| 1 | |
| 2 | |
| 3 | * builtin-annotate.c
|
| 4 | *
|
| 5 | * Builtin annotate command: Analyze the perf.data input file,
|
| 6 | * look up and read DSOs and symbol information and display
|
| 7 | * a histogram of results, along various sorting keys.
|
| 8 |
|
| 9 | #include "builtin.h"
|
| 10 | #include "perf.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 |
|
| 18 | #include "util/debug.h"
|
| 19 |
|
| 20 | #include "util/evlist.h"
|
| 21 | #include "util/evsel.h"
|
| 22 | #include "util/annotate.h"
|
| 23 | #include "util/annotate-data.h"
|
| 24 | #include "util/event.h"
|
| 25 | #include <subcmd/parse-options.h>
|
| 26 | #include "util/parse-events.h"
|
| 27 | #include "util/sort.h"
|
| 28 | #include "util/hist.h"
|
| 29 | #include "util/dso.h"
|
| 30 | #include "util/machine.h"
|
| 31 | #include "util/map.h"
|
| 32 | #include "util/session.h"
|
| 33 | #include "util/tool.h"
|
| 34 | #include "util/data.h"
|
| 35 | #include "arch/common.h"
|
| 36 | #include "util/block-range.h"
|
| 37 | #include "util/map_symbol.h"
|
| 38 | #include "util/branch.h"
|
| 39 | #include "util/util.h"
|
| 40 | #include "ui/progress.h"
|
| 41 |
|
| 42 | #include <dlfcn.h>
|
| 43 | #include <errno.h>
|
| 44 | #include <linux/bitmap.h>
|
| 45 | #include <linux/err.h>
|
| 46 | #include <inttypes.h>
|
| 47 |
|
| 48 | struct perf_annotate {
|
| 49 | struct perf_tool tool;
|
| 50 | struct perf_session *session;
|
| 51 | #ifdef HAVE_SLANG_SUPPORT
|
| 52 | bool use_tui;
|
| 53 | #endif
|
| 54 | bool use_stdio, use_stdio2;
|
| 55 | #ifdef HAVE_GTK2_SUPPORT
|
| 56 | bool use_gtk;
|
| 57 | #endif
|
| 58 | bool skip_missing;
|
| 59 | bool has_br_stack;
|
| 60 | bool group_set;
|
| 61 | bool data_type;
|
| 62 | bool type_stat;
|
| 63 | bool insn_stat;
|
| 64 | float min_percent;
|
| 65 | const char *sym_hist_filter;
|
| 66 | const char *cpu_list;
|
| 67 | const char *target_data_type;
|
| 68 | DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
|
| 69 | };
|
| 70 |
|
| 71 | |
| 72 | * Given one basic block:
|
| 73 | *
|
| 74 | * from to branch_i
|
| 75 | * * ----> *
|
| 76 | * |
|
| 77 | * | block
|
| 78 | * v
|
| 79 | * * ----> *
|
| 80 | * from to branch_i+1
|
| 81 | *
|
| 82 | * where the horizontal are the branches and the vertical is the executed
|
| 83 | * block of instructions.
|
| 84 | *
|
| 85 | * We count, for each 'instruction', the number of blocks that covered it as
|
| 86 | * well as count the ratio each branch is taken.
|
| 87 | *
|
| 88 | * We can do this without knowing the actual instruction stream by keeping
|
| 89 | * track of the address ranges. We break down ranges such that there is no
|
| 90 | * overlap and iterate from the start until the end.
|
| 91 | *
|
| 92 | * @acme: once we parse the objdump output _before_ processing the samples,
|
| 93 | * we can easily fold the branch.cycles IPC bits in.
|
| 94 |
|
| 95 | static void process_basic_block(struct addr_map_symbol *start,
|
| 96 | struct addr_map_symbol *end,
|
| 97 | struct branch_flags *flags)
|
| 98 | {
|
| 99 | struct symbol *sym = start->ms.sym;
|
| 100 | struct annotation *notes = sym ? symbol__annotation(sym) : NULL;
|
| 101 | struct block_range_iter iter;
|
| 102 | struct block_range *entry;
|
| 103 | struct annotated_branch *branch;
|
| 104 |
|
| 105 | |
| 106 | * Sanity; NULL isn't executable and the CPU cannot execute backwards
|
| 107 |
|
| 108 | if (!start->addr || start->addr > end->addr)
|
| 109 | return;
|
| 110 |
|
| 111 | iter = block_range__create(start->addr, end->addr);
|
| 112 | if (!block_range_iter__valid(&iter))
|
| 113 | return;
|
| 114 |
|
| 115 | branch = annotation__get_branch(notes);
|
| 116 |
|
| 117 | |
| 118 | * First block in range is a branch target.
|
| 119 |
|
| 120 | entry = block_range_iter(&iter);
|
| 121 | assert(entry->is_target);
|
| 122 | entry->entry++;
|
| 123 |
|
| 124 | do {
|
| 125 | entry = block_range_iter(&iter);
|
| 126 |
|
| 127 | entry->coverage++;
|
| 128 | entry->sym = sym;
|
| 129 |
|
| 130 | if (branch)
|
| 131 | branch->max_coverage = max(branch->max_coverage, entry->coverage);
|
| 132 |
|
| 133 | } while (block_range_iter__next(&iter));
|
| 134 |
|
| 135 | |
| 136 | * Last block in rage is a branch.
