usr.sbin/pmcstat/pmcpl_calltree.c

   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        256
 124 struct pmcpl_ct_node  *pmcpl_ct_topscreen[PMCPL_CT_MAXCOL][PMCPL_CT_MAXLINE];
 125
 126 /*
 127  * All nodes indexed by function/image name are placed in a hash table.
 128  */
 129 static LIST_HEAD(,pmcpl_ct_node_hash) pmcpl_ct_node_hash[PMCSTAT_NHASH];
 130
 131 /*
 132  * Root node for the graph.
 133  */
 134 static struct pmcpl_ct_node *pmcpl_ct_root;
 135
 136 /*
 137  * Prototypes
 138  */
 139
 140 /*
 141  * Initialize a samples.
 142  */
 143
 144 static void
 145 pmcpl_ct_samples_init(struct pmcpl_ct_sample *samples)
 146 {
 147
 148         samples->npmcs = 0;
 149         samples->sb = NULL;
 150 }
 151
 152 /*
 153  * Free a samples.
 154  */
 155
 156 static void
 157 pmcpl_ct_samples_free(struct pmcpl_ct_sample *samples)
 158 {
 159
 160         samples->npmcs = 0;
 161         free(samples->sb);
 162         samples->sb = NULL;
 163 }
 164
 165 /*
 166  * Grow a sample block to store pmcstat_npmcs PMCs.
 167  */
 168
 169 static void
 170 pmcpl_ct_samples_grow(struct pmcpl_ct_sample *samples)
 171 {
 172         int npmcs;
 173
 174         /* Enough storage. */
 175         if (pmcstat_npmcs <= samples->npmcs)
 176                 return;
 177
 178         npmcs = samples->npmcs +
 179             max(pmcstat_npmcs - samples->npmcs, PMCPL_CT_GROWSIZE);
 180         samples->sb = realloc(samples->sb, npmcs * sizeof(unsigned));
 181         if (samples->sb == NULL)
 182                 errx(EX_SOFTWARE, "ERROR: out of memory");
 183         bzero((char *)samples->sb + samples->npmcs * sizeof(unsigned),
 184             (npmcs - samples->npmcs) * sizeof(unsigned));
 185         samples->npmcs = npmcs;
 186 }
 187
 188 /*
 189  * Compute the sum of all root arcs.
 190  */
 191
 192 static void
 193 pmcpl_ct_samples_root(struct pmcpl_ct_sample *samples)
 194 {
 195         int i, pmcin;
 196
 197         pmcpl_ct_samples_init(samples);
 198         pmcpl_ct_samples_grow(samples);
 199
 200         for (i = 0; i < pmcpl_ct_root->pct_narc; i++)
 201                 for (pmcin = 0; pmcin < pmcstat_npmcs; pmcin++)
 202                         samples->sb[pmcin] += PMCPL_CT_SAMPLE(pmcin,
 203                             &pmcpl_ct_root->pct_arc[i].pcta_samples);
 204 }
 205
 206 /*
 207  * Grow the arc table.
 208  */
 209
 210 static void
 211 pmcpl_ct_arc_grow(int cursize, int *maxsize, struct pmcpl_ct_arc **items)
 212 {
 213         int nmaxsize;
 214
 215         if (cursize < *maxsize)
 216                 return;
 217
 218         nmaxsize = *maxsize + max(cursize + 1 - *maxsize, PMCPL_CT_GROWSIZE);
 219         *items = realloc(*items, nmaxsize * sizeof(struct pmcpl_ct_arc));
 220         if (*items == NULL)
 221                 errx(EX_SOFTWARE, "ERROR: out of memory");
 222         bzero((char *)*items + *maxsize * sizeof(struct pmcpl_ct_arc),
 223             (nmaxsize - *maxsize) * sizeof(struct pmcpl_ct_arc));
 224         *maxsize = nmaxsize;
 225 }
 226
 227 /*
 228  * Compare two arc by samples value.
 229  */
 230 static int
 231 pmcpl_ct_arc_compare(void *thunk, const void *a, const void *b)
 232 {
 233         const struct pmcpl_ct_arc *ct1, *ct2;
 234         int pmcin = *(int *)thunk;
 235
 236         ct1 = (const struct pmcpl_ct_arc *) a;
 237         ct2 = (const struct pmcpl_ct_arc *) b;
 238
 239         /* Sort in reverse order */
 240         if (PMCPL_CT_SAMPLE(pmcin, &ct1->pcta_samples) <
 241             PMCPL_CT_SAMPLE(pmcin, &ct2->pcta_samples))
 242                 return (1);
 243         if (PMCPL_CT_SAMPLE(pmcin, &ct1->pcta_samples) >
 244             PMCPL_CT_SAMPLE(pmcin, &ct2->pcta_samples))
 245                 return (-1);
 246         return (0);
 247 }
 248
 249 /*
 250  * Grow the instr table.
