tools/perf/util/hist.c

   1 #include "hist.h"
   2 #include "session.h"
   3 #include "sort.h"
   4 #include <math.h>
   5
   6 struct callchain_param  callchain_param = {
   7         .mode   = CHAIN_GRAPH_REL,
   8         .min_percent = 0.5
   9 };
  10
  11 /*
  12  * histogram, sorted on item, collects counts
  13  */
  14
  15 struct hist_entry *__perf_session__add_hist_entry(struct rb_root *hists,
  16                                                   struct addr_location *al,
  17                                                   struct symbol *sym_parent,
  18                                                   u64 count, bool *hit)
  19 {
  20         struct rb_node **p = &hists->rb_node;
  21         struct rb_node *parent = NULL;
  22         struct hist_entry *he;
  23         struct hist_entry entry = {
  24                 .thread = al->thread,
  25                 .map    = al->map,
  26                 .sym    = al->sym,
  27                 .ip     = al->addr,
  28                 .level  = al->level,
  29                 .count  = count,
  30                 .parent = sym_parent,
  31         };
  32         int cmp;
  33
  34         while (*p != NULL) {
  35                 parent = *p;
  36                 he = rb_entry(parent, struct hist_entry, rb_node);
  37
  38                 cmp = hist_entry__cmp(&entry, he);
  39
  40                 if (!cmp) {
  41                         *hit = true;
  42                         return he;
  43                 }
  44
  45                 if (cmp < 0)
  46                         p = &(*p)->rb_left;
  47                 else
  48                         p = &(*p)->rb_right;
  49         }
  50
  51         he = malloc(sizeof(*he));
  52         if (!he)
  53                 return NULL;
  54         *he = entry;
  55         rb_link_node(&he->rb_node, parent, p);
  56         rb_insert_color(&he->rb_node, hists);
  57         *hit = false;
  58         return he;
  59 }
  60
  61 int64_t
  62 hist_entry__cmp(struct hist_entry *left, struct hist_entry *right)
  63 {
  64         struct sort_entry *se;
  65         int64_t cmp = 0;
  66
  67         list_for_each_entry(se, &hist_entry__sort_list, list) {
  68                 cmp = se->cmp(left, right);
  69                 if (cmp)
  70                         break;
  71         }
  72
  73         return cmp;
  74 }
  75
  76 int64_t
  77 hist_entry__collapse(struct hist_entry *left, struct hist_entry *right)
  78 {
  79         struct sort_entry *se;
  80         int64_t cmp = 0;
  81
  82         list_for_each_entry(se, &hist_entry__sort_list, list) {
  83                 int64_t (*f)(struct hist_entry *, struct hist_entry *);
  84
  85                 f = se->collapse ?: se->cmp;
  86
  87                 cmp = f(left, right);
  88                 if (cmp)
  89                         break;
  90         }
  91
  92         return cmp;
  93 }
  94
  95 void hist_entry__free(struct hist_entry *he)
  96 {
  97         free(he);
  98 }
  99
 100 /*
 101  * collapse the histogram
 102  */
 103
 104 static void collapse__insert_entry(struct rb_root *root, struct hist_entry *he)
 105 {
 106         struct rb_node **p = &root->rb_node;
 107         struct rb_node *parent = NULL;
 108         struct hist_entry *iter;
 109         int64_t cmp;
 110
 111         while (*p != NULL) {
 112                 parent = *p;
 113                 iter = rb_entry(parent, struct hist_entry, rb_node);
 114
 115                 cmp = hist_entry__collapse(iter, he);
 116
 117                 if (!cmp) {
 118                         iter->count += he->count;
 119                         hist_entry__free(he);
 120                         return;
 121                 }
 122
 123                 if (cmp < 0)
 124                         p = &(*p)->rb_left;
 125                 else
 126                         p = &(*p)->rb_right;
 127         }
 128
 129         rb_link_node(&he->rb_node, parent, p);
 130         rb_insert_color(&he->rb_node, root);
 131 }
 132
 133 void perf_session__collapse_resort(struct rb_root *hists)
 134 {
 135         struct rb_root tmp;
 136         struct rb_node *next;
 137         struct hist_entry *n;
 138
 139         if (!sort__need_collapse)
 140                 return;
 141
 142         tmp = RB_ROOT;
 143         next = rb_first(hists);
 144
 145         while (next) {
 146                 n = rb_entry(next, struct hist_entry, rb_node);
 147                 next = rb_next(&n->rb_node);
 148
 149                 rb_erase(&n->rb_node, hists);
 150                 collapse__insert_entry(&tmp, n);
 151         }
 152
 153         *hists = tmp;
 154 }
 155
 156 /*
 157  * reverse the map, sort on count.