|
| 137 |
|
| 138 | entry = block_range_iter(&iter);
|
| 139 | assert(entry->is_branch);
|
| 140 | entry->taken++;
|
| 141 | if (flags->predicted)
|
| 142 | entry->pred++;
|
| 143 | }
|
| 144 |
|
| 145 | static void process_branch_stack(struct branch_stack *bs, struct addr_location *al,
|
| 146 | struct perf_sample *sample)
|
| 147 | {
|
| 148 | struct addr_map_symbol *prev = NULL;
|
| 149 | struct branch_info *bi;
|
| 150 | int i;
|
| 151 |
|
| 152 | if (!bs || !bs->nr)
|
| 153 | return;
|
| 154 |
|
| 155 | bi = sample__resolve_bstack(sample, al);
|
| 156 | if (!bi)
|
| 157 | return;
|
| 158 |
|
| 159 | for (i = bs->nr - 1; i >= 0; i--) {
|
| 160 | |
| 161 | * XXX filter against symbol
|
| 162 |
|
| 163 | if (prev)
|
| 164 | process_basic_block(prev, &bi[i].from, &bi[i].flags);
|
| 165 | prev = &bi[i].to;
|
| 166 | }
|
| 167 |
|
| 168 | free(bi);
|
| 169 | }
|
| 170 |
|
| 171 | static int hist_iter__branch_callback(struct hist_entry_iter *iter,
|
| 172 | struct addr_location *al __maybe_unused,
|
| 173 | bool single __maybe_unused,
|
| 174 | void *arg __maybe_unused)
|
| 175 | {
|
| 176 | struct hist_entry *he = iter->he;
|
| 177 | struct branch_info *bi;
|
| 178 | struct perf_sample *sample = iter->sample;
|
| 179 | int err;
|
| 180 |
|
| 181 | bi = he->branch_info;
|
| 182 | err = addr_map_symbol__inc_samples(&bi->from, sample);
|
| 183 |
|
| 184 | if (err)
|
| 185 | goto out;
|
| 186 |
|
| 187 | err = addr_map_symbol__inc_samples(&bi->to, sample);
|
| 188 |
|
| 189 | out:
|
| 190 | return err;
|
| 191 | }
|
| 192 |
|
| 193 | static int process_branch_callback(struct perf_sample *sample,
|
| 194 | struct addr_location *al,
|
| 195 | struct perf_annotate *ann,
|
| 196 | struct machine *machine)
|
| 197 | {
|
| 198 | struct hist_entry_iter iter = {
|
| 199 | .sample = sample,
|
| 200 | .add_entry_cb = hist_iter__branch_callback,
|
| 201 | .hide_unresolved = symbol_conf.hide_unresolved,
|
| 202 | .ops = &hist_iter_branch,
|
| 203 | };
|
| 204 | struct addr_location a;
|
| 205 | int ret;
|
| 206 |
|
| 207 | addr_location__init(&a);
|
| 208 | if (machine__resolve(machine, &a, sample) < 0) {
|
| 209 | ret = -1;
|
| 210 | goto out;
|
| 211 | }
|
| 212 |
|
| 213 | if (a.sym == NULL) {
|
| 214 | ret = 0;
|
| 215 | goto out;
|
| 216 | }
|
| 217 |
|
| 218 | if (a.map != NULL)
|
| 219 | dso__set_hit(map__dso(a.map));
|
| 220 |
|
| 221 | hist__account_cycles(sample->branch_stack, al, sample, false,
|
| 222 | NULL);
|
| 223 |
|
| 224 | ret = hist_entry_iter__add(&iter, &a, PERF_MAX_STACK_DEPTH, ann);
|
| 225 | out:
|
| 226 | addr_location__exit(&a);
|
| 227 | return ret;
|
| 228 | }
|
| 229 |
|
| 230 | static bool has_annotation(struct perf_annotate *ann)
|
| 231 | {
|
| 232 | return ui__has_annotation() || ann->use_stdio2;
|
| 233 | }
|
| 234 |
|
| 235 | static int add_sample(struct perf_sample *sample,
|
| 236 | struct addr_location *al, struct perf_annotate *ann,
|
| 237 | struct machine *machine)
|
| 238 | {
|
| 239 | struct hists *hists = evsel__hists(sample->evsel);
|
| 240 | struct hist_entry *he;
|
| 241 | int ret;
|
| 242 |
|
| 243 | if ((!ann->has_br_stack || !has_annotation(ann)) &&
|
| 244 | ann->sym_hist_filter != NULL &&
|
| 245 | (al->sym == NULL ||
|
| 246 | strcmp(ann->sym_hist_filter, al->sym->name) != 0)) {
|
| 247 |
|
| 248 | |
| 249 | * FIXME: why isn't this done in the symbol_filter when loading
|
| 250 | * the DSO?
|
| 251 |
|
| 252 | if (al->sym != NULL) {
|
| 253 | struct dso *dso = map__dso(al->map);
|
| 254 |
|
| 255 | rb_erase_cached(&al->sym->rb_node, dso__symbols(dso));
|
| 256 | symbol__delete(al->sym);
|
| 257 | dso__reset_find_symbol_cache(dso);
|
| 258 | }
|
| 259 | return 0;
|
| 260 | }
|
| 261 |
|
| 262 | |
| 263 | * XXX filtered samples can still have branch entries pointing into our
|
| 264 | * symbol and are missed.