 251  */
 252
 253 static void
 254 pmcpl_ct_instr_grow(int cursize, int *maxsize, struct pmcpl_ct_instr **items)
 255 {
 256         int nmaxsize;
 257
 258         if (cursize < *maxsize)
 259                 return;
 260
 261         nmaxsize = *maxsize + max(cursize + 1 - *maxsize, PMCPL_CT_GROWSIZE);
 262         *items = realloc(*items, nmaxsize * sizeof(struct pmcpl_ct_instr));
 263         if (*items == NULL)
 264                 errx(EX_SOFTWARE, "ERROR: out of memory");
 265         bzero((char *)*items + *maxsize * sizeof(struct pmcpl_ct_instr),
 266             (nmaxsize - *maxsize) * sizeof(struct pmcpl_ct_instr));
 267         *maxsize = nmaxsize;
 268 }
 269
 270 /*
 271  * Add a new instruction sample to given node.
 272  */
 273
 274 static void
 275 pmcpl_ct_instr_add(struct pmcpl_ct_node *ct, int pmcin, uintfptr_t pc)
 276 {
 277         int i;
 278         struct pmcpl_ct_instr *in;
 279
 280         for (i = 0; i<ct->pct_ninstr; i++) {
 281                 if (ct->pct_instr[i].pctf_func == pc) {
 282                         in = &ct->pct_instr[i];
 283                         pmcpl_ct_samples_grow(&in->pctf_samples);
 284                         in->pctf_samples.sb[pmcin]++;
 285                         return;
 286                 }
 287         }
 288
 289         pmcpl_ct_instr_grow(ct->pct_ninstr, &ct->pct_instr_c, &ct->pct_instr);
 290         in = &ct->pct_instr[ct->pct_ninstr];
 291         in->pctf_func = pc;
 292         pmcpl_ct_samples_init(&in->pctf_samples);
 293         pmcpl_ct_samples_grow(&in->pctf_samples);
 294         in->pctf_samples.sb[pmcin] = 1;
 295         ct->pct_ninstr++;
 296 }
 297
 298 /*
 299  * Allocate a new node.
 300  */
 301
 302 static struct pmcpl_ct_node *
 303 pmcpl_ct_node_allocate(struct pmcstat_image *image, uintfptr_t pc)
 304 {
 305         struct pmcpl_ct_node *ct;
 306
 307         if ((ct = malloc(sizeof(*ct))) == NULL)
 308                 err(EX_OSERR, "ERROR: Cannot allocate callgraph node");
 309
 310         ct->pct_flags   = 0;
 311         ct->pct_image   = image;
 312         ct->pct_func    = pc;
 313
 314         pmcpl_ct_samples_init(&ct->pct_samples);
 315
 316         ct->pct_narc    = 0;
 317         ct->pct_arc_c   = 0;
 318         ct->pct_arc     = NULL;
 319
 320         ct->pct_ninstr  = 0;
 321         ct->pct_instr_c = 0;
 322         ct->pct_instr   = NULL;
 323
 324         return (ct);
 325 }
 326
 327 /*
 328  * Free a node.
 329  */
 330
 331 static void
 332 pmcpl_ct_node_free(struct pmcpl_ct_node *ct)
 333 {
 334         int i;
 335
 336         for (i = 0; i < ct->pct_narc; i++) {
 337                 pmcpl_ct_samples_free(&ct->pct_arc[i].pcta_samples);
 338                 pmcpl_ct_samples_free(&ct->pct_arc[i].pcta_callid);
 339         }
 340
 341         pmcpl_ct_samples_free(&ct->pct_samples);
 342         free(ct->pct_arc);
 343         free(ct->pct_instr);
 344         free(ct);
 345 }
 346
 347 /*
 348  * Clear the graph tag on each node.
