perf tools: Put common histogram functions in their own file
[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         offset = ip - sym->start;
67
68         if (offset >= sym_size)
69                 return;
70
71         sym->hist_sum++;
72         sym->hist[offset]++;
73
74         if (verbose >= 3)
75                 printf("%p %s: count++ [ip: %p, %08Lx] => %Ld\n",
76                         (void *)(unsigned long)he->sym->start,
77                         he->sym->name,
78                         (void *)(unsigned long)ip, ip - he->sym->start,
79                         sym->hist[offset]);
80 }
81
82 static int
83 hist_entry__add(struct thread *thread, struct map *map, struct dso *dso,
84                 struct symbol *sym, u64 ip, char level)
85 {
86         struct rb_node **p = &hist.rb_node;
87         struct rb_node *parent = NULL;
88         struct hist_entry *he;
89         struct hist_entry entry = {
90                 .thread = thread,
91                 .map    = map,
92                 .dso    = dso,
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 dso *dso = NULL;
134         struct thread *thread;
135         u64 ip = event->ip.ip;
136         struct map *map = 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
159                 dso = kernel_dso;
160
161                 dump_printf(" ...... dso: %s\n", dso->name);
162
163         } else if (event->header.misc & PERF_RECORD_MISC_USER) {
164
165                 show = SHOW_USER;
166                 level = '.';
167
168                 map = thread__find_map(thread, ip);
169                 if (map != NULL) {
170                         ip = map->map_ip(map, ip);
171                         dso = map->dso;
172                 } else {
173                         /*
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):
178                          */
179                         if ((long long)ip < 0)
180                                 dso = kernel_dso;
181                 }
182                 dump_printf(" ...... dso: %s\n", dso ? dso->name : "<not found>");
183
184         } else {
185                 show = SHOW_HV;
186                 level = 'H';
187                 dump_printf(" ...... dso: [hypervisor]\n");
188         }
189
190         if (show & show_mask) {
191                 struct symbol *sym = NULL;
192
193                 if (dso)
194                         sym = dso->find_symbol(dso, ip);
195
196                 if (hist_entry__add(thread, map, dso, sym, ip, level)) {
197                         fprintf(stderr,
198                 "problem incrementing symbol count, skipping event\n");
199                         return -1;
200                 }
201         }
202         total++;
203
204         return 0;
205 }
206
207 static int
208 process_mmap_event(event_t *event, unsigned long offset, unsigned long head)
209 {
210         struct thread *thread;
211         struct map *map = map__new(&event->mmap, NULL, 0);
212
213         thread = threads__findnew(event->mmap.pid, &threads, &last_match);
214
215         dump_printf("%p [%p]: PERF_RECORD_MMAP %d: [%p(%p) @ %p]: %s\n",
216                 (void *)(offset + head),
217                 (void *)(long)(event->header.size),
218                 event->mmap.pid,
219                 (void *)(long)event->mmap.start,
220                 (void *)(long)event->mmap.len,
221                 (void *)(long)event->mmap.pgoff,
222                 event->mmap.filename);
223
224         if (thread == NULL || map == NULL) {
225                 dump_printf("problem processing PERF_RECORD_MMAP, skipping event.\n");
226                 return 0;
227         }
228
229         thread__insert_map(thread, map);
230         total_mmap++;
231
232         return 0;
233 }
234
235 static int
236 process_comm_event(event_t *event, unsigned long offset, unsigned long head)
237 {
238         struct thread *thread;
239
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);
245
246         if (thread == NULL ||
247             thread__set_comm(thread, event->comm.comm)) {
248                 dump_printf("problem processing PERF_RECORD_COMM, skipping event.\n");
249                 return -1;
250         }
251         total_comm++;
252
253         return 0;
254 }
255
256 static int
257 process_fork_event(event_t *event, unsigned long offset, unsigned long head)
258 {
259         struct thread *thread;
260         struct thread *parent;
261
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);
268
269         /*
270          * A thread clone will have the same PID for both
271          * parent and child.