|
| 265 |
|
| 266 | process_branch_stack(sample->branch_stack, al, sample);
|
| 267 |
|
| 268 | if (ann->has_br_stack && has_annotation(ann))
|
| 269 | return process_branch_callback(sample, al, ann, machine);
|
| 270 |
|
| 271 | he = hists__add_entry(hists, al, NULL, NULL, NULL, NULL, sample, true);
|
| 272 | if (he == NULL)
|
| 273 | return -ENOMEM;
|
| 274 |
|
| 275 | ret = hist_entry__inc_addr_samples(he, sample, al->addr);
|
| 276 | hists__inc_nr_samples(hists, true);
|
| 277 | return ret;
|
| 278 | }
|
| 279 |
|
| 280 | static int process_sample_event(const struct perf_tool *tool,
|
| 281 | union perf_event *event,
|
| 282 | struct perf_sample *sample,
|
| 283 | struct machine *machine)
|
| 284 | {
|
| 285 | struct perf_annotate *ann = container_of(tool, struct perf_annotate, tool);
|
| 286 | struct addr_location al;
|
| 287 | int ret = 0;
|
| 288 |
|
| 289 | addr_location__init(&al);
|
| 290 | if (machine__resolve(machine, &al, sample) < 0) {
|
| 291 | pr_warning("problem processing %s (%u) event at offset %#" PRIx64 ", skipping it.\n",
|
| 292 | perf_event__name(event->header.type), event->header.type,
|
| 293 | sample->file_offset);
|
| 294 | ret = -1;
|
| 295 | goto out_put;
|
| 296 | }
|
| 297 |
|
| 298 | if (ann->cpu_list && (sample->cpu >= MAX_NR_CPUS ||
|
| 299 | !test_bit(sample->cpu, ann->cpu_bitmap)))
|
| 300 | goto out_put;
|
| 301 |
|
| 302 | if (!al.filtered &&
|
| 303 | add_sample(sample, &al, ann, machine)) {
|
| 304 | pr_warning("problem incrementing symbol count, "
|
| 305 | "skipping event\n");
|
| 306 | ret = -1;
|
| 307 | }
|
| 308 | out_put:
|
| 309 | addr_location__exit(&al);
|
| 310 | return ret;
|
| 311 | }
|
| 312 |
|
| 313 | static int hist_entry__stdio_annotate(struct hist_entry *he,
|
| 314 | struct evsel *evsel,
|
| 315 | struct perf_annotate *ann)
|
| 316 | {
|
| 317 | if (ann->use_stdio2)
|
| 318 | return hist_entry__tty_annotate2(he, evsel);
|
| 319 |
|
| 320 | return hist_entry__tty_annotate(he, evsel);
|
| 321 | }
|
| 322 |
|
| 323 | static void print_annotate_data_stat(struct annotated_data_stat *s)
|
| 324 | {
|
| 325 | #define PRINT_STAT(fld) if (s->fld) printf("%10d : %s\n", s->fld, #fld)
|
| 326 |
|
| 327 | int bad = s->no_sym +
|
| 328 | s->no_insn +
|
| 329 | s->no_insn_ops +
|
| 330 | s->no_mem_ops +
|
| 331 | s->no_reg +
|
| 332 | s->no_dbginfo +
|
| 333 | s->no_cuinfo +
|
| 334 | s->no_var +
|
| 335 | s->no_typeinfo +
|
| 336 | s->invalid_size +
|
| 337 | s->bad_offset;
|
| 338 | int ok = s->total - bad;
|
| 339 |
|
| 340 | printf("Annotate data type stats:\n");
|
| 341 | printf("total %d, ok %d (%.1f%%), bad %d (%.1f%%)\n",
|
| 342 | s->total, ok, 100.0 * ok / (s->total ?: 1), bad, 100.0 * bad / (s->total ?: 1));
|
| 343 | printf("-----------------------------------------------------------\n");
|
| 344 | PRINT_STAT(no_sym);
|
| 345 | PRINT_STAT(no_insn);
|
| 346 | PRINT_STAT(no_insn_ops);
|
| 347 | PRINT_STAT(no_mem_ops);
|
| 348 | PRINT_STAT(no_reg);
|
| 349 | PRINT_STAT(no_dbginfo);
|
| 350 | PRINT_STAT(no_cuinfo);
|
| 351 | PRINT_STAT(no_var);
|
| 352 | PRINT_STAT(no_typeinfo);
|
| 353 | PRINT_STAT(invalid_size);
|
| 354 | PRINT_STAT(bad_offset);
|
| 355 | PRINT_STAT(insn_track);
|
| 356 | printf("\n");
|
| 357 |
|
| 358 | #undef PRINT_STAT
|
| 359 | }
|
| 360 |
|
| 361 | static void print_annotate_item_stat(struct list_head *head, const char *title)
|
| 362 | {
|
| 363 | struct annotated_item_stat *istat, *pos, *iter;
|
| 364 | int total_good, total_bad, total;
|
| 365 | int sum1, sum2;
|
| 366 | LIST_HEAD(tmp);
|
| 367 |
|
| 368 |
|
| 369 | list_splice_init(head, &tmp);
|
| 370 | total_good = total_bad = 0;
|
| 371 |
|
| 372 | list_for_each_entry_safe(istat, pos, &tmp, list) {
|
| 373 | total_good += istat->good;
|
| 374 | total_bad += istat->bad;
|
| 375 | sum1 = istat->good + istat->bad;
|
| 376 |
|
| 377 | list_for_each_entry(iter, head, list) {
|
| 378 | sum2 = iter->good + iter->bad;
|
| 379 | if (sum1 > sum2)
|
| 380 | break;
|
| 381 | }
|
| 382 | list_move_tail(&istat->list, &iter->list);
|
| 383 | }
|
| 384 | total = total_good + total_bad;
|
| 385 |
|