 349  */
 350 static void
 351 pmcpl_ct_node_cleartag(void)
 352 {
 353         int i;
 354         struct pmcpl_ct_node_hash *pch;
 355
 356         for (i = 0; i < PMCSTAT_NHASH; i++)
 357                 LIST_FOREACH(pch, &pmcpl_ct_node_hash[i], pch_next)
 358                         pch->pch_ctnode->pct_flags &= ~PMCPL_PCT_TAG;
 359
 360         pmcpl_ct_root->pct_flags &= ~PMCPL_PCT_TAG;
 361 }
 362
 363 /*
 364  * Print the callchain line by line with maximum cost at top.
 365  */
 366
 367 static int
 368 pmcpl_ct_node_dumptop(int pmcin, struct pmcpl_ct_node *ct,
 369     struct pmcpl_ct_sample *rsamples, int x, int *y, int maxy)
 370 {
 371         int i;
 372
 373         if (ct->pct_flags & PMCPL_PCT_TAG)
 374                 return 0;
 375
 376         ct->pct_flags |= PMCPL_PCT_TAG;
 377
 378         if (x >= PMCPL_CT_MAXCOL) {
 379                 pmcpl_ct_topscreen[x][*y] = NULL;
 380                 return 1;
 381         }
 382         pmcpl_ct_topscreen[x][*y] = ct;
 383
 384         /*
 385          * This is a terminal node
 386          */
 387         if (ct->pct_narc == 0) {
 388                 pmcpl_ct_topscreen[x+1][*y] = NULL;
 389                 if (*y >= PMCPL_CT_MAXLINE ||
 390                     *y >= maxy)
 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         /*
 400          * Quicksort the arcs.
 401          */
 402         qsort_r(ct->pct_arc, ct->pct_narc, sizeof(struct pmcpl_ct_arc),
 403             &pmcin, pmcpl_ct_arc_compare);
 404
 405         for (i = 0; i < ct->pct_narc; i++) {
 406                 if (PMCPL_CT_SAMPLEP(pmcin,
 407                     &ct->pct_arc[i].pcta_samples) > pmcstat_threshold) {
 408                         if (pmcpl_ct_node_dumptop(pmcin,
 409                                 ct->pct_arc[i].pcta_child,
 410                                 rsamples, x+1, y, maxy))
 411                                 return 1;
 412                 }
 413         }
 414
 415         return 0;
 416 }
 417
 418 /*
 419  * Format and display given PMC index.
 420  */
 421
 422 static void
 423 pmcpl_ct_node_printtop(struct pmcpl_ct_sample *rsamples, int pmcin, int maxy)
 424 {
 425         int v_attrs, ns_len, vs_len, is_len, width, indentwidth, x, y;
 426         float v;
 427         char ns[30], vs[10], is[20];
 428         struct pmcpl_ct_node *ct;
 429         struct pmcstat_symbol *sym;
 430         const char *space = " ";
 431
 432         for (y = 0; y < maxy; y++) {
 433                 /* Output image. */
 434                 ct = pmcpl_ct_topscreen[0][y];
 435                 snprintf(is, sizeof(is), "%-10.10s",
 436                     pmcstat_string_unintern(ct->pct_image->pi_name));
 437                 PMCSTAT_PRINTW("%s ", is);
 438                 width = indentwidth = 11;
 439
 440                 for (x = 0; pmcpl_ct_topscreen[x][y] !=NULL; x++) {
 441
 442                         ct = pmcpl_ct_topscreen[x][y];
 443
 444                         ns[0] = '\0'; ns_len = 0;
 445                         vs[0] = '\0'; vs_len = 0;
 446                         is[0] = '\0'; is_len = 0;
 447
 448                         /* Format value. */
 449                         v = PMCPL_CT_SAMPLEP(pmcin, &ct->pct_samples);
 450                         if (v > pmcstat_threshold)
 451                                 vs_len  = snprintf(vs, sizeof(vs), "(%.1f%%)", v);
 452                         v_attrs = PMCSTAT_ATTRPERCENT(v);
 453
 454                         if (pmcstat_skiplink && v <= pmcstat_threshold) {
 455                                 PMCSTAT_PRINTW(". ");
 456                                 width += 2;
 457                                 continue;
 458                         }
 459                         sym = pmcstat_symbol_search(ct->pct_image, ct->pct_func);
 460                         if (sym != NULL) {
 461                                 ns_len = snprintf(ns, sizeof(ns), "%s",
 462                                     pmcstat_string_unintern(sym->ps_name));
 463                         } else
 464                                 ns_len = snprintf(ns, sizeof(ns), "%p",
 465                                     (void *)ct->pct_func);
 466
 467                         /* Format image. */
 468                         if (x > 0 && pmcpl_ct_topscreen[x-1][y]->pct_image != ct->pct_image)
 469                                 is_len = snprintf(is, sizeof(is), "@%s",
 470                                     pmcstat_string_unintern(ct->pct_image->pi_name));
 471
 472                         /* Check for line wrap. */
 473                         width += ns_len + is_len + vs_len + 1;
 474                         if (width >= pmcstat_displaywidth) {
 475                                 maxy--;
 476                                 if (y >= maxy)
 477                                         break;
 478                                 PMCSTAT_PRINTW("\n%*s", indentwidth, space);
 479                                 width = indentwidth + ns_len + is_len + vs_len;
 480                         }
 481
 482                         PMCSTAT_ATTRON(v_attrs);
 483                         PMCSTAT_PRINTW("%s%s%s ", ns, is, vs);
 484                         PMCSTAT_ATTROFF(v_attrs);
 485                 }
 486                 PMCSTAT_PRINTW("\n");
 487         }
 488 }
 489
 490 /*
 491  * Output top mode snapshot.