272          */
273         if (thread == parent)
274                 return 0;
275
276         if (!thread || !parent || thread__fork(thread, parent)) {
277                 dump_printf("problem processing PERF_RECORD_FORK, skipping event.\n");
278                 return -1;
279         }
280         total_fork++;
281
282         return 0;
283 }
284
285 static int
286 process_event(event_t *event, unsigned long offset, unsigned long head)
287 {
288         switch (event->header.type) {
289         case PERF_RECORD_SAMPLE:
290                 return process_sample_event(event, offset, head);
291
292         case PERF_RECORD_MMAP:
293                 return process_mmap_event(event, offset, head);
294
295         case PERF_RECORD_COMM:
296                 return process_comm_event(event, offset, head);
297
298         case PERF_RECORD_FORK:
299                 return process_fork_event(event, offset, head);
300         /*
301          * We dont process them right now but they are fine:
302          */
303
304         case PERF_RECORD_THROTTLE:
305         case PERF_RECORD_UNTHROTTLE:
306                 return 0;
307
308         default:
309                 return -1;
310         }
311
312         return 0;
313 }
314
315 static int
316 parse_line(FILE *file, struct symbol *sym, u64 start, u64 len)
317 {
318         char *line = NULL, *tmp, *tmp2;
319         static const char *prev_line;
320         static const char *prev_color;
321         unsigned int offset;
322         size_t line_len;
323         s64 line_ip;
324         int ret;
325         char *c;
326
327         if (getline(&line, &line_len, file) < 0)
328                 return -1;
329         if (!line)
330                 return -1;
331
332         c = strchr(line, '\n');
333         if (c)
334                 *c = 0;
335
336         line_ip = -1;
337         offset = 0;
338         ret = -2;
339
340         /*
341          * Strip leading spaces:
342          */
343         tmp = line;
344         while (*tmp) {
345                 if (*tmp != ' ')
346                         break;
347                 tmp++;
348         }
349
350         if (*tmp) {
351                 /*
352                  * Parse hexa addresses followed by ':'
353                  */
354                 line_ip = strtoull(tmp, &tmp2, 16);
355                 if (*tmp2 != ':')
356                         line_ip = -1;
357         }
358
359         if (line_ip != -1) {
360                 const char *path = NULL;
361                 unsigned int hits = 0;
362                 double percent = 0.0;
363                 const char *color;
364                 struct sym_ext *sym_ext = sym->priv;
365
366                 offset = line_ip - start;
367                 if (offset < len)
368                         hits = sym->hist[offset];
369
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;
375
376                 color = get_percent_color(percent);
377
378                 /*
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
382                  */
383                 if (path) {
384                         if (!prev_line || strcmp(prev_line, path)
385                                        || color != prev_color) {
386                                 color_fprintf(stdout, color, " %s", path);
387                                 prev_line = path;
388                                 prev_color = color;
389                         }
390                 }
391
392                 color_fprintf(stdout, color, " %7.2f", percent);
393                 printf(" :      ");
394                 color_fprintf(stdout, PERF_COLOR_BLUE, "%s\n", line);
395         } else {
396                 if (!*line)
397                         printf("         :\n");
398                 else
399                         printf("         :      %s\n", line);
400         }
401
402         return 0;
403 }
404
405 static struct rb_root root_sym_ext;
406
407 static void insert_source_line(struct sym_ext *sym_ext)
408 {
409         struct sym_ext *iter;
410         struct rb_node **p = &root_sym_ext.rb_node;
411         struct rb_node *parent = NULL;
412
413         while (*p != NULL) {
414                 parent = *p;
415                 iter = rb_entry(parent, struct sym_ext, node);
416
417                 if (sym_ext->percent > iter->percent)
418                         p = &(*p)->rb_left;
419                 else
420                         p = &(*p)->rb_right;
421         }
422
423         rb_link_node(&sym_ext->node, parent, p);
424         rb_insert_color(&sym_ext->node, &root_sym_ext);
425 }
426
427 static void free_source_line(struct symbol *sym, int len)
428 {
429         struct sym_ext *sym_ext = sym->priv;
430         int i;
431
432         if (!sym_ext)
433                 return;
434
435         for (i = 0; i < len; i++)
436                 free(sym_ext[i].path);
437         free(sym_ext);
438
439         sym->priv = NULL;
440         root_sym_ext = RB_ROOT;
441 }
442
443 /* Get the filename:line for the colored entries */
