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; 357 358 if (ct->pct_flags & PMCPL_PCT_TAG) 359 return 0; 360 361 ct->pct_flags |= PMCPL_PCT_TAG; 362 363 if (x >= PMCPL_CT_MAXCOL) { 364 pmcpl_ct_topscreen[x][*y] = NULL; 365 return 1; 366 } 367 pmcpl_ct_topscreen[x][*y] = ct; 368 369 /* 370 * Check if this is a terminal node. 371 * We need to check that some samples exist 372 * for at least one arc for that PMC. 373 */ 374 terminal = 1; 375 for (i = 0; i < ct->pct_narc; i++) 376 if (PMCPL_CT_SAMPLE(pmcin, 377 &ct->pct_arc[i].pcta_samples) != 0) { 378 terminal = 0; 379 break; 380 } 381 382 if (ct->pct_narc == 0 || terminal) { 383 pmcpl_ct_topscreen[x+1][*y] = NULL; 384 if (*y >= PMCPL_CT_MAXLINE) 385 return 1; 386 *y = *y + 1; 387 for (i=0; i < x; i++) 388 pmcpl_ct_topscreen[i][*y] = 389 pmcpl_ct_topscreen[i][*y - 1]; 390 return 0; 391 } 392 393 for (i = 0; i < ct->pct_narc; i++) { 394 if (PMCPL_CT_SAMPLE(pmcin, 395 &ct->pct_arc[i].pcta_samples) == 0) 396 continue; 397 if (PMCPL_CT_SAMPLEP(pmcin, 398 &ct->pct_arc[i].pcta_samples) > pmcstat_threshold) { 399 if (pmcpl_ct_node_dumptop(pmcin, 400 ct->pct_arc[i].pcta_child, 401 rsamples, x+1, y)) 402 return 1; 403 } 404 } 405 406 return 0; 407 } 408 409 /* 410 * Compare two top line by sum. 411 */ 412 static int 413 pmcpl_ct_line_compare(const void *a, const void *b) 414 { 415 const struct pmcpl_ct_line *ct1, *ct2; 416 417 ct1 = (const struct pmcpl_ct_line *) a; 418 ct2 = (const struct pmcpl_ct_line *) b; 419 420 /* Sort in reverse order */ 421 if (ct1->ln_sum < ct2->ln_sum) 422 return (1); 423 if (ct1->ln_sum > ct2->ln_sum) 424 return (-1); 425 return (0); 426 } 427 428 /* 429 * Format and display given PMC index. 430 */ 431 432 static void 433 pmcpl_ct_node_printtop(struct pmcpl_ct_sample *rsamples, int pmcin, int maxy) 434 { 435 #undef TS 436 #undef TSI 437 #define TS(x, y) (pmcpl_ct_topscreen[x][y]) 438 #define TSI(x, y) (pmcpl_ct_topscreen[x][pmcpl_ct_topmax[y].ln_index]) 439 440 int v_attrs, ns_len, vs_len, is_len, width, indentwidth, x, y; 441 float v; 442 char ns[30], vs[10], is[20]; 443 struct pmcpl_ct_node *ct; 444 struct pmcstat_symbol *sym; 445 const char *space = " "; 446 447 /* 448 * Sort by line cost. 449 */ 450 for (y = 0; ; y++) { 451 ct = TS(1, y); 452 if (ct == NULL) 453 break; 454 455 pmcpl_ct_topmax[y].ln_sum = 0; 456 pmcpl_ct_topmax[y].ln_index = y; 457 for (x = 1; TS(x, y) != NULL; x++) { 458 pmcpl_ct_topmax[y].ln_sum += 459 PMCPL_CT_SAMPLE(pmcin, &TS(x, y)->pct_samples); 460 } 461 } 462 qsort(pmcpl_ct_topmax, y, sizeof(pmcpl_ct_topmax[0]), 463 pmcpl_ct_line_compare); 464 pmcpl_ct_topmax[y].ln_index = y; 465 466 for (y = 0; y < maxy; y++) { 467 ct = TSI(1, y); 468 if (ct == NULL) 469 break; 470 471 if (y > 0) 472 PMCSTAT_PRINTW("\n"); 473 474 /* Output sum. */ 475 v = pmcpl_ct_topmax[y].ln_sum * 100.0 / 476 rsamples->sb[pmcin]; 477 snprintf(vs, sizeof(vs), "%.1f", v); 478 v_attrs = PMCSTAT_ATTRPERCENT(v); 