 492  */
 493
 494 void
 495 pmcpl_ct_topdisplay(void)
 496 {
 497         int i, x, y, pmcin;
 498         struct pmcpl_ct_sample rsamples;
 499
 500         pmcpl_ct_samples_root(&rsamples);
 501
 502         PMCSTAT_PRINTW("%-10.10s %s\n", "IMAGE", "CALLTREE");
 503
 504         for (pmcin = 0; pmcin < pmcstat_npmcs; pmcin++) {
 505                 /* Filter PMCs. */
 506                 if (pmcstat_pmcinfilter != pmcin)
 507                         continue;
 508
 509                 pmcpl_ct_node_cleartag();
 510
 511                 /* Quicksort the arcs. */
 512                 qsort_r(pmcpl_ct_root->pct_arc,
 513                     pmcpl_ct_root->pct_narc,
 514                     sizeof(struct pmcpl_ct_arc),
 515                     &pmcin, pmcpl_ct_arc_compare);
 516
 517                 x = y = 0;
 518                 for (i = 0; i < pmcpl_ct_root->pct_narc; i++) {
 519                         if (pmcpl_ct_node_dumptop(pmcin,
 520                                 pmcpl_ct_root->pct_arc[i].pcta_child,
 521                                 &rsamples, x, &y, pmcstat_displayheight - 2)) {
 522                                 break;
 523                         }
 524                 }
 525
 526                 pmcpl_ct_node_printtop(&rsamples, pmcin, y);
 527         }
 528         pmcpl_ct_samples_free(&rsamples);
 529 }
 530
 531 /*
 532  * Handle top mode keypress.
 533  */
 534
 535 int
 536 pmcpl_ct_topkeypress(int c, WINDOW *w)
 537 {
 538
 539         switch (c) {
 540         case 'f':
 541                 pmcstat_skiplink = !pmcstat_skiplink;
 542                 wprintw(w, "skip empty link %s", pmcstat_skiplink ? "on" : "off");
 543                 break;
 544         }
 545
 546         return 0;
 547 }
 548
 549 /*
 550  * Look for a callgraph node associated with pmc `pmcid' in the global
 551  * hash table that corresponds to the given `pc' value in the process map
 552  * `ppm'.
 553  */
 554
 555 static struct pmcpl_ct_node *
 556 pmcpl_ct_node_hash_lookup_pc(struct pmcpl_ct_node *parent,
 557     struct pmcstat_pcmap *ppm, uintfptr_t pc, int pmcin)
 558 {
 559         struct pmcstat_symbol *sym;
 560         struct pmcstat_image *image;
 561         struct pmcpl_ct_node *ct;
 562         struct pmcpl_ct_node_hash *h;
 563         struct pmcpl_ct_arc *arc;
 564         uintfptr_t loadaddress;
 565         int i;
 566         unsigned int hash;
 567
 568         assert(parent != NULL);
 569
 570         image = ppm->ppm_image;
 571
 572         loadaddress = ppm->ppm_lowpc + image->pi_vaddr - image->pi_start;
 573         pc -= loadaddress;      /* Convert to an offset in the image. */
 574
 575         /*
 576          * Try determine the function at this offset.  If we can't
 577          * find a function round leave the `pc' value alone.