| 386 | printf("Annotate %s stats\n", title);
|
| 387 | printf("total %d, ok %d (%.1f%%), bad %d (%.1f%%)\n\n", total,
|
| 388 | total_good, 100.0 * total_good / (total ?: 1),
|
| 389 | total_bad, 100.0 * total_bad / (total ?: 1));
|
| 390 | printf(" %-20s: %5s %5s\n", "Name/opcode", "Good", "Bad");
|
| 391 | printf("-----------------------------------------------------------\n");
|
| 392 | list_for_each_entry(istat, head, list)
|
| 393 | printf(" %-20s: %5d %5d\n", istat->name, istat->good, istat->bad);
|
| 394 | printf("\n");
|
| 395 | }
|
| 396 |
|
| 397 | static void hists__find_annotations(struct hists *hists,
|
| 398 | struct evsel *evsel,
|
| 399 | struct perf_annotate *ann)
|
| 400 | {
|
| 401 | struct rb_node *nd = rb_first_cached(&hists->entries), *next;
|
| 402 | int key = K_RIGHT;
|
| 403 |
|
| 404 | if (ann->type_stat)
|
| 405 | print_annotate_data_stat(&ann_data_stat);
|
| 406 | if (ann->insn_stat)
|
| 407 | print_annotate_item_stat(&ann_insn_stat, "Instruction");
|
| 408 |
|
| 409 | while (nd) {
|
| 410 | struct hist_entry *he = rb_entry(nd, struct hist_entry, rb_node);
|
| 411 | struct annotation *notes;
|
| 412 |
|
| 413 | if (he->ms.sym == NULL || dso__annotate_warned(map__dso(he->ms.map)))
|
| 414 | goto find_next;
|
| 415 |
|
| 416 | if (ann->sym_hist_filter &&
|
| 417 | (strcmp(he->ms.sym->name, ann->sym_hist_filter) != 0))
|
| 418 | goto find_next;
|
| 419 |
|
| 420 | if (ann->min_percent) {
|
| 421 | float percent = 0;
|
| 422 | u64 total = hists__total_period(hists);
|
| 423 |
|
| 424 | if (total)
|
| 425 | percent = 100.0 * he->stat.period / total;
|
| 426 |
|
| 427 | if (percent < ann->min_percent)
|
| 428 | goto find_next;
|
| 429 | }
|
| 430 |
|
| 431 | notes = symbol__annotation(he->ms.sym);
|
| 432 | if (notes->src == NULL) {
|
| 433 | find_next:
|
| 434 | if (key == K_LEFT || key == '<')
|
| 435 | nd = rb_prev(nd);
|
| 436 | else
|
| 437 | nd = rb_next(nd);
|
| 438 | continue;
|
| 439 | }
|
| 440 |
|
| 441 | if (ann->data_type) {
|
| 442 |
|
| 443 | if (he->mem_type->histograms == NULL)
|
| 444 | goto find_next;
|
| 445 |
|
| 446 | if (ann->target_data_type) {
|
| 447 | const char *type_name = he->mem_type->self.type_name;
|
| 448 |
|
| 449 |
|
| 450 | if (strncmp(ann->target_data_type, "struct ", 7) &&
|
| 451 | !strncmp(type_name, "struct ", 7))
|
| 452 | type_name += 7;
|
| 453 |
|
| 454 |
|
| 455 | if (strncmp(ann->target_data_type, "union ", 6) &&
|
| 456 | !strncmp(type_name, "union ", 6))
|
| 457 | type_name += 6;
|
| 458 |
|
| 459 | if (strcmp(ann->target_data_type, type_name))
|
| 460 | goto find_next;
|
| 461 | }
|
| 462 |
|
| 463 | if (use_browser == 1)
|
| 464 | key = hist_entry__annotate_data_tui(he, evsel, NULL);
|
| 465 | else
|
| 466 | key = hist_entry__annotate_data_tty(he, evsel);
|
| 467 |
|
| 468 | switch (key) {
|
| 469 | case -1:
|
| 470 | if (!ann->skip_missing)
|
| 471 | return;
|
| 472 |
|
| 473 | case K_RIGHT:
|
| 474 | case '>':
|
| 475 | next = rb_next(nd);
|
| 476 | break;
|
| 477 | case K_LEFT:
|
| 478 | case '<':
|
| 479 | next = rb_prev(nd);
|
| 480 | break;
|
| 481 | default:
|
| 482 | return;
|
| 483 | }
|
| 484 |
|
| 485 | if (use_browser == 0 || next != NULL)
|
| 486 | nd = next;
|
| 487 |
|
| 488 | continue;
|
| 489 | }
|
| 490 |
|
| 491 | if (use_browser == 2) {
|
| 492 | int ret;
|
| 493 | int (*annotate)(struct hist_entry *he,
|
| 494 | struct evsel *evsel,
|
| 495 | struct hist_browser_timer *hbt);
|
| 496 |
|
| 497 | annotate = dlsym(perf_gtk_handle,
|
| 498 | "hist_entry__gtk_annotate");
|
| 499 | if (annotate == NULL) {
|
| 500 | ui__error("GTK browser not found!\n");
|
| 501 | return;
|
| 502 | }
|
| 503 |
|
| 504 | ret = annotate(he, evsel, NULL);
|
| 505 | if (!ret || !ann->skip_missing)
|
| 506 | return;
|
| 507 |
|
| 508 |
|
| 509 | nd = rb_next(nd);
|
| 510 | } else if (use_browser == 1) {
|
| 511 | key = hist_entry__tui_annotate(he, evsel, NULL, NO_ADDR);
|
| 512 |
|
| 513 | switch (key) {
|
| 514 | case -1:
|
| 515 | if (!ann->skip_missing)
|
| 516 | return;
|
| 517 |
|
| 518 | case K_RIGHT:
|
| 519 | case '>':
|
| 520 | next = rb_next(nd);
|
| 521 | break;