479 PMCSTAT_ATTRON(v_attrs); 480 PMCSTAT_PRINTW("%5.5s ", vs); 481 PMCSTAT_ATTROFF(v_attrs); 482 483 width = indentwidth = 5 + 1; 484 485 for (x = 1; (ct = TSI(x, y)) != NULL; x++) { 486 487 vs[0] = '\0'; vs_len = 0; 488 is[0] = '\0'; is_len = 0; 489 490 /* Format value. */ 491 v = PMCPL_CT_SAMPLEP(pmcin, &ct->pct_samples); 492 if (v > pmcstat_threshold) 493 vs_len = snprintf(vs, sizeof(vs), 494 "(%.1f%%)", v); 495 v_attrs = PMCSTAT_ATTRPERCENT(v); 496 497 if (pmcstat_skiplink && v <= pmcstat_threshold) { 498 strlcpy(ns, ".", sizeof(ns)); 499 ns_len = 1; 500 } else { 501 sym = pmcstat_symbol_search(ct->pct_image, ct->pct_func); 502 if (sym != NULL) { 503 ns_len = snprintf(ns, sizeof(ns), "%s", 504 pmcstat_string_unintern(sym->ps_name)); 505 } else 506 ns_len = snprintf(ns, sizeof(ns), "%p", 507 (void *)ct->pct_func); 508 509 /* Format image. */ 510 if (x == 1 || 511 TSI(x-1, y)->pct_image != ct->pct_image) 512 is_len = snprintf(is, sizeof(is), "@%s", 513 pmcstat_string_unintern(ct->pct_image->pi_name)); 514 515 /* Check for line wrap. */ 516 width += ns_len + is_len + vs_len + 1; 517 } 518 if (width >= pmcstat_displaywidth) { 519 maxy--; 520 if (y >= maxy) 521 break; 522 PMCSTAT_PRINTW("\n%*s", indentwidth, space); 523 width = indentwidth + ns_len + is_len + vs_len; 524 } 525 526 PMCSTAT_ATTRON(v_attrs); 527 PMCSTAT_PRINTW("%s%s%s ", ns, is, vs); 528 PMCSTAT_ATTROFF(v_attrs); 529 } 530 } 531 } 532 533 /* 534 * Output top mode snapshot. 535 */ 536 537 void 538 pmcpl_ct_topdisplay(void) 539 { 540 int y; 541 struct pmcpl_ct_sample r, *rsamples; 542 543 rsamples = &r; 544 pmcpl_ct_samples_root(rsamples); 545 546 pmcpl_ct_node_cleartag(); 547 548 PMCSTAT_PRINTW("%5.5s %s\n", "%SAMP", "CALLTREE"); 549 550 y = 0; 551 if (pmcpl_ct_node_dumptop(pmcstat_pmcinfilter, 552 pmcpl_ct_root, rsamples, 0, &y)) 553 PMCSTAT_PRINTW("...\n"); 554 pmcpl_ct_topscreen[1][y] = NULL; 555 556 pmcpl_ct_node_printtop(rsamples, 557 pmcstat_pmcinfilter, pmcstat_displayheight - 2); 558 559 pmcpl_ct_samples_free(rsamples); 560 } 561 562 /* 563 * Handle top mode keypress. 564 */ 565 566 int 567 pmcpl_ct_topkeypress(int c, WINDOW *w) 568 { 569 570 switch (c) { 571 case 'f': 572 pmcstat_skiplink = !pmcstat_skiplink; 573 wprintw(w, "skip empty link %s", pmcstat_skiplink ? "on" : "off"); 574 break; 575 } 576 577 return 0; 578 } 579 580 /* 581 * Look for a callgraph node associated with pmc `pmcid' in the global 582 * hash table that corresponds to the given `pc' value in the process map 583 * `ppm'. 584 */ 585 586 static struct pmcpl_ct_node * 587 pmcpl_ct_node_hash_lookup_pc(struct pmcpl_ct_node *parent, 588 struct pmcstat_pcmap *ppm, uintfptr_t pc, int pmcin) 589 { 590 struct pmcstat_symbol *sym; 591 struct pmcstat_image *image; 592 struct pmcpl_ct_node *ct; 593 struct pmcpl_ct_node_hash *h; 594 struct pmcpl_ct_arc *arc; 595 uintfptr_t loadaddress; 596 int i; 597 unsigned int hash; 598 599 assert(parent != NULL); 600 601 image = ppm->ppm_image; 602 603 loadaddress = ppm->ppm_lowpc + image->pi_vaddr - image->pi_start; 604 pc -= loadaddress; /* Convert to an offset in the image. */ 605 606 /* 607 * Try determine the function at this offset. If we can't 608 * find a function round leave the `pc' value alone. 