 578          */
 579         if ((sym = pmcstat_symbol_search(image, pc)) != NULL)
 580                 pc = sym->ps_start;
 581
 582         for (hash = i = 0; i < (int)sizeof(uintfptr_t); i++)
 583                 hash += (pc >> i) & 0xFF;
 584
 585         hash &= PMCSTAT_HASH_MASK;
 586
 587         ct = NULL;
 588         LIST_FOREACH(h, &pmcpl_ct_node_hash[hash], pch_next) {
 589                 ct = h->pch_ctnode;
 590
 591                 assert(ct != NULL);
 592
 593                 if (ct->pct_image == image && ct->pct_func == pc) {
 594                         /*
 595                          * Find related arc in parent node and
 596                          * increment the sample count.
 597                          */
 598                         for (i = 0; i < parent->pct_narc; i++) {
 599                                 if (parent->pct_arc[i].pcta_child == ct) {
 600                                         arc = &parent->pct_arc[i];
 601                                         pmcpl_ct_samples_grow(&arc->pcta_samples);
 602                                         arc->pcta_samples.sb[pmcin]++;
 603                                         /* Estimate call count. */
 604                                         pmcpl_ct_samples_grow(&arc->pcta_callid);
 605                                         if (pmcpl_ct_callid.sb[pmcin] -
 606                                             arc->pcta_callid.sb[pmcin] > 1)
 607                                                 arc->pcta_call++;
 608                                         arc->pcta_callid.sb[pmcin] =
 609                                             pmcpl_ct_callid.sb[pmcin];
 610                                         return (ct);
 611                                 }
 612                         }
 613
 614                         /*
 615                          * No arc found for us, add ourself to the parent.
 616                          */
 617                         pmcpl_ct_arc_grow(parent->pct_narc,
 618                             &parent->pct_arc_c, &parent->pct_arc);
 619                         arc = &parent->pct_arc[parent->pct_narc];
 620                         pmcpl_ct_samples_grow(&arc->pcta_samples);
 621                         arc->pcta_samples.sb[pmcin] = 1;
 622                         arc->pcta_call = 1;
 623                         pmcpl_ct_samples_grow(&arc->pcta_callid);
 624                         arc->pcta_callid.sb[pmcin] = pmcpl_ct_callid.sb[pmcin];
 625                         arc->pcta_child = ct;
 626                         parent->pct_narc++;
 627                         return (ct);
 628                 }
 629         }
 630
 631         /*
 632          * We haven't seen this (pmcid, pc) tuple yet, so allocate a
 633          * new callgraph node and a new hash table entry for it.
 634          */
 635         ct = pmcpl_ct_node_allocate(image, pc);
 636         if ((h = malloc(sizeof(*h))) == NULL)
 637                 err(EX_OSERR, "ERROR: Could not allocate callgraph node");
 638
 639         h->pch_ctnode = ct;
 640         LIST_INSERT_HEAD(&pmcpl_ct_node_hash[hash], h, pch_next);
 641
 642         pmcpl_ct_arc_grow(parent->pct_narc,
 643             &parent->pct_arc_c, &parent->pct_arc);
 644         arc = &parent->pct_arc[parent->pct_narc];
 645         pmcpl_ct_samples_grow(&arc->pcta_samples);
 646         arc->pcta_samples.sb[pmcin] = 1;
 647         arc->pcta_call = 1;
 648         pmcpl_ct_samples_grow(&arc->pcta_callid);
 649         arc->pcta_callid.sb[pmcin] = pmcpl_ct_callid.sb[pmcin];
 650         arc->pcta_child = ct;
 651         parent->pct_narc++;
 652         return (ct);
 653 }
 654
 655 /*
 656  * Record a callchain.
 657  */
 658
 659 void
 660 pmcpl_ct_process(struct pmcstat_process *pp, struct pmcstat_pmcrecord *pmcr,
 661     uint32_t nsamples, uintfptr_t *cc, int usermode, uint32_t cpu)
 662 {
 663         int n, pmcin;
 664         struct pmcstat_pcmap *ppm[PMC_CALLCHAIN_DEPTH_MAX];
 665         struct pmcstat_process *km;
 666         struct pmcpl_ct_node *parent, *child;
 667
 668         (void) cpu;
 669
 670         assert(nsamples>0 && nsamples<=PMC_CALLCHAIN_DEPTH_MAX);
 671
 672         /* Get the PMC index. */
 673         pmcin = pmcr->pr_pmcin;
 674
 675         /*
 676          * Validate mapping for the callchain.
 677          * Go from bottom to first invalid entry.
 678          */
 679         km = pmcstat_kernproc;
 680         for (n = 0; n < (int)nsamples; n++) {
 681                 ppm[n] = pmcstat_process_find_map(usermode ?
