1 /*- 2 * Copyright (c) 2009, Fabien Thomas 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 */ 26 27 /* 28 * Process hwpmc(4) samples as calltree. 29 * 30 * Output file format compatible with Kcachegrind (kdesdk). 31 * Handle top mode with a sorted tree display. 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include <sys/param.h> 38 #include <sys/endian.h> 39 #include <sys/queue.h> 40 41 #include <assert.h> 42 #include <curses.h> 43 #include <ctype.h> 44 #include <err.h> 45 #include <errno.h> 46 #include <fcntl.h> 47 #include <pmc.h> 48 #include <pmclog.h> 49 #include <sysexits.h> 50 #include <stdint.h> 51 #include <stdio.h> 52 #include <stdlib.h> 53 #include <string.h> 54 #include <unistd.h> 55 #include <sysexits.h> 56 57 #include "pmcstat.h" 58 #include "pmcstat_log.h" 59 #include "pmcstat_top.h" 60 #include "pmcpl_calltree.h" 61 62 #define PMCPL_CT_GROWSIZE 4 63 64 static pmcstat_interned_string pmcpl_ct_prevfn; 65 66 static int pmcstat_skiplink = 0; 67 68 struct pmcpl_ct_node; 69 70 /* Get the sample value for PMC a. */ 71 #define PMCPL_CT_SAMPLE(a, b) \ 72 ((a) < (b)->npmcs ? (b)->sb[a] : 0) 73 74 /* Get the sample value in percent related to rsamples. */ 75 #define PMCPL_CT_SAMPLEP(a, b) \ 76 (PMCPL_CT_SAMPLE(a, b) * 100.0 / rsamples->sb[a]) 77 78 struct pmcpl_ct_sample { 79 int npmcs; /* Max pmc index available. */ 80 unsigned *sb; /* Sample buffer for 0..npmcs. */ 81 }; 82 83 struct pmcpl_ct_arc { 84 struct pmcpl_ct_sample pcta_samples; 85 struct pmcpl_ct_sample pcta_callid; 86 unsigned pcta_call; 87 struct pmcpl_ct_node *pcta_child; 88 }; 89 90 struct pmcpl_ct_instr { 91 uintfptr_t pctf_func; 92 struct pmcpl_ct_sample pctf_samples; 93 }; 94 95 /* 96 * Each calltree node is tracked by a pmcpl_ct_node struct. 97 */ 98 struct pmcpl_ct_node { 99 #define PMCPL_PCT_TAG 0x00000001 /* Loop detection. */ 100 uint32_t pct_flags; 101 struct pmcstat_image *pct_image; 102 uintfptr_t pct_func; 103 struct pmcpl_ct_sample pct_samples; 104 105 int pct_narc; 106 int pct_arc_c; 107 struct pmcpl_ct_arc *pct_arc; 108 109 /* TODO: optimize for large number of items. */ 110 int pct_ninstr; 111 int pct_instr_c; 112 struct pmcpl_ct_instr *pct_instr; 113 }; 114 115 struct pmcpl_ct_node_hash { 116 struct pmcpl_ct_node *pch_ctnode; 117 LIST_ENTRY(pmcpl_ct_node_hash) pch_next; 118 }; 119 120 struct pmcpl_ct_sample pmcpl_ct_callid; 121 122 #define PMCPL_CT_MAXCOL PMC_CALLCHAIN_DEPTH_MAX 123 #define PMCPL_CT_MAXLINE 1024 /* TODO: dynamic. */ 124 125 struct pmcpl_ct_line { 126 unsigned ln_sum; 127 unsigned ln_index; 128 }; 129 130 struct pmcpl_ct_line pmcpl_ct_topmax[PMCPL_CT_MAXLINE+1]; 131 struct pmcpl_ct_node *pmcpl_ct_topscreen[PMCPL_CT_MAXCOL+1][PMCPL_CT_MAXLINE+1]; 132 133 /* 134 * All nodes indexed by function/image name are placed in a hash table. 135 */ 136 static LIST_HEAD(,pmcpl_ct_node_hash) pmcpl_ct_node_hash[PMCSTAT_NHASH]; 137 138 /* 139 * Root node for the graph. 140 */ 141 static struct pmcpl_ct_node *pmcpl_ct_root; 142 143 /* 144 * Prototypes 145 */ 146 147 /* 148 * Initialize a samples. 149 */ 150 151 static void 152 pmcpl_ct_samples_init(struct pmcpl_ct_sample *samples) 153 { 154 155 samples->npmcs = 0; 156 samples->sb = NULL; 157 } 158 159 /* 160 * Free a samples. 