 158  */
 159
 160 static void perf_session__insert_output_hist_entry(struct rb_root *root,
 161                                                    struct hist_entry *he,
 162                                                    u64 min_callchain_hits)
 163 {
 164         struct rb_node **p = &root->rb_node;
 165         struct rb_node *parent = NULL;
 166         struct hist_entry *iter;
 167
 168         if (symbol_conf.use_callchain)
 169                 callchain_param.sort(&he->sorted_chain, &he->callchain,
 170                                       min_callchain_hits, &callchain_param);
 171
 172         while (*p != NULL) {
 173                 parent = *p;
 174                 iter = rb_entry(parent, struct hist_entry, rb_node);
 175
 176                 if (he->count > iter->count)
 177                         p = &(*p)->rb_left;
 178                 else
 179                         p = &(*p)->rb_right;
 180         }
 181
 182         rb_link_node(&he->rb_node, parent, p);
 183         rb_insert_color(&he->rb_node, root);
 184 }
 185
 186 void perf_session__output_resort(struct rb_root *hists, u64 total_samples)
 187 {
 188         struct rb_root tmp;
 189         struct rb_node *next;
 190         struct hist_entry *n;
 191         u64 min_callchain_hits;
 192
 193         min_callchain_hits =
 194                 total_samples * (callchain_param.min_percent / 100);
 195
 196         tmp = RB_ROOT;
 197         next = rb_first(hists);
 198
 199         while (next) {
 200                 n = rb_entry(next, struct hist_entry, rb_node);
 201                 next = rb_next(&n->rb_node);
 202
 203                 rb_erase(&n->rb_node, hists);
 204                 perf_session__insert_output_hist_entry(&tmp, n,
 205                                                        min_callchain_hits);
 206         }
 207
 208         *hists = tmp;
 209 }
 210
 211 static size_t callchain__fprintf_left_margin(FILE *fp, int left_margin)
 212 {
 213         int i;
 214         int ret = fprintf(fp, "            ");
 215
 216         for (i = 0; i < left_margin; i++)
 217                 ret += fprintf(fp, " ");
 218
 219         return ret;
 220 }
 221
 222 static size_t ipchain__fprintf_graph_line(FILE *fp, int depth, int depth_mask,
 223                                           int left_margin)
 224 {
 225         int i;
 226         size_t ret = callchain__fprintf_left_margin(fp, left_margin);
 227
 228         for (i = 0; i < depth; i++)
 229                 if (depth_mask & (1 << i))
 230                         ret += fprintf(fp, "|          ");
 231                 else
 232                         ret += fprintf(fp, "           ");
 233
 234         ret += fprintf(fp, "\n");
 235
 236         return ret;
 237 }
 238
 239 static size_t ipchain__fprintf_graph(FILE *fp, struct callchain_list *chain,
 240                                      int depth, int depth_mask, int count,
 241                                      u64 total_samples, int hits,
 242                                      int left_margin)
 243 {
 244         int i;
 245         size_t ret = 0;
 246
 247         ret += callchain__fprintf_left_margin(fp, left_margin);
 248         for (i = 0; i < depth; i++) {
 249                 if (depth_mask & (1 << i))
 250                         ret += fprintf(fp, "|");
 251                 else
 252                         ret += fprintf(fp, " ");
 253                 if (!count && i == depth - 1) {
 254                         double percent;
 255
 256                         percent = hits * 100.0 / total_samples;
 257                         ret += percent_color_fprintf(fp, "--%2.2f%%-- ", percent);
 258                 } else
 259                         ret += fprintf(fp, "%s", "          ");
 260         }
 261         if (chain->sym)
 262                 ret += fprintf(fp, "%s\n", chain->sym->name);
 263         else
 264                 ret += fprintf(fp, "%p\n", (void *)(long)chain->ip);
 265
 266         return ret;
 267 }
 268
 269 static struct symbol *rem_sq_bracket;
 270 static struct callchain_list rem_hits;
 271
 272 static void init_rem_hits(void)
 273 {
 274         rem_sq_bracket = malloc(sizeof(*rem_sq_bracket) + 6);
 275         if (!rem_sq_bracket) {
 276                 fprintf(stderr, "Not enough memory to display remaining hits\n");
 277                 return;
 278         }
 279
 280         strcpy(rem_sq_bracket->name, "[...]");
 281         rem_hits.sym = rem_sq_bracket;
 282 }
 283
 284 static size_t __callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
 285                                          u64 total_samples, int depth,
 286                                          int depth_mask, int left_margin)
 287 {
 288         struct rb_node *node, *next;
 289         struct callchain_node *child;
 290         struct callchain_list *chain;
 291         int new_depth_mask = depth_mask;
 292         u64 new_total;
 293         u64 remaining;
 294         size_t ret = 0;
 295         int i;
 296
 297         if (callchain_param.mode == CHAIN_GRAPH_REL)
 298                 new_total = self->children_hit;
 299         else
 300                 new_total = total_samples;
 301
 302         remaining = new_total;
 303
 304         node = rb_first(&self->rb_root);
 305         while (node) {
 306                 u64 cumul;
 307
 308                 child = rb_entry(node, struct callchain_node, rb_node);
 309                 cumul = cumul_hits(child);
 310                 remaining -= cumul;
 311
 312                 /*
 313                  * The depth mask manages the output of pipes that show
 314                  * the depth. We don't want to keep the pipes of the current
 315                  * level for the last child of this depth.