|
| 522 | case K_LEFT:
|
| 523 | case '<':
|
| 524 | next = rb_prev(nd);
|
| 525 | break;
|
| 526 | default:
|
| 527 | return;
|
| 528 | }
|
| 529 |
|
| 530 | if (next != NULL)
|
| 531 | nd = next;
|
| 532 | } else {
|
| 533 | hist_entry__stdio_annotate(he, evsel, ann);
|
| 534 | nd = rb_next(nd);
|
| 535 | }
|
| 536 | }
|
| 537 | }
|
| 538 |
|
| 539 | static int __cmd_annotate(struct perf_annotate *ann)
|
| 540 | {
|
| 541 | int ret;
|
| 542 | struct perf_session *session = ann->session;
|
| 543 | struct evsel *pos;
|
| 544 | u64 total_nr_samples;
|
| 545 |
|
| 546 | if (ann->cpu_list) {
|
| 547 | ret = perf_session__cpu_bitmap(session, ann->cpu_list,
|
| 548 | ann->cpu_bitmap);
|
| 549 | if (ret)
|
| 550 | goto out;
|
| 551 | }
|
| 552 |
|
| 553 | if (!annotate_opts.objdump_path) {
|
| 554 | ret = perf_env__lookup_objdump(perf_session__env(session),
|
| 555 | &annotate_opts.objdump_path);
|
| 556 | if (ret)
|
| 557 | goto out;
|
| 558 | }
|
| 559 |
|
| 560 | ret = perf_session__process_events(session);
|
| 561 | if (ret)
|
| 562 | goto out;
|
| 563 |
|
| 564 | if ((use_browser == 1 || ann->use_stdio2) && ann->has_br_stack)
|
| 565 | if (evlist__nr_br_cntr(session->evlist) > 0)
|
| 566 | annotate_opts.show_br_cntr = true;
|
| 567 |
|
| 568 | if (dump_trace) {
|
| 569 | perf_session__fprintf_nr_events(session, stdout);
|
| 570 | evlist__fprintf_nr_events(session->evlist, stdout);
|
| 571 | goto out;
|
| 572 | }
|
| 573 |
|
| 574 | if (verbose > 3)
|
| 575 | perf_session__fprintf(session, stdout);
|
| 576 |
|
| 577 | if (verbose > 2)
|
| 578 | perf_session__fprintf_dsos(session, stdout);
|
| 579 |
|
| 580 | total_nr_samples = 0;
|
| 581 | evlist__for_each_entry(session->evlist, pos) {
|
| 582 | struct hists *hists = evsel__hists(pos);
|
| 583 | u32 nr_samples = hists->stats.nr_samples;
|
| 584 | struct ui_progress prog;
|
| 585 |
|
| 586 | if (nr_samples > 0) {
|
| 587 | total_nr_samples += nr_samples;
|
| 588 |
|
| 589 | ui_progress__init(&prog, nr_samples,
|
| 590 | "Merging related events...");
|
| 591 | hists__collapse_resort(hists, &prog);
|
| 592 | ui_progress__finish();
|
| 593 |
|
| 594 |
|
| 595 | evsel__reset_sample_bit(pos, CALLCHAIN);
|
| 596 |
|
| 597 | ui_progress__init(&prog, nr_samples,
|
| 598 | "Sorting events for output...");
|
| 599 | evsel__output_resort(pos, &prog);
|
| 600 | ui_progress__finish();
|
| 601 |
|
| 602 | |
| 603 | * An event group needs to display other events too.
|
| 604 | * Let's delay printing until other events are processed.
|
| 605 |
|
| 606 | if (symbol_conf.event_group) {
|
| 607 | if (!evsel__is_group_leader(pos)) {
|
| 608 | struct hists *leader_hists;
|
| 609 |
|
| 610 | leader_hists = evsel__hists(evsel__leader(pos));
|
| 611 | hists__match(leader_hists, hists);
|
| 612 | hists__link(leader_hists, hists);
|
| 613 | }
|
| 614 | continue;
|
| 615 | }
|
| 616 |
|
| 617 | hists__find_annotations(hists, pos, ann);
|
| 618 | }
|
| 619 | }
|
| 620 |
|
| 621 | if (total_nr_samples == 0) {
|
| 622 | ui__error("The %s data has no samples!\n", session->data->path);
|
| 623 | goto out;
|
| 624 | }
|
| 625 |
|
| 626 |
|
| 627 | evlist__for_each_entry(session->evlist, pos) {
|
| 628 | struct hists *hists = evsel__hists(pos);
|
| 629 | u32 nr_samples = hists->stats.nr_samples;
|
| 630 | struct ui_progress prog;
|
| 631 | struct evsel *evsel;
|
| 632 |
|
| 633 | if (!symbol_conf.event_group || !evsel__is_group_leader(pos))
|
| 634 | continue;
|
| 635 |
|
| 636 | for_each_group_member(evsel, pos)
|
| 637 | nr_samples += evsel__hists(evsel)->stats.nr_samples;
|
| 638 |
|
| 639 | if (nr_samples == 0)
|
| 640 | continue;
|
| 641 |
|
| 642 | ui_progress__init(&prog, nr_samples,
|
| 643 | "Sorting group events for output...");
|
| 644 | evsel__output_resort(pos, &prog);
|
| 645 | ui_progress__finish();
|
| 646 |
|
| 647 | hists__find_annotations(hists, pos, ann);
|
| 648 | }
|
| 649 |
|
| 650 | if (use_browser == 2) {
|
| 651 | void (*show_annotations)(void);
|
| 652 |
|
| 653 | show_annotations = dlsym(perf_gtk_handle,
|
| 654 | "perf_gtk__show_annotations");
|
| 655 | if (show_annotations == NULL) {