161 */ 162 163 static void 164 pmcpl_ct_samples_free(struct pmcpl_ct_sample *samples) 165 { 166 167 samples->npmcs = 0; 168 free(samples->sb); 169 samples->sb = NULL; 170 } 171 172 /* 173 * Grow a sample block to store pmcstat_npmcs PMCs. 174 */ 175 176 static void 177 pmcpl_ct_samples_grow(struct pmcpl_ct_sample *samples) 178 { 179 int npmcs; 180 181 /* Enough storage. */ 182 if (pmcstat_npmcs <= samples->npmcs) 183 return; 184 185 npmcs = samples->npmcs + 186 max(pmcstat_npmcs - samples->npmcs, PMCPL_CT_GROWSIZE); 187 samples->sb = realloc(samples->sb, npmcs * sizeof(unsigned)); 188 if (samples->sb == NULL) 189 errx(EX_SOFTWARE, "ERROR: out of memory"); 190 bzero((char *)samples->sb + samples->npmcs * sizeof(unsigned), 191 (npmcs - samples->npmcs) * sizeof(unsigned)); 192 samples->npmcs = npmcs; 193 } 194 195 /* 196 * Compute the sum of all root arcs. 197 */ 198 199 static void 200 pmcpl_ct_samples_root(struct pmcpl_ct_sample *samples) 201 { 202 int i, pmcin; 203 204 pmcpl_ct_samples_init(samples); 205 pmcpl_ct_samples_grow(samples); 206 207 for (i = 0; i < pmcpl_ct_root->pct_narc; i++) 208 for (pmcin = 0; pmcin < pmcstat_npmcs; pmcin++) 209 samples->sb[pmcin] += PMCPL_CT_SAMPLE(pmcin, 210 &pmcpl_ct_root->pct_arc[i].pcta_samples); 211 } 212 213 /* 214 * Grow the arc table. 215 */ 216 217 static void 218 pmcpl_ct_arc_grow(int cursize, int *maxsize, struct pmcpl_ct_arc **items) 219 { 220 int nmaxsize; 221 222 if (cursize < *maxsize) 223 return; 224 225 nmaxsize = *maxsize + max(cursize + 1 - *maxsize, PMCPL_CT_GROWSIZE); 226 *items = realloc(*items, nmaxsize * sizeof(struct pmcpl_ct_arc)); 227 if (*items == NULL) 228 errx(EX_SOFTWARE, "ERROR: out of memory"); 229 bzero((char *)*items + *maxsize * sizeof(struct pmcpl_ct_arc), 230 (nmaxsize - *maxsize) * sizeof(struct pmcpl_ct_arc)); 231 *maxsize = nmaxsize; 232 } 233 234 /* 235 * Grow the instr table. 236 */ 237 238 static void 239 pmcpl_ct_instr_grow(int cursize, int *maxsize, struct pmcpl_ct_instr **items) 240 { 241 int nmaxsize; 242 243 if (cursize < *maxsize) 244 return; 245 246 nmaxsize = *maxsize + max(cursize + 1 - *maxsize, PMCPL_CT_GROWSIZE); 247 *items = realloc(*items, nmaxsize * sizeof(struct pmcpl_ct_instr)); 248 if (*items == NULL) 249 errx(EX_SOFTWARE, "ERROR: out of memory"); 250 bzero((char *)*items + *maxsize * sizeof(struct pmcpl_ct_instr), 251 (nmaxsize - *maxsize) * sizeof(struct pmcpl_ct_instr)); 252 *maxsize = nmaxsize; 253 } 254 255 /* 256 * Add a new instruction sample to given node. 257 */ 258 259 static void 260 pmcpl_ct_instr_add(struct pmcpl_ct_node *ct, int pmcin, uintfptr_t pc) 261 { 262 int i; 263 struct pmcpl_ct_instr *in; 264 265 for (i = 0; i<ct->pct_ninstr; i++) { 266 if (ct->pct_instr[i].pctf_func == pc) { 267 in = &ct->pct_instr[i]; 268 pmcpl_ct_samples_grow(&in->pctf_samples); 269 in->pctf_samples.sb[pmcin]++; 270 return; 271 } 272 } 273 274 pmcpl_ct_instr_grow(ct->pct_ninstr, &ct->pct_instr_c, &ct->pct_instr); 275 in = &ct->pct_instr[ct->pct_ninstr]; 276 in->pctf_func = pc; 277 pmcpl_ct_samples_init(&in->pctf_samples); 278 pmcpl_ct_samples_grow(&in->pctf_samples); 279 in->pctf_samples.sb[pmcin] = 1; 280 ct->pct_ninstr++; 281 } 282 283 /* 284 * Allocate a new