 316                  * Except if we have remaining filtered hits. They will
 317                  * supersede the last child
 318                  */
 319                 next = rb_next(node);
 320                 if (!next && (callchain_param.mode != CHAIN_GRAPH_REL || !remaining))
 321                         new_depth_mask &= ~(1 << (depth - 1));
 322
 323                 /*
 324                  * But we keep the older depth mask for the line separator
 325                  * to keep the level link until we reach the last child
 326                  */
 327                 ret += ipchain__fprintf_graph_line(fp, depth, depth_mask,
 328                                                    left_margin);
 329                 i = 0;
 330                 list_for_each_entry(chain, &child->val, list) {
 331                         ret += ipchain__fprintf_graph(fp, chain, depth,
 332                                                       new_depth_mask, i++,
 333                                                       new_total,
 334                                                       cumul,
 335                                                       left_margin);
 336                 }
 337                 ret += __callchain__fprintf_graph(fp, child, new_total,
 338                                                   depth + 1,
 339                                                   new_depth_mask | (1 << depth),
 340                                                   left_margin);
 341                 node = next;
 342         }
 343
 344         if (callchain_param.mode == CHAIN_GRAPH_REL &&
 345                 remaining && remaining != new_total) {
 346
 347                 if (!rem_sq_bracket)
 348                         return ret;
 349
 350                 new_depth_mask &= ~(1 << (depth - 1));
 351
 352                 ret += ipchain__fprintf_graph(fp, &rem_hits, depth,
 353                                               new_depth_mask, 0, new_total,
 354                                               remaining, left_margin);
 355         }
 356
 357         return ret;
 358 }
 359
 360 static size_t callchain__fprintf_graph(FILE *fp, struct callchain_node *self,
 361                                        u64 total_samples, int left_margin)
 362 {
 363         struct callchain_list *chain;
 364         bool printed = false;
 365         int i = 0;
 366         int ret = 0;
 367
 368         list_for_each_entry(chain, &self->val, list) {
 369                 if (!i++ && sort__first_dimension == SORT_SYM)
 370                         continue;
 371
 372                 if (!printed) {
 373                         ret += callchain__fprintf_left_margin(fp, left_margin);
 374                         ret += fprintf(fp, "|\n");
 375                         ret += callchain__fprintf_left_margin(fp, left_margin);
 376                         ret += fprintf(fp, "---");
 377
 378                         left_margin += 3;
 379                         printed = true;
 380                 } else
 381                         ret += callchain__fprintf_left_margin(fp, left_margin);
 382
 383                 if (chain->sym)
 384                         ret += fprintf(fp, " %s\n", chain->sym->name);
 385                 else
 386                         ret += fprintf(fp, " %p\n", (void *)(long)chain->ip);
 387         }
 388
 389         ret += __callchain__fprintf_graph(fp, self, total_samples, 1, 1, left_margin);
 390
 391         return ret;
 392 }
 393
 394 static size_t callchain__fprintf_flat(FILE *fp, struct callchain_node *self,
 395                                       u64 total_samples)
 396 {
 397         struct callchain_list *chain;
 398         size_t ret = 0;
 399
 400         if (!self)
 401                 return 0;
 402
 403         ret += callchain__fprintf_flat(fp, self->parent, total_samples);
 404
 405
 406         list_for_each_entry(chain, &self->val, list) {
 407                 if (chain->ip >= PERF_CONTEXT_MAX)
 408                         continue;
 409                 if (chain->sym)
 410                         ret += fprintf(fp, "                %s\n", chain->sym->name);
 411                 else
 412                         ret += fprintf(fp, "                %p\n",
 413                                         (void *)(long)chain->ip);
 414         }
 415
 416         return ret;
 417 }
 418
 419 static size_t hist_entry_callchain__fprintf(FILE *fp, struct hist_entry *self,
 420                                             u64 total_samples, int left_margin)
 421 {
 422         struct rb_node *rb_node;
 423         struct callchain_node *chain;
 424         size_t ret = 0;
 425
 426         rb_node = rb_first(&self->sorted_chain);
 427         while (rb_node) {
 428                 double percent;
 429