|
| 656 | ui__error("GTK browser not found!\n");
|
| 657 | goto out;
|
| 658 | }
|
| 659 | show_annotations();
|
| 660 | }
|
| 661 |
|
| 662 | out:
|
| 663 | return ret;
|
| 664 | }
|
| 665 |
|
| 666 | static int parse_percent_limit(const struct option *opt, const char *str,
|
| 667 | int unset __maybe_unused)
|
| 668 | {
|
| 669 | struct perf_annotate *ann = opt->value;
|
| 670 | double pcnt = strtof(str, NULL);
|
| 671 |
|
| 672 | ann->min_percent = pcnt;
|
| 673 | return 0;
|
| 674 | }
|
| 675 |
|
| 676 | static int parse_data_type(const struct option *opt, const char *str, int unset)
|
| 677 | {
|
| 678 | struct perf_annotate *ann = opt->value;
|
| 679 |
|
| 680 | ann->data_type = !unset;
|
| 681 | if (str)
|
| 682 | ann->target_data_type = strdup(str);
|
| 683 |
|
| 684 | return 0;
|
| 685 | }
|
| 686 |
|
| 687 | static const char * const annotate_usage[] = {
|
| 688 | "perf annotate [<options>]",
|
| 689 | NULL
|
| 690 | };
|
| 691 |
|
| 692 | int cmd_annotate(int argc, const char **argv)
|
| 693 | {
|
| 694 | struct perf_annotate annotate = {};
|
| 695 | struct perf_data data = {
|
| 696 | .mode = PERF_DATA_MODE_READ,
|
| 697 | };
|
| 698 | struct itrace_synth_opts itrace_synth_opts = {
|
| 699 | .set = 0,
|
| 700 | };
|
| 701 | const char *disassembler_style = NULL, *objdump_path = NULL, *addr2line_path = NULL;
|
| 702 | struct option options[] = {
|
| 703 | OPT_STRING('i', "input", &input_name, "file",
|
| 704 | "input file name"),
|
| 705 | OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
|
| 706 | "only consider symbols in these dsos"),
|
| 707 | OPT_STRING('s', "symbol", &annotate.sym_hist_filter, "symbol",
|
| 708 | "symbol to annotate"),
|
| 709 | OPT_BOOLEAN('f', "force", &data.force, "don't complain, do it"),
|
| 710 | OPT_INCR('v', "verbose", &verbose,
|
| 711 | "be more verbose (show symbol address, etc)"),
|
| 712 | OPT_BOOLEAN('q', "quiet", &quiet, "do now show any warnings or messages"),
|
| 713 | OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
|
| 714 | "dump raw trace in ASCII"),
|
| 715 | #ifdef HAVE_GTK2_SUPPORT
|
| 716 | OPT_BOOLEAN(0, "gtk", &annotate.use_gtk, "Use the GTK interface"),
|
| 717 | #endif
|
| 718 | #ifdef HAVE_SLANG_SUPPORT
|
| 719 | OPT_BOOLEAN(0, "tui", &annotate.use_tui, "Use the TUI interface"),
|
| 720 | #endif
|
| 721 | OPT_BOOLEAN(0, "stdio", &annotate.use_stdio, "Use the stdio interface"),
|
| 722 | OPT_BOOLEAN(0, "stdio2", &annotate.use_stdio2, "Use the stdio interface"),
|
| 723 | OPT_BOOLEAN(0, "ignore-vmlinux", &symbol_conf.ignore_vmlinux,
|
| 724 | "don't load vmlinux even if found"),
|
| 725 | OPT_STRING('k', "vmlinux", &symbol_conf.vmlinux_name,
|
| 726 | "file", "vmlinux pathname"),
|
| 727 | OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
|
| 728 | "load module symbols - WARNING: use only with -k and LIVE kernel"),
|
| 729 | OPT_BOOLEAN('l', "print-line", &annotate_opts.print_lines,
|
| 730 | "print matching source lines (may be slow)"),
|
| 731 | OPT_BOOLEAN('P', "full-paths", &annotate_opts.full_path,
|
| 732 | "Don't shorten the displayed pathnames"),
|
| 733 | OPT_BOOLEAN(0, "skip-missing", &annotate.skip_missing,
|
| 734 | "Skip symbols that cannot be annotated"),
|
| 735 | OPT_BOOLEAN_SET(0, "group", &symbol_conf.event_group,
|
| 736 | &annotate.group_set,
|
| 737 | "Show event group information together"),
|
| 738 | OPT_STRING('C', "cpu", &annotate.cpu_list, "cpu", "list of cpus to profile"),
|
| 739 | OPT_CALLBACK(0, "symfs", NULL, "directory[,layout]", SYMFS_HELP,
|
| 740 | symbol__config_symfs),
|
| 741 | OPT_BOOLEAN(0, "source", &annotate_opts.annotate_src,
|
| 742 | "Interleave source code with assembly code (default)"),
|
| 743 | OPT_BOOLEAN(0, "asm-raw", &annotate_opts.show_asm_raw,
|
| 744 | "Display raw encoding of assembly instructions (default)"),
|
| 745 | OPT_STRING('M', "disassembler-style", &disassembler_style, "disassembler style",
|
| 746 | "Specify disassembler style (e.g. -M intel for intel syntax)"),
|
| 747 | OPT_STRING(0, "prefix", &annotate_opts.prefix, "prefix",
|
| 748 | "Add prefix to source file path names in programs (with --prefix-strip)"),