node. 285 */ 286 287 static struct pmcpl_ct_node * 288 pmcpl_ct_node_allocate(struct pmcstat_image *image, uintfptr_t pc) 289 { 290 struct pmcpl_ct_node *ct; 291 292 if ((ct = malloc(sizeof(*ct))) == NULL) 293 err(EX_OSERR, "ERROR: Cannot allocate callgraph node"); 294 295 ct->pct_flags = 0; 296 ct->pct_image = image; 297 ct->pct_func = pc; 298 299 pmcpl_ct_samples_init(&ct->pct_samples); 300 301 ct->pct_narc = 0; 302 ct->pct_arc_c = 0; 303 ct->pct_arc = NULL; 304 305 ct->pct_ninstr = 0; 306 ct->pct_instr_c = 0; 307 ct->pct_instr = NULL; 308 309 return (ct); 310 } 311 312 /* 313 * Free a node. 314 */ 315 316 static void 317 pmcpl_ct_node_free(struct pmcpl_ct_node *ct) 318 { 319 int i; 320 321 for (i = 0; i < ct->pct_narc; i++) { 322 pmcpl_ct_samples_free(&ct->pct_arc[i].pcta_samples); 323 pmcpl_ct_samples_free(&ct->pct_arc[i].pcta_callid); 324 } 325 326 pmcpl_ct_samples_free(&ct->pct_samples); 327 free(ct->pct_arc); 328 free(ct->pct_instr); 329 free(ct); 330 } 331 332 /* 333 * Clear the graph tag on each node. 334 */ 335 static void 336 pmcpl_ct_node_cleartag(void) 337 { 338 int i; 339 struct pmcpl_ct_node_hash *pch; 340 341 for (i = 0; i < PMCSTAT_NHASH; i++) 342 LIST_FOREACH(pch, &pmcpl_ct_node_hash[i], pch_next) 343 pch->pch_ctnode->pct_flags &= ~PMCPL_PCT_TAG; 344 345 pmcpl_ct_root->pct_flags &= ~PMCPL_PCT_TAG; 346 } 347 348 /* 349 * Print the callchain line by line with maximum cost at top. 350 */ 351 352 static int 353 pmcpl_ct_node_dumptop(int pmcin, struct pmcpl_ct_node *ct, 354 struct pmcpl_ct_sample *rsamples, int x, int *y) 355 { 356 int i, terminal; 357 struct pmcpl_ct_arc *arc; 358 359 if (ct->pct_flags & PMCPL_PCT_TAG) 360 return 0; 361 362 ct->pct_flags |= PMCPL_PCT_TAG; 363 364 if (x >= PMCPL_CT_MAXCOL) { 365 pmcpl_ct_topscreen[x][*y] = NULL; 366 return 1; 367 } 368 pmcpl_ct_topscreen[x][*y] = ct; 369 370 /* 371 * Check if this is a terminal node. 372 * We need to check that some samples exist 373 * for at least one arc for that PMC. 374 */ 375 terminal = 1; 376 for (i = 0; i < ct->pct_narc; i++) { 377 arc = &ct->pct_arc[i]; 378 if (PMCPL_CT_SAMPLE(pmcin, 379 &arc->pcta_samples) != 0 && 380 PMCPL_CT_SAMPLEP(pmcin, 381 &arc->pcta_samples) > pmcstat_threshold && 382 (arc->pcta_child->pct_flags & PMCPL_PCT_TAG) == 0) { 383 terminal = 0; 384 break; 385 } 386 } 387 388 if (ct->pct_narc == 0 || terminal) { 389 pmcpl_ct_topscreen[x+1][*y] = NULL; 390 if (*y >= PMCPL_CT_MAXLINE) 391 return 1; 392 *y = *y + 1; 393 for (i=0; i < x; i++) 394 pmcpl_ct_topscreen[i][*y] = 395 pmcpl_ct_topscreen[i][*y - 1]; 396 return 0; 397 } 398 399 for (i = 0; i < ct->pct_narc; i++) { 400 if (PMCPL_CT_SAMPLE(pmcin, 401 &ct->pct_arc[i].pcta_samples) == 0) 402 continue; 403 if (PMCPL_CT_SAMPLEP(pmcin, 404 &ct->pct_arc[i].pcta_samples) > pmcstat_threshold) { 405 if (pmcpl_ct_node_dumptop(pmcin, 406 ct->pct_arc[i].pcta_child, 407 rsamples, x+1, y)) 408 return 1; 409 } 410 } 411 412 return 0; 413 } 414 415 /* 416 * Compare two top line by sum. 417 */ 418 static int 419 pmcpl_ct_line_compare(const void *a, const void *b) 420 { 421 const struct pmcpl_ct_line *ct1, *ct2; 422 423 ct1 = (const struct pmcpl_ct_line *) a; 424 ct2 = (const struct pmcpl_ct_line *) b; 425 426 /* Sort in reverse order */ 427 if (ct1->ln_sum < ct2->ln_sum) 428 return (1); 429 if (ct1->ln_sum > ct2->ln_sum) 430 return (-1); 431 return (0); 432 } 433 434 /* 435 * Format and display given PMC index. 