609 */ 610 if ((sym = pmcstat_symbol_search(image, pc)) != NULL) 611 pc = sym->ps_start; 612 613 for (hash = i = 0; i < (int)sizeof(uintfptr_t); i++) 614 hash += (pc >> i) & 0xFF; 615 616 hash &= PMCSTAT_HASH_MASK; 617 618 ct = NULL; 619 LIST_FOREACH(h, &pmcpl_ct_node_hash[hash], pch_next) { 620 ct = h->pch_ctnode; 621 622 assert(ct != NULL); 623 624 if (ct->pct_image == image && ct->pct_func == pc) { 625 /* 626 * Find related arc in parent node and 627 * increment the sample count. 628 */ 629 for (i = 0; i < parent->pct_narc; i++) { 630 if (parent->pct_arc[i].pcta_child == ct) { 631 arc = &parent->pct_arc[i]; 632 pmcpl_ct_samples_grow(&arc->pcta_samples); 633 arc->pcta_samples.sb[pmcin]++; 634 /* Estimate call count. */ 635 pmcpl_ct_samples_grow(&arc->pcta_callid); 636 if (pmcpl_ct_callid.sb[pmcin] - 637 arc->pcta_callid.sb[pmcin] > 1) 638 arc->pcta_call++; 639 arc->pcta_callid.sb[pmcin] = 640 pmcpl_ct_callid.sb[pmcin]; 641 return (ct); 642 } 643 } 644 645 /* 646 * No arc found for us, add ourself to the parent. 647 */ 648 pmcpl_ct_arc_grow(parent->pct_narc, 649 &parent->pct_arc_c, &parent->pct_arc); 650 arc = &parent->pct_arc[parent->pct_narc]; 651 pmcpl_ct_samples_grow(&arc->pcta_samples); 652 arc->pcta_samples.sb[pmcin] = 1; 653 arc->pcta_call = 1; 654 pmcpl_ct_samples_grow(&arc->pcta_callid); 655 arc->pcta_callid.sb[pmcin] = pmcpl_ct_callid.sb[pmcin]; 656 arc->pcta_child = ct; 657 parent->pct_narc++; 658 return (ct); 659 } 660 } 661 662 /* 663 * We haven't seen this (pmcid, pc) tuple yet, so allocate a 664 * new callgraph node and a new hash table entry for it. 665 */ 666 ct = pmcpl_ct_node_allocate(image, pc); 667 if ((h = malloc(sizeof(*h))) == NULL) 668 err(EX_OSERR, "ERROR: Could not allocate callgraph node"); 669 670 h->pch_ctnode = ct; 671 LIST_INSERT_HEAD(&pmcpl_ct_node_hash[hash], h, pch_next); 672 673 pmcpl_ct_arc_grow(parent->pct_narc, 674 &parent->pct_arc_c, &parent->pct_arc); 675 arc = &parent->pct_arc[parent->pct_narc]; 676 pmcpl_ct_samples_grow(&arc->pcta_samples); 677 arc->pcta_samples.sb[pmcin] = 1; 678 arc->pcta_call = 1; 679 pmcpl_ct_samples_grow(&arc->pcta_callid); 680 arc->pcta_callid.sb[pmcin] = pmcpl_ct_callid.sb[pmcin]; 681 arc->pcta_child = ct; 682 parent->pct_narc++; 683 return (ct); 684 } 685 686 /* 687 * Record a callchain. 688 */ 689 690 void 691 pmcpl_ct_process(struct pmcstat_process *pp, struct pmcstat_pmcrecord *pmcr, 692 uint32_t nsamples, uintfptr_t *cc, int usermode, uint32_t cpu) 693 { 694 int n, pmcin; 695 struct pmcstat_pcmap *ppm[PMC_CALLCHAIN_DEPTH_MAX]; 696 struct pmcstat_process *km; 697 struct pmcpl_ct_node *parent, *child; 698 699 (void) cpu; 700 701 assert(nsamples>0 && nsamples<=PMC_CALLCHAIN_DEPTH_MAX); 702 703 /* Get the PMC index. */ 704 pmcin = pmcr->pr_pmcin; 705 706 /* 707 * Validate mapping for the callchain. 