|
| 749 | OPT_STRING(0, "prefix-strip", &annotate_opts.prefix_strip, "N",
|
| 750 | "Strip first N entries of source file path name in programs (with --prefix)"),
|
| 751 | OPT_STRING(0, "objdump", &objdump_path, "path",
|
| 752 | "objdump binary to use for disassembly and annotations"),
|
| 753 | OPT_STRING(0, "addr2line", &addr2line_path, "path",
|
| 754 | "addr2line binary to use for line numbers"),
|
| 755 | OPT_BOOLEAN(0, "demangle", &symbol_conf.demangle,
|
| 756 | "Enable symbol demangling"),
|
| 757 | OPT_BOOLEAN(0, "demangle-kernel", &symbol_conf.demangle_kernel,
|
| 758 | "Enable kernel symbol demangling"),
|
| 759 | OPT_BOOLEAN(0, "show-total-period", &symbol_conf.show_total_period,
|
| 760 | "Show a column with the sum of periods"),
|
| 761 | OPT_BOOLEAN('n', "show-nr-samples", &symbol_conf.show_nr_samples,
|
| 762 | "Show a column with the number of samples"),
|
| 763 | OPT_CALLBACK_DEFAULT(0, "stdio-color", NULL, "mode",
|
| 764 | "'always' (default), 'never' or 'auto' only applicable to --stdio mode",
|
| 765 | stdio__config_color, "always"),
|
| 766 | OPT_CALLBACK(0, "percent-type", &annotate_opts, "local-period",
|
| 767 | "Set percent type local/global-period/hits",
|
| 768 | annotate_parse_percent_type),
|
| 769 | OPT_CALLBACK(0, "percent-limit", &annotate, "percent",
|
| 770 | "Don't show entries under that percent", parse_percent_limit),
|
| 771 | OPT_CALLBACK_OPTARG(0, "itrace", &itrace_synth_opts, NULL, "opts",
|
| 772 | "Instruction Tracing options\n" ITRACE_HELP,
|
| 773 | itrace_parse_synth_opts),
|
| 774 | OPT_CALLBACK_OPTARG(0, "data-type", &annotate, NULL, "name",
|
| 775 | "Show data type annotate for the memory accesses",
|
| 776 | parse_data_type),
|
| 777 | OPT_BOOLEAN(0, "type-stat", &annotate.type_stat,
|
| 778 | "Show stats for the data type annotation"),
|
| 779 | OPT_BOOLEAN(0, "insn-stat", &annotate.insn_stat,
|
| 780 | "Show instruction stats for the data type annotation"),
|
| 781 | OPT_BOOLEAN(0, "skip-empty", &symbol_conf.skip_empty,
|
| 782 | "Do not display empty (or dummy) events in the output"),
|
| 783 | OPT_BOOLEAN(0, "code-with-type", &annotate_opts.code_with_type,
|
| 784 | "Show data type info in code annotation (memory instructions only)"),
|
| 785 | OPT_END()
|
| 786 | };
|
| 787 | int ret;
|
| 788 |
|
| 789 | set_option_flag(options, 0, "show-total-period", PARSE_OPT_EXCLUSIVE);
|
| 790 | set_option_flag(options, 0, "show-nr-samples", PARSE_OPT_EXCLUSIVE);
|
| 791 |
|
| 792 | annotation_options__init();
|
| 793 |
|
| 794 | ret = hists__init();
|
| 795 | if (ret < 0)
|
| 796 | return ret;
|
| 797 |
|
| 798 | annotation_config__init();
|
| 799 |
|
| 800 | argc = parse_options(argc, argv, options, annotate_usage, 0);
|
| 801 | if (argc) {
|
| 802 | |
| 803 | * Special case: if there's an argument left then assume that
|
| 804 | * it's a symbol filter:
|
| 805 |
|
| 806 | if (argc > 1)
|
| 807 | usage_with_options(annotate_usage, options);
|
| 808 |
|
| 809 | annotate.sym_hist_filter = argv[0];
|
| 810 | }
|
| 811 |
|
| 812 | if (disassembler_style) {
|
| 813 | annotate_opts.disassembler_style = strdup(disassembler_style);
|
| 814 | if (!annotate_opts.disassembler_style)
|
| 815 | return -ENOMEM;
|
| 816 | }
|
| 817 | if (objdump_path) {
|
| 818 | annotate_opts.objdump_path = strdup(objdump_path);
|
| 819 | if (!annotate_opts.objdump_path)
|
| 820 | return -ENOMEM;
|
| 821 | }
|
| 822 | if (addr2line_path) {
|
| 823 | symbol_conf.addr2line_path = strdup(addr2line_path);
|
| 824 | if (!symbol_conf.addr2line_path)
|
| 825 | return -ENOMEM;
|
| 826 | }
|
| 827 |
|
| 828 | if (annotate_check_args() < 0)
|
| 829 | return -EINVAL;
|
| 830 |
|
| 831 | #ifdef HAVE_GTK2_SUPPORT
|
| 832 | if (symbol_conf.show_nr_samples && annotate.use_gtk) {
|
| 833 | pr_err("--show-nr-samples is not available in --gtk mode at this time\n");
|
| 834 | return ret;
|
| 835 | }
|
| 836 | #endif
|
| 837 |
|
| 838 | #ifndef HAVE_LIBDW_SUPPORT
|
| 839 | if (annotate.data_type) {
|
| 840 | pr_err("Error: Data type profiling is disabled due to missing DWARF support\n");