436 */ 437 438 static void 439 pmcpl_ct_node_printtop(struct pmcpl_ct_sample *rsamples, int pmcin, int maxy) 440 { 441 #undef TS 442 #undef TSI 443 #define TS(x, y) (pmcpl_ct_topscreen[x][y]) 444 #define TSI(x, y) (pmcpl_ct_topscreen[x][pmcpl_ct_topmax[y].ln_index]) 445 446 int v_attrs, ns_len, vs_len, is_len, width, indentwidth, x, y; 447 float v; 448 char ns[30], vs[10], is[20]; 449 struct pmcpl_ct_node *ct; 450 struct pmcstat_symbol *sym; 451 const char *space = " "; 452 453 /* 454 * Sort by line cost. 455 */ 456 for (y = 0; ; y++) { 457 ct = TS(1, y); 458 if (ct == NULL) 459 break; 460 461 pmcpl_ct_topmax[y].ln_sum = 0; 462 pmcpl_ct_topmax[y].ln_index = y; 463 for (x = 1; TS(x, y) != NULL; x++) { 464 pmcpl_ct_topmax[y].ln_sum += 465 PMCPL_CT_SAMPLE(pmcin, &TS(x, y)->pct_samples); 466 } 467 } 468 qsort(pmcpl_ct_topmax, y, sizeof(pmcpl_ct_topmax[0]), 469 pmcpl_ct_line_compare); 470 pmcpl_ct_topmax[y].ln_index = y; 471 472 for (y = 0; y < maxy; y++) { 473 ct = TSI(1, y); 474 if (ct == NULL) 475 break; 476 477 if (y > 0) 478 PMCSTAT_PRINTW("\n"); 479 480 /* Output sum. */ 481 v = pmcpl_ct_topmax[y].ln_sum * 100.0 / 482 rsamples->sb[pmcin]; 483 snprintf(vs, sizeof(vs), "%.1f", v); 484 v_attrs = PMCSTAT_ATTRPERCENT(v); 485 PMCSTAT_ATTRON(v_attrs); 486 PMCSTAT_PRINTW("%5.5s ", vs); 487 PMCSTAT_ATTROFF(v_attrs); 488 489 width = indentwidth = 5 + 1; 490 491 for (x = 1; (ct = TSI(x, y)) != NULL; x++) { 492 493 vs[0] = '\0'; vs_len = 0; 494 is[0] = '\0'; is_len = 0; 495 496 /* Format value. */ 497 v = PMCPL_CT_SAMPLEP(pmcin, &ct->pct_samples); 498 if (v > pmcstat_threshold) 499 vs_len = snprintf(vs, sizeof(vs), 500 "(%.1f%%)", v); 501 v_attrs = PMCSTAT_ATTRPERCENT(v); 502 503 if (pmcstat_skiplink && v <= pmcstat_threshold) { 504 strlcpy(ns, ".", sizeof(ns)); 505 ns_len = 1; 506 } else { 507 sym = pmcstat_symbol_search(ct->pct_image, ct->pct_func); 508 if (sym != NULL) { 509 ns_len = snprintf(ns, sizeof(ns), "%s", 510 pmcstat_string_unintern(sym->ps_name)); 511 } else 512 ns_len = snprintf(ns, sizeof(ns), "%p", 513 (void *)ct->pct_func); 514 515 /* Format image. */ 516 if (x == 1 || 517 TSI(x-1, y)->pct_image != ct->pct_image) 518 is_len = snprintf(is, sizeof(is), "@%s", 519 pmcstat_string_unintern(ct->pct_image->pi_name)); 520 521 /* Check for line wrap. */ 522 width += ns_len + is_len + vs_len + 1; 523 } 524 if (width >= pmcstat_displaywidth) { 525 maxy--; 526 if (y >= maxy) 527 break; 528 PMCSTAT_PRINTW("\n%*s", indentwidth, space); 529 width = indentwidth + ns_len + is_len + vs_len; 530 } 531 532 PMCSTAT_ATTRON(v_attrs); 533 PMCSTAT_PRINTW("%s%s%s ", ns, is, vs); 534 PMCSTAT_ATTROFF(v_attrs); 535 } 536 } 537 } 538 539 /* 540 * Output top mode snapshot. 541 */ 542 543 void 544 pmcpl_ct_topdisplay(void) 545 { 546 int y; 547 struct pmcpl_ct_sample r, *rsamples; 548 549 rsamples = &r; 550 pmcpl_ct_samples_root(rsamples); 551 552 pmcpl_ct_node_cleartag(); 553 554 PMCSTAT_PRINTW("%5.5s %s\n", "%SAMP", "CALLTREE"); 555 556 y = 0; 557 if (pmcpl_ct_node_dumptop(pmcstat_pmcinfilter, 558 pmcpl_ct_root, rsamples, 0, &y)) 559 PMCSTAT_PRINTW("...