 430                 chain = rb_entry(rb_node, struct callchain_node, rb_node);
 431                 percent = chain->hit * 100.0 / total_samples;
 432                 switch (callchain_param.mode) {
 433                 case CHAIN_FLAT:
 434                         ret += percent_color_fprintf(fp, "           %6.2f%%\n",
 435                                                      percent);
 436                         ret += callchain__fprintf_flat(fp, chain, total_samples);
 437                         break;
 438                 case CHAIN_GRAPH_ABS: /* Falldown */
 439                 case CHAIN_GRAPH_REL:
 440                         ret += callchain__fprintf_graph(fp, chain, total_samples,
 441                                                         left_margin);
 442                 case CHAIN_NONE:
 443                 default:
 444                         break;
 445                 }
 446                 ret += fprintf(fp, "\n");
 447                 rb_node = rb_next(rb_node);
 448         }
 449
 450         return ret;
 451 }
 452
 453 size_t hist_entry__fprintf(struct hist_entry *self,
 454                            struct perf_session *pair_session,
 455                            bool show_displacement,
 456                            long displacement, FILE *fp,
 457                            u64 session_total)
 458 {
 459         struct sort_entry *se;
 460         u64 count, total;
 461         const char *sep = symbol_conf.field_sep;
 462         size_t ret;
 463
 464         if (symbol_conf.exclude_other && !self->parent)
 465                 return 0;
 466
 467         if (pair_session) {
 468                 count = self->pair ? self->pair->count : 0;
 469                 total = pair_session->events_stats.total;
 470         } else {
 471                 count = self->count;
 472                 total = session_total;
 473         }
 474
 475         if (total)
 476                 ret = percent_color_fprintf(fp, sep ? "%.2f" : "   %6.2f%%",
 477                                             (count * 100.0) / total);
 478         else
 479                 ret = fprintf(fp, sep ? "%lld" : "%12lld ", count);
 480
 481         if (symbol_conf.show_nr_samples) {
 482                 if (sep)
 483                         ret += fprintf(fp, "%c%lld", *sep, count);
 484                 else
 485                         ret += fprintf(fp, "%11lld", count);
 486         }
 487
 488         if (pair_session) {
 489                 char bf[32];
 490                 double old_percent = 0, new_percent = 0, diff;
 491
 492                 if (total > 0)
 493                         old_percent = (count * 100.0) / total;
 494                 if (session_total > 0)
 495                         new_percent = (self->count * 100.0) / session_total;
 496
 497                 diff = new_percent - old_percent;
 498
 499                 if (fabs(diff) >= 0.01)
 500                         snprintf(bf, sizeof(bf), "%+4.2F%%", diff);
 501                 else
 502                         snprintf(bf, sizeof(bf), " ");
 503
 504                 if (sep)
 505                         ret += fprintf(fp, "%c%s", *sep, bf);
 506                 else
 507                         ret += fprintf(fp, "%11.11s", bf);
 508
 509                 if (show_displacement) {
 510                         if (displacement)
 511                                 snprintf(bf, sizeof(bf), "%+4ld", displacement);
 512                         else
 513                                 snprintf(bf, sizeof(bf), " ");
 514
 515                         if (sep)
 516                                 ret += fprintf(fp, "%c%s", *sep, bf);
 517                         else
 518                                 ret += fprintf(fp, "%6.6s", bf);
 519                 }
 520         }
 521
 522         list_for_each_entry(se, &hist_entry__sort_list, list) {
 523                 if (se->elide)
 524                         continue;
 525
 526                 ret += fprintf(fp, "%s", sep ?: "  ");
 527                 ret += se->print(fp, self, se->width ? *se->width : 0);
 528         }
 529
 530         return ret + fprintf(fp, "\n");
 531 }
 532
 533 static size_t hist_entry__fprintf_callchain(struct hist_entry *self, FILE *fp,
 534                                             u64 session_total)
 535 {
 536         int left_margin = 0;
 537
 538         if (sort__first_dimension == SORT_COMM) {
 539                 struct sort_entry *se = list_first_entry(&hist_entry__sort_list,
 540                                                          typeof(*se), list);
 541                 left_margin = se->width ? *se->width : 0;
 542                 left_margin -= thread__comm_len(self->thread);
 543         }
 544
 545         return hist_entry_callchain__fprintf(fp, self, session_total,