 682                     pp : km, cc[n]);
 683                 if (ppm[n] == NULL) {
 684                         /* Detect full frame capture (kernel + user). */
 685                         if (!usermode) {
 686                                 ppm[n] = pmcstat_process_find_map(pp, cc[n]);
 687                                 if (ppm[n] != NULL)
 688                                         km = pp;
 689                         }
 690                 }
 691                 if (ppm[n] == NULL)
 692                         break;
 693         }
 694         if (n-- == 0) {
 695                 pmcstat_stats.ps_callchain_dubious_frames++;
 696                 return;
 697         }
 698
 699         /* Increase the call generation counter. */
 700         pmcpl_ct_samples_grow(&pmcpl_ct_callid);
 701         pmcpl_ct_callid.sb[pmcin]++;
 702
 703         /*
 704          * Iterate remaining addresses.
 705          */
 706         for (parent = pmcpl_ct_root, child = NULL; n >= 0; n--) {
 707                 child = pmcpl_ct_node_hash_lookup_pc(parent, ppm[n], cc[n],
 708                     pmcin);
 709                 if (child == NULL) {
 710                         pmcstat_stats.ps_callchain_dubious_frames++;
 711                         continue;
 712                 }
 713                 parent = child;
 714         }
 715
 716         /*
 717          * Increment the sample count for this PMC.
 718          */
 719         if (child != NULL) {
 720                 pmcpl_ct_samples_grow(&child->pct_samples);
 721                 child->pct_samples.sb[pmcin]++;
 722
 723                 /* Update per instruction sample if required. */
 724                 if (args.pa_ctdumpinstr)
 725                         pmcpl_ct_instr_add(child, pmcin, cc[0] -
 726                             (ppm[0]->ppm_lowpc + ppm[0]->ppm_image->pi_vaddr -
 727                              ppm[0]->ppm_image->pi_start));
 728         }
 729 }
 730
 731 /*
 732  * Print node self cost.
 733  */
 734
 735 static void
 736 pmcpl_ct_node_printself(struct pmcpl_ct_node *ct)
 737 {
 738         int i, j, line;
 739         uintptr_t addr;
 740         struct pmcstat_symbol *sym;
 741         char sourcefile[PATH_MAX];
 742         char funcname[PATH_MAX];
 743
 744         /*
 745          * Object binary.
 746          */
 747 #ifdef PMCPL_CT_OPTIMIZEFN
 748         if (pmcpl_ct_prevfn != ct->pct_image->pi_fullpath) {
 749 #endif
 750                 pmcpl_ct_prevfn = ct->pct_image->pi_fullpath;
 751                 fprintf(args.pa_graphfile, "ob=%s\n",
 752                     pmcstat_string_unintern(pmcpl_ct_prevfn));
 753 #ifdef PMCPL_CT_OPTIMIZEFN
 754         }
 755 #endif
 756
 757         /*
 758          * Function name.
 759          */
 760         if (pmcstat_image_addr2line(ct->pct_image, ct->pct_func,
 761             sourcefile, sizeof(sourcefile), &line,
 762             funcname, sizeof(funcname))) {
 763                 fprintf(args.pa_graphfile, "fn=%s\n",
 764                     funcname);
 765         } else {
 766                 sym = pmcstat_symbol_search(ct->pct_image, ct->pct_func);
 767                 if (sym != NULL)
 768                         fprintf(args.pa_graphfile, "fn=%s\n",
 769                             pmcstat_string_unintern(sym->ps_name));
 770                 else
 771                         fprintf(args.pa_graphfile, "fn=%p\n",
 772                             (void *)(ct->pct_image->pi_vaddr + ct->pct_func));
 773         }
 774
 775         /*
 776          * Self cost.