\n"); 560 pmcpl_ct_topscreen[1][y] = NULL; 561 562 pmcpl_ct_node_printtop(rsamples, 563 pmcstat_pmcinfilter, pmcstat_displayheight - 2); 564 565 pmcpl_ct_samples_free(rsamples); 566 } 567 568 /* 569 * Handle top mode keypress. 570 */ 571 572 int 573 pmcpl_ct_topkeypress(int c, WINDOW *w) 574 { 575 576 switch (c) { 577 case 'f': 578 pmcstat_skiplink = !pmcstat_skiplink; 579 wprintw(w, "skip empty link %s", pmcstat_skiplink ? "on" : "off"); 580 break; 581 } 582 583 return 0; 584 } 585 586 /* 587 * Look for a callgraph node associated with pmc `pmcid' in the global 588 * hash table that corresponds to the given `pc' value in the process map 589 * `ppm'. 590 */ 591 592 static struct pmcpl_ct_node * 593 pmcpl_ct_node_hash_lookup_pc(struct pmcpl_ct_node *parent, 594 struct pmcstat_pcmap *ppm, uintfptr_t pc, int pmcin) 595 { 596 struct pmcstat_symbol *sym; 597 struct pmcstat_image *image; 598 struct pmcpl_ct_node *ct; 599 struct pmcpl_ct_node_hash *h; 600 struct pmcpl_ct_arc *arc; 601 uintfptr_t loadaddress; 602 int i; 603 unsigned int hash; 604 605 assert(parent != NULL); 606 607 image = ppm->ppm_image; 608 609 loadaddress = ppm->ppm_lowpc + image->pi_vaddr - image->pi_start; 610 pc -= loadaddress; /* Convert to an offset in the image. */ 611 612 /* 613 * Try determine the function at this offset. If we can't 614 * find a function round leave the `pc' value alone. 615 */ 616 if ((sym = pmcstat_symbol_search(image, pc)) != NULL) 617 pc = sym->ps_start; 618 else 619 pmcstat_stats.ps_samples_unknown_function++; 620 621 for (hash = i = 0; i < (int)sizeof(uintfptr_t); i++) 622 hash += (pc >> i) & 0xFF; 623 624 hash &= PMCSTAT_HASH_MASK; 625 626 ct = NULL; 627 LIST_FOREACH(h, &pmcpl_ct_node_hash[hash], pch_next) { 628 ct = h->pch_ctnode; 629 630 assert(ct != NULL); 631 632 if (ct->pct_image == image && ct->pct_func == pc) { 633 /* 634 * Find related arc in parent node and 635 * increment the sample count. 636 */ 637 for (i = 0; i < parent->pct_narc; i++) { 638 if (parent->pct_arc[i].pcta_child == ct) { 639 arc = &parent->pct_arc[i]; 640 pmcpl_ct_samples_grow(&arc->pcta_samples); 641 arc->pcta_samples.sb[pmcin]++; 642 /* Estimate call count. */ 643 pmcpl_ct_samples_grow(&arc->pcta_callid); 644 if (pmcpl_ct_callid.sb[pmcin] - 645 arc->pcta_callid.sb[pmcin] > 1) 646 arc->pcta_call++; 647 arc->pcta_callid.sb[pmcin] = 648 pmcpl_ct_callid.sb[pmcin]; 649 return (ct); 650 } 651 } 652 653 /* 654 * No arc found for us, add ourself to the parent. 655 */ 656 pmcpl_ct_arc_grow(parent->pct_narc, 657 &parent->pct_arc_c, &parent->pct_arc); 658 arc = &parent->pct_arc[parent->pct_narc]; 659 pmcpl_ct_samples_grow(&arc->pcta_samples); 660 arc->pcta_samples.sb[pmcin] = 1; 661 arc->pcta_call = 1; 662 pmcpl_ct_samples_grow(&arc->pcta_callid); 663 arc->pcta_callid.sb[pmcin] = pmcpl_ct_callid.sb[pmcin]; 664 arc->pcta_child = ct; 665 parent->pct_narc++; 666 return (ct); 667 } 668 } 669 670 /* 671 * We haven't seen this (pmcid, pc) tuple yet, so allocate a 672 * new callgraph node and a new hash table entry for it. 673 */ 674 ct = pmcpl_ct_node_allocate(image, pc); 675 if ((h = malloc(sizeof(*h))) == NULL) 676 err(EX_OSERR, "ERROR: Could not allocate