(file);
578         if (print_line)
579                 free_source_line(he, len);
580 }
581
582 static void find_annotations(void)
583 {
584         struct rb_node *nd;
585
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;
589
590                 if (he->sym == NULL)
591                         continue;
592
593                 priv = dso__sym_priv(he->map->dso, he->sym);
594                 if (priv->hist == NULL)
595                         continue;
596
597                 annotate_sym(he);
598                 /*
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.
601                  */
602                 free(priv->hist);
603                 priv->hist = NULL;
604         }
605 }
606
607 static int __cmd_annotate(void)
608 {
609         int ret, rc = EXIT_FAILURE;
610         unsigned long offset = 0;
611         unsigned long head = 0;
612         struct stat input_stat;
613         event_t *event;
614         uint32_t size;
615         char *buf;
616
617         register_idle_thread();
618
619         input = open(input_name, O_RDONLY);
620         if (input < 0) {
621                 perror("failed to open file");
622                 exit(-1);
623         }
624
625         ret = fstat(input, &input_stat);
626         if (ret < 0) {
627                 perror("failed to stat file");
628                 exit(-1);
629         }
630
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);
633                 exit(-1);
634         }
635
636         if (!input_stat.st_size) {
637                 fprintf(stderr, "zero-sized file, nothing to do!\n");
638                 exit(0);
639         }
640
641         if (load_kernel(sizeof(struct sym_priv), symbol_filter) < 0) {
642                 perror("failed to load kernel symbols");
643                 return EXIT_FAILURE;
644         }
645
646 remap:
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");
651                 exit(-1);
652         }
653
654 more:
655         event = (event_t *)(buf + head);
656
657         size = event->header.size;
658         if (!size)
659                 size = 8;
660
661         if (head + event->header.size >= page_size * mmap_window) {
662                 unsigned long shift = page_size * (head / page_size);
663                 int munmap_ret;
664
665                 munmap_ret = munmap(buf, page_size * mmap_window);
666                 assert(munmap_ret == 0);
667
668                 offset += shift;
669                 head -= shift;
670                 goto remap;
671         }
672
673         size = event->header.size;
674
675         dump_printf("%p [%p]: event: %d\n",
676                         (void *)(offset + head),
677                         (void *)(long)event->header.size,
678                         event->header.type);
679
680         if (!size || process_event(event, offset, head) < 0) {
681
682                 dump_printf("%p [%p]: skipping unknown header type: %d\n",
683                         (void *)(offset + head),
684                         (void *)(long)(event->header.size),
685                         event->header.type);
686
687                 total_unknown++;
688
689                 /*
690                  * assume we lost track of the stream, check alignment, and
691                  * increment a single u64 in the hope to catch on again 'soon'.
692                  */
693
694                 if (unlikely(head & 7))
695                         head &= ~7ULL;
696
697                 size = 8;
698         }
699
700         head += size;
701
702         if (offset + head < (unsigned long)input_stat.st_size)
703                 goto more;
704
705         rc = EXIT_SUCCESS;
706         close(input);
707
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);
713
714         if (dump_trace)
715                 return 0;
716
717         if (verbose > 3)
718                 threads__fprintf(stdout);
719
720         if (verbose > 2)
721                 dsos__fprintf(stdout);
722
723         collapse__resort();
724         output__resort(total);
725
726         find_annotations();
727
728         return rc;
729 }
730
731 static const char * const annotate_usage[] = {
732         "perf annotate [<options>] <command>",
733         NULL
734 };
735
736 static const struct option options[] = {
737         OPT_STRING('i', "input", &input_name, "file",
738                     "input file name"),
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"),
753         OPT_END()
754 };
755
756 static void setup_sorting(void)
757 {
758         char *tmp, *tok, *str = strdup(sort_order);
759
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);
765                 }
766         }
767
768         free(str);
769 }
770
771 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
772 {
773         symbol__init();
774
775         page_size = getpagesize();
776
777         argc = parse_options(argc, argv, options, annotate_usage, 0);
778
779         setup_sorting();
780
781         if (argc) {
782                 /*
783                  * Special case: if there's an argument left then assume tha
784                  * it's a symbol filter:
785                  */
786                 if (argc > 1)
787                         usage_with_options(annotate_usage, options);
788
789                 sym_hist_filter = argv[0];
790         }
791
792         setup_pager();
793
794         if (field_sep && *field_sep == '.') {
795                 fputs("'.' is the only non valid --field-separator argument\n",
796                                 stderr);
797                 exit(129);
798         }
799
800         return __cmd_annotate();
801 }