 777          */
 778         if (ct->pct_ninstr > 0) {
 779                 for (i = 0; i < ct->pct_ninstr; i++) {
 780                         addr = ct->pct_image->pi_vaddr +
 781                             ct->pct_instr[i].pctf_func;
 782                         line = 0;
 783                         if (pmcstat_image_addr2line(ct->pct_image, addr,
 784                             sourcefile, sizeof(sourcefile), &line,
 785                             funcname, sizeof(funcname)))
 786                                 fprintf(args.pa_graphfile, "fl=%s\n", sourcefile);
 787                         fprintf(args.pa_graphfile, "%p %u", (void *)addr, line);
 788                         for (j = 0; j<pmcstat_npmcs; j++)
 789                                 fprintf(args.pa_graphfile, " %u",
 790                                     PMCPL_CT_SAMPLE(j,
 791                                     &ct->pct_instr[i].pctf_samples));
 792                         fprintf(args.pa_graphfile, "\n");
 793                 }
 794         } else {
 795                 addr = ct->pct_image->pi_vaddr + ct->pct_func;
 796                 line = 0;
 797                 if (pmcstat_image_addr2line(ct->pct_image, addr,
 798                     sourcefile, sizeof(sourcefile), &line,
 799                     funcname, sizeof(funcname)))
 800                         fprintf(args.pa_graphfile, "fl=%s\n", sourcefile);
 801                 fprintf(args.pa_graphfile, "* *");
 802                 for (i = 0; i<pmcstat_npmcs ; i++)
 803                         fprintf(args.pa_graphfile, " %u",
 804                             PMCPL_CT_SAMPLE(i, &ct->pct_samples));
 805                 fprintf(args.pa_graphfile, "\n");
 806         }
 807 }
 808
 809 /*
 810  * Print node child cost.
 811  */
 812
 813 static void
 814 pmcpl_ct_node_printchild(struct pmcpl_ct_node *ct)
 815 {
 816         int i, j, line;
 817         uintptr_t addr;
 818         struct pmcstat_symbol *sym;
 819         struct pmcpl_ct_node *child;
 820         char sourcefile[PATH_MAX];
 821         char funcname[PATH_MAX];
 822
 823         /*
 824          * Child cost.
 825          * TODO: attach child cost to the real position in the funtion.
 826          * TODO: cfn=<fn> / call <ncall> addr(<fn>) / addr(call <fn>) <arccost>
 827          */
 828         for (i=0 ; i<ct->pct_narc; i++) {
 829                 child = ct->pct_arc[i].pcta_child;
 830
 831                 /* Object binary. */
 832 #ifdef PMCPL_CT_OPTIMIZEFN
 833                 if (pmcpl_ct_prevfn != child->pct_image->pi_fullpath) {
 834 #endif
 835                         pmcpl_ct_prevfn = child->pct_image->pi_fullpath;
 836                         fprintf(args.pa_graphfile, "cob=%s\n",
 837                             pmcstat_string_unintern(pmcpl_ct_prevfn));
 838 #if PMCPL_CT_OPTIMIZEFN
 839                 }
 840 #endif
 841                 /* Child function name. */
 842                 addr = child->pct_image->pi_vaddr + child->pct_func;
 843                 /* Child function source file. */
 844                 if (pmcstat_image_addr2line(child->pct_image, addr,
 845                     sourcefile, sizeof(sourcefile), &line,
 846                     funcname, sizeof(funcname))) {
 847                         fprintf(args.pa_graphfile, "cfn=%s\n", funcname);
 848                         fprintf(args.pa_graphfile, "cfl=%s\n", sourcefile);
 849                 } else {
 850                         sym = pmcstat_symbol_search(child->pct_image,
 851                             child->pct_func);
 852                         if (sym != NULL)
 853                                 fprintf(args.pa_graphfile, "cfn=%s\n",
 854                                     pmcstat_string_unintern(sym->ps_name));
 855                         else
 856                                 fprintf(args.pa_graphfile, "cfn=%p\n", (void *)addr);
 857                 }
 858
 859                 /* Child function address, line and call count. */
 860                 fprintf(args.pa_graphfile, "calls=%u %p %u\n",
 861                     ct->pct_arc[i].pcta_call, (void *)addr, line);
 862
 863                 if (ct->pct_image != NULL) {
 864                         /* Call address, line, sample. */
 865                         addr = ct->pct_image->pi_vaddr + ct->pct_func;
 866                         line = 0;
 867                         pmcstat_image_addr2line(ct->pct_image, addr, sourcefile,
 868                             sizeof(sourcefile), &line,
 869                             funcname, sizeof(funcname));
 870                         fprintf(args.pa_graphfile, "%p %u", (void *)addr, line);
 871                 }
 872                 else
 873                         fprintf(args.pa_graphfile, "* *");
 874                 for (j = 0; j<pmcstat_npmcs; j++)
 875                         fprintf(args.pa_graphfile, " %u",
 876                             PMCPL_CT_SAMPLE(j, &ct->pct_arc[i].pcta_samples));
 877                 fprintf(args.pa_graphfile, "\n");
 878         }
 879 }
 880
 881 /*
 882  * Clean the PMC name for Kcachegrind formula
 883  */
 884
 885 static void
 886 pmcpl_ct_fixup_pmcname(char *s)
 887 {
 888         char *p;
 889
 890         for (p = s; *p; p++)
 891                 if (!isalnum(*p))
 892                         *p = '_';
 893 }
 894
 895 /*
 896  * Print a calltree (KCachegrind) for all PMCs.