|
| 841 | return -ENOTSUP;
|
| 842 | }
|
| 843 | #endif
|
| 844 |
|
| 845 | ret = symbol__validate_sym_arguments();
|
| 846 | if (ret)
|
| 847 | return ret;
|
| 848 |
|
| 849 | if (quiet)
|
| 850 | perf_quiet_option();
|
| 851 |
|
| 852 | data.path = input_name;
|
| 853 |
|
| 854 | perf_tool__init(&annotate.tool, true);
|
| 855 | annotate.tool.sample = process_sample_event;
|
| 856 | annotate.tool.mmap = perf_event__process_mmap;
|
| 857 | annotate.tool.mmap2 = perf_event__process_mmap2;
|
| 858 | annotate.tool.comm = perf_event__process_comm;
|
| 859 | annotate.tool.exit = perf_event__process_exit;
|
| 860 | annotate.tool.fork = perf_event__process_fork;
|
| 861 | annotate.tool.namespaces = perf_event__process_namespaces;
|
| 862 | annotate.tool.attr = perf_event__process_attr;
|
| 863 | annotate.tool.build_id = perf_event__process_build_id;
|
| 864 | #ifdef HAVE_LIBTRACEEVENT
|
| 865 | annotate.tool.tracing_data = perf_event__process_tracing_data;
|
| 866 | #endif
|
| 867 | annotate.tool.id_index = perf_event__process_id_index;
|
| 868 | annotate.tool.auxtrace_info = perf_event__process_auxtrace_info;
|
| 869 | annotate.tool.auxtrace = perf_event__process_auxtrace;
|
| 870 | annotate.tool.feature = perf_event__process_feature;
|
| 871 | annotate.tool.ordering_requires_timestamps = true;
|
| 872 |
|
| 873 | annotate.session = perf_session__new(&data, &annotate.tool);
|
| 874 | if (IS_ERR(annotate.session))
|
| 875 | return PTR_ERR(annotate.session);
|
| 876 |
|
| 877 | annotate.session->itrace_synth_opts = &itrace_synth_opts;
|
| 878 |
|
| 879 | annotate.has_br_stack = perf_header__has_feat(&annotate.session->header,
|
| 880 | HEADER_BRANCH_STACK);
|
| 881 |
|
| 882 | if (annotate.group_set)
|
| 883 | evlist__force_leader(annotate.session->evlist);
|
| 884 |
|
| 885 | ret = symbol__annotation_init();
|
| 886 | if (ret < 0)
|
| 887 | goto out_delete;
|
| 888 |
|
| 889 | symbol_conf.try_vmlinux_path = true;
|
| 890 |
|
| 891 | ret = symbol__init(perf_session__env(annotate.session));
|
| 892 | if (ret < 0)
|
| 893 | goto out_delete;
|
| 894 |
|
| 895 | if (annotate.use_stdio || annotate.use_stdio2)
|
| 896 | use_browser = 0;
|
| 897 | #ifdef HAVE_SLANG_SUPPORT
|
| 898 | else if (annotate.use_tui)
|
| 899 | use_browser = 1;
|
| 900 | #endif
|
| 901 | #ifdef HAVE_GTK2_SUPPORT
|
| 902 | else if (annotate.use_gtk)
|
| 903 | use_browser = 2;
|
| 904 | #endif
|
| 905 |
|
| 906 | if (annotate.data_type) {
|
| 907 | annotate_opts.annotate_src = false;
|
| 908 | symbol_conf.annotate_data_member = true;
|
| 909 | symbol_conf.annotate_data_sample = true;
|
| 910 | } else if (annotate_opts.code_with_type) {
|
| 911 | symbol_conf.annotate_data_member = true;
|
| 912 | }
|
| 913 |
|
| 914 | setup_browser(true);
|
| 915 |
|
| 916 | |
| 917 | * Events of different processes may correspond to the same
|
| 918 | * symbol, we do not care about the processes in annotate,
|
| 919 | * set sort order to avoid repeated output.
|
| 920 |
|
| 921 | if (annotate.data_type)
|
| 922 | sort_order = "dso,type";
|
| 923 | else
|
| 924 | sort_order = "dso,symbol";
|
| 925 |
|
| 926 | |
| 927 | * Set SORT_MODE__BRANCH so that annotate displays IPC/Cycle and
|
| 928 | * branch counters, if the corresponding branch info is available
|
| 929 | * in the perf data in the TUI mode.
|
| 930 |
|
| 931 | if ((use_browser == 1 || annotate.use_stdio2) && annotate.has_br_stack) {
|
| 932 | sort__mode = SORT_MODE__BRANCH;
|
| 933 | if (evlist__nr_br_cntr(annotate.session->evlist) > 0)
|
| 934 | annotate_opts.show_br_cntr = true;
|
| 935 | }
|
| 936 |
|
| 937 | if (setup_sorting(NULL, perf_session__env(annotate.session)) < 0)
|
| 938 | usage_with_options(annotate_usage, options);
|
| 939 |
|
| 940 | ret = __cmd_annotate(&annotate);
|
| 941 |
|
| 942 | out_delete:
|
| 943 | |
| 944 | * Speed up the exit process by only deleting for debug builds. For
|
| 945 | * large files this can save time.
|
| 946 |
|
| 947 | #ifndef NDEBUG
|
| 948 | perf_session__delete(annotate.session);
|
| 949 | #endif
|
| 950 | annotation_options__exit();
|
| 951 |
|
| 952 | return ret;
|
| 953 | }
|