callgraph node"); 677 678 h->pch_ctnode = ct; 679 LIST_INSERT_HEAD(&pmcpl_ct_node_hash[hash], h, pch_next); 680 681 pmcpl_ct_arc_grow(parent->pct_narc, 682 &parent->pct_arc_c, &parent->pct_arc); 683 arc = &parent->pct_arc[parent->pct_narc]; 684 pmcpl_ct_samples_grow(&arc->pcta_samples); 685 arc->pcta_samples.sb[pmcin] = 1; 686 arc->pcta_call = 1; 687 pmcpl_ct_samples_grow(&arc->pcta_callid); 688 arc->pcta_callid.sb[pmcin] = pmcpl_ct_callid.sb[pmcin]; 689 arc->pcta_child = ct; 690 parent->pct_narc++; 691 return (ct); 692 } 693 694 /* 695 * Record a callchain. 696 */ 697 698 void 699 pmcpl_ct_process(struct pmcstat_process *pp, struct pmcstat_pmcrecord *pmcr, 700 uint32_t nsamples, uintfptr_t *cc, int usermode, uint32_t cpu) 701 { 702 int n, pmcin; 703 struct pmcstat_pcmap *ppm[PMC_CALLCHAIN_DEPTH_MAX]; 704 struct pmcstat_process *km; 705 struct pmcpl_ct_node *parent, *child; 706 707 (void) cpu; 708 709 assert(nsamples>0 && nsamples<=PMC_CALLCHAIN_DEPTH_MAX); 710 711 /* Get the PMC index. */ 712 pmcin = pmcr->pr_pmcin; 713 714 /* 715 * Validate mapping for the callchain. 716 * Go from bottom to first invalid entry. 717 */ 718 km = pmcstat_kernproc; 719 for (n = 0; n < (int)nsamples; n++) { 720 ppm[n] = pmcstat_process_find_map(usermode ? 721 pp : km, cc[n]); 722 if (ppm[n] == NULL) { 723 /* Detect full frame capture (kernel + user). */ 724 if (!usermode) { 725 ppm[n] = pmcstat_process_find_map(pp, cc[n]); 726 if (ppm[n] != NULL) 727 km = pp; 728 } 729 } 730 if (ppm[n] == NULL) 731 break; 732 } 733 if (n-- == 0) { 734 pmcstat_stats.ps_callchain_dubious_frames++; 735 pmcr->pr_dubious_frames++; 736 return; 737 } 738 739 /* Increase the call generation counter. */ 740 pmcpl_ct_samples_grow(&pmcpl_ct_callid); 741 pmcpl_ct_callid.sb[pmcin]++; 742 743 /* 744 * Iterate remaining addresses. 745 */ 746 for (parent = pmcpl_ct_root, child = NULL; n >= 0; n--) { 747 child = pmcpl_ct_node_hash_lookup_pc(parent, ppm[n], cc[n], 748 pmcin); 749 if (child == NULL) { 750 pmcstat_stats.ps_callchain_dubious_frames++; 751 continue; 752 } 753 parent = child; 754 } 755 756 /* 757 * Increment the sample count for this PMC. 758 */ 759 if (child != NULL) { 760 pmcpl_ct_samples_grow(&child->pct_samples); 761 child->pct_samples.sb[pmcin]++; 762 763 /* Update per instruction sample if required. */ 764 if (args.pa_ctdumpinstr) 765 pmcpl_ct_instr_add(child, pmcin, cc[0] - 766 (ppm[0]->ppm_lowpc + ppm[0]->ppm_image->pi_vaddr - 767 ppm[0]->ppm_image->pi_start)); 768 } 769 } 770 771 /* 772 * Print node self cost. 773 */ 774 775 static void 776 pmcpl_ct_node_printself(struct pmcpl_ct_node *ct) 777 { 778 int i, j, line; 779 uintptr_t addr; 780 struct pmcstat_symbol *sym; 781 char sourcefile[PATH_MAX]; 782 char funcname[PATH_MAX]; 783 784 /* 785 * Object binary. 