 897  */
 898
 899 static void
 900 pmcpl_ct_print(void)
 901 {
 902         int n, i;
 903         struct pmcpl_ct_node_hash *pch;
 904         struct pmcpl_ct_sample rsamples;
 905         char name[40];
 906
 907         pmcpl_ct_samples_root(&rsamples);
 908         pmcpl_ct_prevfn = NULL;
 909
 910         fprintf(args.pa_graphfile,
 911                 "version: 1\n"
 912                 "creator: pmcstat\n"
 913                 "positions: instr line\n"
 914                 "events:");
 915         for (i=0; i<pmcstat_npmcs; i++) {
 916                 snprintf(name, sizeof(name), "%s_%d",
 917                     pmcstat_pmcindex_to_name(i), i);
 918                 pmcpl_ct_fixup_pmcname(name);
 919                 fprintf(args.pa_graphfile, " %s", name);
 920         }
 921         fprintf(args.pa_graphfile, "\nsummary:");
 922         for (i=0; i<pmcstat_npmcs ; i++)
 923                 fprintf(args.pa_graphfile, " %u",
 924                     PMCPL_CT_SAMPLE(i, &rsamples));
 925         fprintf(args.pa_graphfile, "\n\n");
 926
 927         /*
 928          * Fake root node
 929          */
 930         fprintf(args.pa_graphfile, "ob=FreeBSD\n");
 931         fprintf(args.pa_graphfile, "fn=ROOT\n");
 932         fprintf(args.pa_graphfile, "* *");
 933         for (i = 0; i<pmcstat_npmcs ; i++)
 934                 fprintf(args.pa_graphfile, " 0");
 935         fprintf(args.pa_graphfile, "\n");
 936         pmcpl_ct_node_printchild(pmcpl_ct_root);
 937
 938         for (n = 0; n < PMCSTAT_NHASH; n++)
 939                 LIST_FOREACH(pch, &pmcpl_ct_node_hash[n], pch_next) {
 940                         pmcpl_ct_node_printself(pch->pch_ctnode);
 941                         pmcpl_ct_node_printchild(pch->pch_ctnode);
 942         }
 943
 944         pmcpl_ct_samples_free(&rsamples);
 945 }
 946
 947 int
 948 pmcpl_ct_configure(char *opt)
 949 {
 950
 951         if (strncmp(opt, "skiplink=", 9) == 0) {
 952                 pmcstat_skiplink = atoi(opt+9);
 953         } else
 954                 return (0);
 955
 956         return (1);
 957 }
 958
 959 int
 960 pmcpl_ct_init(void)
 961 {
 962         int i;
 963
 964         pmcpl_ct_prevfn = NULL;
 965         pmcpl_ct_root = pmcpl_ct_node_allocate(NULL, 0);
 966
 967         for (i = 0; i < PMCSTAT_NHASH; i++)
 968                 LIST_INIT(&pmcpl_ct_node_hash[i]);
 969
 970         pmcpl_ct_samples_init(&pmcpl_ct_callid);
 971
 972         return (0);
 973 }
 974
 975 void
 976 pmcpl_ct_shutdown(FILE *mf)
 977 {
 978         int i;
 979         struct pmcpl_ct_node_hash *pch, *pchtmp;
 980
 981         (void) mf;
 982
 983         if (args.pa_flags & FLAG_DO_CALLGRAPHS)
 984                 pmcpl_ct_print();
 985
 986         /*
 987          * Free memory.
 988          */
 989
 990         for (i = 0; i < PMCSTAT_NHASH; i++) {
 991                 LIST_FOREACH_SAFE(pch, &pmcpl_ct_node_hash[i], pch_next,
 992                     pchtmp) {
 993                         pmcpl_ct_node_free(pch->pch_ctnode);
 994                         free(pch);
 995                 }
 996         }
 997
 998         pmcpl_ct_node_free(pmcpl_ct_root);
 999         pmcpl_ct_root = NULL;
1000
1001         pmcpl_ct_samples_free(&pmcpl_ct_callid);
1002 }
1003