708 * Go from bottom to first invalid entry. 709 */ 710 km = pmcstat_kernproc; 711 for (n = 0; n < (int)nsamples; n++) { 712 ppm[n] = pmcstat_process_find_map(usermode ? 713 pp : km, cc[n]); 714 if (ppm[n] == NULL) { 715 /* Detect full frame capture (kernel + user). */ 716 if (!usermode) { 717 ppm[n] = pmcstat_process_find_map(pp, cc[n]); 718 if (ppm[n] != NULL) 719 km = pp; 720 } 721 } 722 if (ppm[n] == NULL) 723 break; 724 } 725 if (n-- == 0) { 726 pmcstat_stats.ps_callchain_dubious_frames++; 727 pmcr->pr_dubious_frames++; 728 return; 729 } 730 731 /* Increase the call generation counter. */ 732 pmcpl_ct_samples_grow(&pmcpl_ct_callid); 733 pmcpl_ct_callid.sb[pmcin]++; 734 735 /* 736 * Iterate remaining addresses. 737 */ 738 for (parent = pmcpl_ct_root, child = NULL; n >= 0; n--) { 739 child = pmcpl_ct_node_hash_lookup_pc(parent, ppm[n], cc[n], 740 pmcin); 741 if (child == NULL) { 742 pmcstat_stats.ps_callchain_dubious_frames++; 743 continue; 744 } 745 parent = child; 746 } 747 748 /* 749 * Increment the sample count for this PMC. 750 */ 751 if (child != NULL) { 752 pmcpl_ct_samples_grow(&child->pct_samples); 753 child->pct_samples.sb[pmcin]++; 754 755 /* Update per instruction sample if required. */ 756 if (args.pa_ctdumpinstr) 757 pmcpl_ct_instr_add(child, pmcin, cc[0] - 758 (ppm[0]->ppm_lowpc + ppm[0]->ppm_image->pi_vaddr - 759 ppm[0]->ppm_image->pi_start)); 760 } 761 } 762 763 /* 764 * Print node self cost. 765 */ 766 767 static void 768 pmcpl_ct_node_printself(struct pmcpl_ct_node *ct) 769 { 770 int i, j, line; 771 uintptr_t addr; 772 struct pmcstat_symbol *sym; 773 char sourcefile[PATH_MAX]; 774 char funcname[PATH_MAX]; 775 776 /* 777 * Object binary. 778 */ 779 #ifdef PMCPL_CT_OPTIMIZEFN 780 if (pmcpl_ct_prevfn != ct->pct_image->pi_fullpath) { 781 #endif 782 pmcpl_ct_prevfn = ct->pct_image->pi_fullpath; 783 fprintf(args.pa_graphfile, "ob=%s\n", 784 pmcstat_string_unintern(pmcpl_ct_prevfn)); 785 #ifdef PMCPL_CT_OPTIMIZEFN 786 } 787 #endif 788 789 /* 790 * Function name. 791 */ 792 if (pmcstat_image_addr2line(ct->pct_image, ct->pct_func, 793 sourcefile, sizeof(sourcefile), &line, 794 funcname, sizeof(funcname))) { 795 fprintf(args.pa_graphfile, "fn=%s\n", 796 funcname); 797 } else { 798 sym = pmcstat_symbol_search(ct->pct_image, ct->pct_func); 799 if (sym != NULL) 800 fprintf(args.pa_graphfile, "fn=%s\n", 801 pmcstat_string_unintern(sym->ps_name)); 802 else 803 fprintf(args.pa_graphfile, "fn=%p\n", 804 (void *)(ct->pct_image->pi_vaddr + ct->pct_func)); 805 } 806 807 /* 808 * Self cost. 809 */ 810 if (ct->pct_ninstr > 0) { 811 for (i = 0; i < ct->pct_ninstr; i++) { 812 addr = ct->pct_image->pi_vaddr + 813 ct->pct_instr[i].pctf_func; 814 line = 0; 815 if (pmcstat_image_addr2line(ct->pct_image, addr, 816 sourcefile, sizeof(sourcefile), &line, 817 funcname, sizeof(funcname))) 818 fprintf(args.pa_graphfile, "fl=%s\n", sourcefile); 819 fprintf(args.pa_graphfile, "%p %u", (void *)addr, line); 820 for (j = 0; j<pmcstat_npmcs; j++) 821 fprintf(args.pa_graphfile, " %u", 822 PMCPL_CT_SAMPLE(j, 823 &ct->pct_instr[i].pctf_samples)); 824 fprintf(args.pa_graphfile, "\n"); 825 } 826 } else { 827 addr = ct->pct_image->pi_vaddr + ct->pct_func; 828 line = 0; 829 if (pmcstat_image_addr2line(ct->pct_image, addr, 830 sourcefile, sizeof(sourcefile), &line, 831 funcname, sizeof(funcname))) 832 fprintf(args.pa_graphfile, "fl=%s\n", sourcefile); 833 fprintf(args.pa_graphfile, "* *"); 834 for (i = 0; i<pmcstat_npmcs ; i++) 835 fprintf(args.pa_graphfile, " %u", 836 PMCPL_CT_SAMPLE(i, &ct->pct_samples)); 837 fprintf(args.pa_graphfile, "\n"); 838 } 839 } 840 841 /* 842 * Print node child cost. 843 */ 844 845 static void 846 pmcpl_ct_node_printchild(struct pmcpl_ct_node *ct) 847 { 848 int i, j, line; 849 uintptr_t addr; 850 struct pmcstat_symbol *sym; 851 struct pmcpl_ct_node *child; 852 char sourcefile[PATH_MAX]; 853 char funcname[PATH_MAX]; 854 855 /* 856 * Child cost. 857 * TODO: attach child cost to the real position in the funtion. 858 * TODO: cfn=<fn> / call <ncall> addr(<fn>) / addr(call <fn>) <arccost> 859 */ 860 for (i=0 ; i<ct->pct_narc; i++) { 861 child = ct->pct_arc[i].pcta_child; 862 863 /* Object binary. */ 864 #ifdef PMCPL_CT_OPTIMIZEFN 865 if (pmcpl_ct_prevfn != child->pct_image->pi_fullpath) { 866 #endif 867 pmcpl_ct_prevfn = child->pct_image->pi_fullpath; 868 fprintf(args.pa_graphfile, "cob=%s\n", 869 pmcstat_string_unintern(pmcpl_ct_prevfn)); 870 #if PMCPL_CT_OPTIMIZEFN 871 } 872 #endif 873 /* Child function name. */ 874 addr = child->pct_image->pi_vaddr + child->pct_func; 875 /* Child function source file. */ 876 if (pmcstat_image_addr2line(child->pct_image, addr, 877 sourcefile, sizeof(sourcefile), &line, 878 funcname, sizeof(funcname))) { 879 fprintf(args.pa_graphfile, "cfn=%s\n", funcname); 880 fprintf(args.pa_graphfile, "cfl=%s\n", sourcefile); 881 } else { 882 sym = pmcstat_symbol_search(child->pct_image, 883 child->pct_func); 884 if (sym != NULL) 885 fprintf(args.pa_graphfile, "cfn=%s\n", 886 pmcstat_string_unintern(sym->ps_name)); 887 else 888 fprintf(args.pa_graphfile, "cfn=%p\n", (void *)addr); 889 } 890 891 /* Child function address, line and call count. */ 892 fprintf(args.pa_graphfile, "calls=%u %p %u\n", 893 ct->pct_arc[i].pcta_call, (void *)addr, line); 894 895 if (ct->pct_image != NULL) { 896 /* Call address, line, sample. */ 897 addr = ct->pct_image->pi_vaddr + ct->pct_func; 898 line = 0; 899 if (pmcstat_image_addr2line(ct->pct_image, addr, sourcefile, 900 sizeof(sourcefile), &line, 901 funcname, sizeof(funcname))) 902 fprintf(args.pa_graphfile, "%p %u", (void *)addr, line); 903 else 904 fprintf(args.pa_graphfile, "* *"); 905 } 906 else 907 fprintf(args.pa_graphfile, "* *"); 908 for (j = 0; j<pmcstat_npmcs; j++) 909 fprintf(args.pa_graphfile, " %u", 910 PMCPL_CT_SAMPLE(j, &ct->pct_arc[i].pcta_samples)); 911 fprintf(args.pa_graphfile, "\n"); 912 } 913 } 914 915 /* 916 * Clean the PMC name for Kcachegrind formula 917 */ 918 919 static void 920 pmcpl_ct_fixup_pmcname(char *s) 921 { 922 char *p; 923 924 for (p = s; *p; p++) 925 if (!isalnum(*p)) 926 *p = '_'; 927 } 928 929 /* 930 * Print a calltree (KCachegrind) for all PMCs. 931 */ 932 933 static void 934 pmcpl_ct_print(void) 935 { 936 int n, i; 937 struct pmcpl_ct_node_hash *pch; 938 struct pmcpl_ct_sample rsamples; 939 char name[40]; 940 941 pmcpl_ct_samples_root(&rsamples); 942 pmcpl_ct_prevfn = NULL; 943 944 fprintf(args.pa_graphfile, 945 "version: 1\n" 946 "creator: pmcstat\n" 947 "positions: instr line\n" 948 "events:"); 949 for (i=0; i<pmcstat_npmcs; i++) { 950 snprintf(name, sizeof(name), "%s_%d", 951 pmcstat_pmcindex_to_name(i), i); 952 pmcpl_ct_fixup_pmcname(name); 953 fprintf(args.pa_graphfile, " %s", name); 954 } 955 fprintf(args.pa_graphfile, "\nsummary:"); 956 for (i=0; i<pmcstat_npmcs ; i++) 957 fprintf(args.pa_graphfile, " %u", 958 PMCPL_CT_SAMPLE(i, &rsamples)); 959 fprintf(args.pa_graphfile, "\n\n"); 960 961 /* 962 * Fake root node 963 */ 964 fprintf(args.pa_graphfile, "ob=FreeBSD\n"); 965 fprintf(args.pa_graphfile, "fn=ROOT\n"); 966 fprintf(args.pa_graphfile, "* *"); 967 for (i = 0; i<pmcstat_npmcs ; i++) 968 fprintf(args.pa_graphfile, " 0"); 969 fprintf(args.pa_graphfile, "\n"); 970 pmcpl_ct_node_printchild(pmcpl_ct_root); 971 972 for (n = 0; n < PMCSTAT_NHASH; n++) 973 LIST_FOREACH(pch, &pmcpl_ct_node_hash[n], pch_next) { 974 pmcpl_ct_node_printself(pch->pch_ctnode); 975 pmcpl_ct_node_printchild(pch->pch_ctnode); 976 } 977 978 pmcpl_ct_samples_free(&rsamples); 979 } 980 981 int 982 pmcpl_ct_configure(char *opt) 983 { 984 985 if (strncmp(opt, "skiplink=", 9) == 0) { 986 pmcstat_skiplink = atoi(opt+9); 987 } else 988 return (0); 989 990 return (1); 991 } 992 993 int 994 pmcpl_ct_init(void) 995 { 996 int i; 997 998 pmcpl_ct_prevfn = NULL; 999 pmcpl_ct_root = pmcpl_ct_node_allocate(NULL, 0); 1000 1001 for (i = 0; i < PMCSTAT_NHASH; i++) 1002 LIST_INIT(&pmcpl_ct_node_hash[i]); 1003 1004 pmcpl_ct_samples_init(&pmcpl_ct_callid); 1005 1006 return (0); 1007 } 1008 1009 void 1010 pmcpl_ct_shutdown(FILE *mf) 1011 { 1012 int i; 1013 struct pmcpl_ct_node_hash *pch, *pchtmp; 1014 1015 (void) mf; 1016 1017 if (args.pa_flags & FLAG_DO_CALLGRAPHS) 1018 pmcpl_ct_print(); 1019 1020 /* 1021 * Free memory. 1022 */ 1023 1024 for (i = 0; i < PMCSTAT_NHASH; i++) { 1025 LIST_FOREACH_SAFE(pch, &pmcpl_ct_node_hash[i], pch_next, 1026 pchtmp) { 1027 pmcpl_ct_node_free(pch->pch_ctnode); 1028 free(pch); 1029 } 1030 } 1031 1032 pmcpl_ct_node_free(pmcpl_ct_root); 1033 pmcpl_ct_root = NULL; 1034 1035 pmcpl_ct_samples_free(&pmcpl_ct_callid); 1036 } 1037 1038