 546                                              left_margin);
 547 }
 548
 549 size_t perf_session__fprintf_hists(struct rb_root *hists,
 550                                    struct perf_session *pair,
 551                                    bool show_displacement, FILE *fp,
 552                                    u64 session_total)
 553 {
 554         struct sort_entry *se;
 555         struct rb_node *nd;
 556         size_t ret = 0;
 557         unsigned long position = 1;
 558         long displacement = 0;
 559         unsigned int width;
 560         const char *sep = symbol_conf.field_sep;
 561         char *col_width = symbol_conf.col_width_list_str;
 562
 563         init_rem_hits();
 564
 565         fprintf(fp, "# %s", pair ? "Baseline" : "Overhead");
 566
 567         if (symbol_conf.show_nr_samples) {
 568                 if (sep)
 569                         fprintf(fp, "%cSamples", *sep);
 570                 else
 571                         fputs("  Samples  ", fp);
 572         }
 573
 574         if (pair) {
 575                 if (sep)
 576                         ret += fprintf(fp, "%cDelta", *sep);
 577                 else
 578                         ret += fprintf(fp, "  Delta    ");
 579
 580                 if (show_displacement) {
 581                         if (sep)
 582                                 ret += fprintf(fp, "%cDisplacement", *sep);
 583                         else
 584                                 ret += fprintf(fp, " Displ");
 585                 }
 586         }
 587
 588         list_for_each_entry(se, &hist_entry__sort_list, list) {
 589                 if (se->elide)
 590                         continue;
 591                 if (sep) {
 592                         fprintf(fp, "%c%s", *sep, se->header);
 593                         continue;
 594                 }
 595                 width = strlen(se->header);
 596                 if (se->width) {
 597                         if (symbol_conf.col_width_list_str) {
 598                                 if (col_width) {
 599                                         *se->width = atoi(col_width);
 600                                         col_width = strchr(col_width, ',');
 601                                         if (col_width)
 602                                                 ++col_width;
 603                                 }
 604                         }
 605                         width = *se->width = max(*se->width, width);
 606                 }
 607                 fprintf(fp, "  %*s", width, se->header);
 608         }
 609         fprintf(fp, "\n");
 610
 611         if (sep)
 612                 goto print_entries;
 613
 614         fprintf(fp, "# ........");
 615         if (symbol_conf.show_nr_samples)
 616                 fprintf(fp, " ..........");
 617         if (pair) {
 618                 fprintf(fp, " ..........");
 619                 if (show_displacement)
 620                         fprintf(fp, " .....");
 621         }
 622         list_for_each_entry(se, &hist_entry__sort_list, list) {
 623                 unsigned int i;
 624
 625                 if (se->elide)
 626                         continue;
 627
 628                 fprintf(fp, "  ");
 629                 if (se->width)
 630                         width = *se->width;
 631                 else
 632                         width = strlen(se->header);
 633                 for (i = 0; i < width; i++)
 634                         fprintf(fp, ".");
 635         }
 636
 637         fprintf(fp, "\n#\n");
 638
 639 print_entries:
 640         for (nd = rb_first(hists); nd; nd = rb_next(nd)) {
 641                 struct hist_entry *h = rb_entry(nd, struct hist_entry, rb_node);
 642
 643                 if (show_displacement) {
 644                         if (h->pair != NULL)
 645                                 displacement = ((long)h->pair->position -
 646                                                 (long)position);
 647                         else
 648                                 displacement = 0;
 649                         ++position;
 650                 }
 651                 ret += hist_entry__fprintf(h, pair, show_displacement,
 652                                            displacement, fp, session_total);
 653
 654                 if (symbol_conf.use_callchain)
 655                         ret += hist_entry__fprintf_callchain(h, fp, session_total);
 656
 657                 if (h->map == NULL && verbose > 1) {
 658                         __map_groups__fprintf_maps(&h->thread->mg,
 659                                                    MAP__FUNCTION, fp);
 660                         fprintf(fp, "%.10s end\n", graph_dotted_line);
 661                 }
 662         }
 663
 664         free(rem_sq_bracket);
 665
 666         return ret;
 667 }