786 */ 787 #ifdef PMCPL_CT_OPTIMIZEFN 788 if (pmcpl_ct_prevfn != ct->pct_image->pi_fullpath) { 789 #endif 790 pmcpl_ct_prevfn = ct->pct_image->pi_fullpath; 791 fprintf(args.pa_graphfile, "ob=%s\n", 792 pmcstat_string_unintern(pmcpl_ct_prevfn)); 793 #ifdef PMCPL_CT_OPTIMIZEFN 794 } 795 #endif 796 797 /* 798 * Function name. 799 */ 800 if (pmcstat_image_addr2line(ct->pct_image, ct->pct_func, 801 sourcefile, sizeof(sourcefile), &line, 802 funcname, sizeof(funcname))) { 803 fprintf(args.pa_graphfile, "fn=%s\n", 804 funcname); 805 } else { 806 sym = pmcstat_symbol_search(ct->pct_image, ct->pct_func); 807 if (sym != NULL) 808 fprintf(args.pa_graphfile, "fn=%s\n", 809 pmcstat_string_unintern(sym->ps_name)); 810 else 811 fprintf(args.pa_graphfile, "fn=%p\n", 812 (void *)(ct->pct_image->pi_vaddr + ct->pct_func)); 813 } 814 815 /* 816 * Self cost. 817 */ 818 if (ct->pct_ninstr > 0) { 819 for (i = 0; i < ct->pct_ninstr; i++) { 820 addr = ct->pct_image->pi_vaddr + 821 ct->pct_instr[i].pctf_func; 822 line = 0; 823 if (pmcstat_image_addr2line(ct->pct_image, addr, 824 sourcefile, sizeof(sourcefile), &line, 825 funcname, sizeof(funcname))) 826 fprintf(args.pa_graphfile, "fl=%s\n", sourcefile); 827 fprintf(args.pa_graphfile, "%p %u", (void *)addr, line); 828 for (j = 0; j<pmcstat_npmcs; j++) 829 fprintf(args.pa_graphfile, " %u", 830 PMCPL_CT_SAMPLE(j, 831 &ct->pct_instr[i].pctf_samples)); 832 fprintf(args.pa_graphfile, "\n"); 833 } 834 } else { 835 addr = ct->pct_image->pi_vaddr + ct->pct_func; 836 line = 0; 837 if (pmcstat_image_addr2line(ct->pct_image, addr, 838 sourcefile, sizeof(sourcefile), &line, 839 funcname, sizeof(funcname))) 840 fprintf(args.pa_graphfile, "fl=%s\n", sourcefile); 841 fprintf(args.pa_graphfile, "* *"); 842 for (i = 0; i<pmcstat_npmcs ; i++) 843 fprintf(args.pa_graphfile, " %u", 844 PMCPL_CT_SAMPLE(i, &ct->pct_samples)); 845 fprintf(args.pa_graphfile, "\n"); 846 } 847 } 848 849 /* 850 * Print node child cost. 851 */ 852 853 static void 854 pmcpl_ct_node_printchild(struct pmcpl_ct_node *ct) 855 { 856 int i, j, line; 857 uintptr_t addr; 858 struct pmcstat_symbol *sym; 859 struct pmcpl_ct_node *child; 860 char sourcefile[PATH_MAX]; 861 char funcname[PATH_MAX]; 862 863 /* 864 * Child cost. 865 * TODO: attach child cost to the real position in the funtion. 866 * TODO: cfn=<fn> / call <ncall> addr(<fn>) / addr(call <fn>) <arccost> 867 */ 868 for (i=0 ; i<ct->pct_narc; i++) { 869 child = ct->pct_arc[i].pcta_child; 870 871 /* Object binary. */ 872 #ifdef PMCPL_CT_OPTIMIZEFN 873 if (pmcpl_ct_prevfn != child->pct_image->pi_fullpath) { 874 #endif 875 pmcpl_ct_prevfn = child->pct_image->pi_fullpath; 876 fprintf(args.pa_graphfile, "cob=%s\n", 877 pmcstat_string_unintern(pmcpl_ct_prevfn)); 878 #if PMCPL_CT_OPTIMIZEFN 879 } 880 #endif 881 /* Child function name. */ 882 addr = child->pct_image->pi_vaddr + child->pct_func; 883 /* Child function source file. */ 884 if (pmcstat_image_addr2line(child->pct_image, addr, 885 sourcefile, sizeof(sourcefile), &line, 886 funcname, sizeof(funcname))) { 887 fprintf(args.pa_graphfile, "cfn=%s\n", funcname); 888 fprintf(args.pa_graphfile, "cfl=%s\n", sourcefile); 889 } else { 890 sym = pmcstat_symbol_search(child->pct_image, 891 child->pct_func); 892 if (sym != NULL) 893 fprintf(args.pa_graphfile, "cfn=%s\n", 894 pmcstat_string_unintern(sym->ps_name)); 895 else 896 fprintf(args.pa_graphfile, "cfn=%p\n", (void *)addr); 897 } 898 899 /* Child function address, line and call count. */ 900 fprintf(args.pa_graphfile, "calls=%u %p %u\n", 901 ct->pct_arc[i].pcta_call, (void *)addr, line); 902 903 if (ct->pct_image != NULL) { 904 /* Call address, line, sample. */ 905 addr = ct->pct_image->pi_vaddr + ct->pct_func; 906 line = 0; 907 if (pmcstat_image_addr2line(ct->pct_image, addr, sourcefile, 908 sizeof(sourcefile), &line, 909 funcname, sizeof(funcname))) 910 fprintf(args.pa_graphfile, "%p %u", (void *)addr, line); 911 else 912 fprintf(args.pa_graphfile, "* *"); 913 } 914 else 915 fprintf(args.pa_graphfile, "* *"); 916 for (j = 0; j<pmcstat_npmcs; j++) 917 fprintf(args.pa_graphfile, " %u", 918 PMCPL_CT_SAMPLE(j, &ct->pct_arc[i].pcta_samples)); 919 fprintf(args.pa_graphfile, "\n"); 920 } 921 } 922 923 /* 924 * Clean the PMC name for Kcachegrind formula 925 */ 926 927 static void 928 pmcpl_ct_fixup_pmcname(char *s) 929 { 930 char *p; 931 932 for (p = s; *p; p++) 933 if (!isalnum(*p)) 934 *p = '_'; 935 } 936 937 /* 938 * Print a calltree (KCachegrind) for all PMCs. 939 */ 940 941 static void 942 pmcpl_ct_print(void) 943 { 944 int n, i; 945 struct pmcpl_ct_node_hash *pch; 946 struct pmcpl_ct_sample rsamples; 947 char name[40]; 948 949 pmcpl_ct_samples_root(&rsamples); 950 pmcpl_ct_prevfn = NULL; 951 952 fprintf(args.pa_graphfile, 953 "version: 1\n" 954 "creator: pmcstat\n" 955 "positions: instr line\n" 956 "events:"); 957 for (i=0; i<pmcstat_npmcs; i++) { 958 snprintf(name, sizeof(name), "%s_%d", 959 pmcstat_pmcindex_to_name(i), i); 960 pmcpl_ct_fixup_pmcname(name); 961 fprintf(args.pa_graphfile, " %s", name); 962 } 963 fprintf(args.pa_graphfile, "\nsummary:"); 964 for (i=0; i<pmcstat_npmcs ; i++) 965 fprintf(args.pa_graphfile, " %u", 966 PMCPL_CT_SAMPLE(i, &rsamples)); 967 fprintf(args.pa_graphfile, "\n\n"); 968 969 /* 970 * Fake root node 971 */ 972 fprintf(args.pa_graphfile, "ob=FreeBSD\n"); 973 fprintf(args.pa_graphfile, "fn=ROOT\n"); 974 fprintf(args.pa_graphfile, "* *"); 975 for (i = 0; i<pmcstat_npmcs ; i++) 976 fprintf(args.pa_graphfile, " 0"); 977 fprintf(args.pa_graphfile, "\n"); 978 pmcpl_ct_node_printchild(pmcpl_ct_root); 979 980 for (n = 0; n < PMCSTAT_NHASH; n++) 981 LIST_FOREACH(pch, &pmcpl_ct_node_hash[n], pch_next) { 982 pmcpl_ct_node_printself(pch->pch_ctnode); 983 pmcpl_ct_node_printchild(pch->pch_ctnode); 984 } 985 986 pmcpl_ct_samples_free(&rsamples); 987 } 988 989 int 990 pmcpl_ct_configure(char *opt) 991 { 992 993 if (strncmp(opt, "skiplink=", 9) == 0) { 994 pmcstat_skiplink = atoi(opt+9); 995 } else 996 return (0); 997 998 return (1); 999 } 1000 1001 int 1002 pmcpl_ct_init(void) 1003 { 1004 int i; 1005 1006 pmcpl_ct_prevfn = NULL; 1007 pmcpl_ct_root = pmcpl_ct_node_allocate(NULL, 0); 1008 1009 for (i = 0; i < PMCSTAT_NHASH; i++) 1010 LIST_INIT(&pmcpl_ct_node_hash[i]); 1011 1012 pmcpl_ct_samples_init(&pmcpl_ct_callid); 1013 1014 return (0); 1015 } 1016 1017 void 1018 pmcpl_ct_shutdown(FILE *mf) 1019 { 1020 int i; 1021 struct pmcpl_ct_node_hash *pch, *pchtmp; 1022 1023 (void) mf; 1024 1025 if (args.pa_flags & FLAG_DO_CALLGRAPHS) 1026 pmcpl_ct_print(); 1027 1028 /* 1029 * Free memory. 1030 */ 1031 1032 for (i = 0; i < PMCSTAT_NHASH; i++) { 1033 LIST_FOREACH_SAFE(pch, &pmcpl_ct_node_hash[i], pch_next, 1034 pchtmp) { 1035 pmcpl_ct_node_free(pch->pch_ctnode); 1036 free(pch); 1037 } 1038 } 1039 1040 pmcpl_ct_node_free(pmcpl_ct_root); 1041 pmcpl_ct_root = NULL; 1042 1043 pmcpl_ct_samples_free(&pmcpl_ct_callid); 1044 } 1045 1046