444 static void
445 get_source_line(struct symbol *sym, u64 start, int len, const char *filename)
446 {
447         int i;
448         char cmd[PATH_MAX * 2];
449         struct sym_ext *sym_ext;
450
451         if (!sym->hist_sum)
452                 return;
453
454         sym->priv = calloc(len, sizeof(struct sym_ext));
455         if (!sym->priv)
456                 return;
457
458         sym_ext = sym->priv;
459
460         for (i = 0; i < len; i++) {
461                 char *path = NULL;
462                 size_t line_len;
463                 u64 offset;
464                 FILE *fp;
465
466                 sym_ext[i].percent = 100.0 * sym->hist[i] / sym->hist_sum;
467                 if (sym_ext[i].percent <= 0.5)
468                         continue;
469
470                 offset = start + i;
471                 sprintf(cmd, "addr2line -e %s %016llx", filename, offset);
472                 fp = popen(cmd, "r");
473                 if (!fp)
474                         continue;
475
476                 if (getline(&path, &line_len, fp) < 0 || !line_len)
477                         goto next;
478
479                 sym_ext[i].path = malloc(sizeof(char) * line_len + 1);
480                 if (!sym_ext[i].path)
481                         goto next;
482
483                 strcpy(sym_ext[i].path, path);
484                 insert_source_line(&sym_ext[i]);
485
486         next:
487                 pclose(fp);
488         }
489 }
490
491 static void print_summary(const char *filename)
492 {
493         struct sym_ext *sym_ext;
494         struct rb_node *node;
495
496         printf("\nSorted summary for file %s\n", filename);
497         printf("----------------------------------------------\n\n");
498
499         if (RB_EMPTY_ROOT(&root_sym_ext)) {
500                 printf(" Nothing higher than %1.1f%%\n", MIN_GREEN);
501                 return;
502         }
503
504         node = rb_first(&root_sym_ext);
505         while (node) {
506                 double percent;
507                 const char *color;
508                 char *path;
509
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;
514
515                 color_fprintf(stdout, color, " %7.2f %s", percent, path);
516                 node = rb_next(node);
517         }
518 }
519
520 static void annotate_sym(struct dso *dso, struct symbol *sym)
521 {
522         const char *filename = dso->name, *d_filename;
523         u64 start, end, len;
524         char command[PATH_MAX*2];
525         FILE *file;
526
527         if (!filename)
528                 return;
529         if (sym->module)
530                 filename = sym->module->path;
531         else if (dso == kernel_dso)
532                 filename = vmlinux_name;
533
534         start = sym->obj_start;
535         if (!start)
536                 start = sym->start;
537         if (full_paths)
538                 d_filename = filename;
539         else
540                 d_filename = basename(filename);
541
542         end = start + sym->end - sym->start + 1;
543         len = sym->end - sym->start;
544
545         if (print_line) {
546                 get_source_line(sym, start, len, filename);
547                 print_summary(filename);
548         }
549
550         printf("\n\n------------------------------------------------\n");
551         printf(" Percent |      Source code & Disassembly of %s\n", d_filename);
552         printf("------------------------------------------------\n");
553
554         if (verbose >= 2)
555                 printf("annotating [%p] %30s : [%p] %30s\n", dso, dso->name, sym, sym->name);
556
557         sprintf(command, "objdump --start-address=0x%016Lx --stop-address=0x%016Lx -dS %s|grep -v %s",
558                         (u64)start, (u64)end, filename, filename);
559
560         if (verbose >= 3)
561                 printf("doing: %s\n", command);
562
563         file = popen(command, "r");
564         if (!file)
565                 return;
566
567         while (!feof(file)) {
568                 if (parse_line(file, sym, start, len) < 0)
569                         break;
570         }
571
572         pclose(file);
573         if (print_line)
574                 free_source_line(sym, len);
575 }
576
577 static void find_annotations(void)
578 {
579         struct rb_node *nd;
580         struct dso *dso;
581         int count = 0;
582
583         list_for_each_entry(dso, &dsos, node) {
584
585                 for (nd = rb_first(&dso->syms); nd; nd = rb_next(nd)) {
586                         struct symbol *sym = rb_entry(nd, struct symbol, rb_node);
587
588                         if (sym->hist) {
589                                 annotate_sym(dso, sym);
590                                 count++;
591                         }
592                 }
593         }
594
595         if (!count)
596                 printf(" Error: symbol '%s' not present amongst the samples.\n", sym_hist_filter);
597 }
598
599 static int __cmd_annotate(void)
600 {
601         int ret, rc = EXIT_FAILURE;
602         unsigned long offset = 0;
603         unsigned long head = 0;
604         struct stat input_stat;
605         event_t *event;
606         uint32_t size;
607         char *buf;
608
609         register_idle_thread(&threads, &last_match);
610
611         input = open(input_name, O_RDONLY);
612         if (input < 0) {
613                 perror("failed to open file");
614                 exit(-1);
615         }
616
617         ret = fstat(input, &input_stat);
618         if (ret < 0) {
619                 perror("failed to stat file");
620                 exit(-1);
621         }
622
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);
625                 exit(-1);
626         }
627
628         if (!input_stat.st_size) {
629                 fprintf(stderr, "zero-sized file, nothing to do!\n");
630                 exit(0);
631         }
632
633         if (load_kernel() < 0) {
634                 perror("failed to load kernel symbols");
635                 return EXIT_FAILURE;
636         }
637
638 remap:
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");
643                 exit(-1);
644         }
645
646 more:
647         event = (event_t *)(buf + head);
648
649         size = event->header.size;
650         if (!size)
651                 size = 8;
652
653         if (head + event->header.size >= page_size * mmap_window) {
654                 unsigned long shift = page_size * (head / page_size);
655                 int munmap_ret;
656
657                 munmap_ret = munmap(buf, page_size * mmap_window);
658                 assert(munmap_ret == 0);
659
660                 offset += shift;
661                 head -= shift;
662                 goto remap;
663         }
664
665         size = event->header.size;
666
667         dump_printf("%p [%p]: event: %d\n",
668                         (void *)(offset + head),
669                         (void *)(long)event->header.size,
670                         event->header.type);
671
672         if (!size || process_event(event, offset, head) < 0) {
673
674                 dump_printf("%p [%p]: skipping unknown header type: %d\n",
675                         (void *)(offset + head),
676                         (void *)(long)(event->header.size),
677                         event->header.type);
678
679                 total_unknown++;
680
681                 /*
682                  * assume we lost track of the stream, check alignment, and
683                  * increment a single u64 in the hope to catch on again 'soon'.
684                  */
685
686                 if (unlikely(head & 7))
687                         head &= ~7ULL;
688
689                 size = 8;
690         }
691
692         head += size;
693
694         if (offset + head < (unsigned long)input_stat.st_size)
695                 goto more;
696
697         rc = EXIT_SUCCESS;
698         close(input);
699
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);
705
706         if (dump_trace)
707                 return 0;
708
709         if (verbose >= 3)
710                 threads__fprintf(stdout, &threads);
711
712         if (verbose >= 2)
713                 dsos__fprintf(stdout);
714
715         collapse__resort();
716         output__resort(total);
717
718         find_annotations();
719
720         return rc;
721 }
722
723 static const char * const annotate_usage[] = {
724         "perf annotate [<options>] <command>",
725         NULL
726 };
727
728 static const struct option options[] = {
729         OPT_STRING('i', "input", &input_name, "file",
730                     "input file name"),
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"),
745         OPT_END()
746 };
747
748 static void setup_sorting(void)
749 {
750         char *tmp, *tok, *str = strdup(sort_order);
751
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);
757                 }
758         }
759
760         free(str);
761 }
762
763 int cmd_annotate(int argc, const char **argv, const char *prefix __used)
764 {
765         symbol__init();
766
767         page_size = getpagesize();
768
769         argc = parse_options(argc, argv, options, annotate_usage, 0);
770
771         setup_sorting();
772
773         if (argc) {
774                 /*
775                  * Special case: if there's an argument left then assume tha
776                  * it's a symbol filter:
777                  */
778                 if (argc > 1)
779                         usage_with_options(annotate_usage, options);
780
781                 sym_hist_filter = argv[0];
782         }
783
784         if (!sym_hist_filter)
785                 usage_with_options(annotate_usage, options);
786
787         setup_pager();
788
789         if (field_sep && *field_sep == '.') {
790                 fputs("'.' is the only non valid --field-separator argument\n",
791                                 stderr);
792                 exit(129);
793         }
794
795         return __cmd_annotate();
796 }