usr.sbin/pmcstat/pmcstat_log.c

   1 /*-
   2  * Copyright (c) 2005-2007, Joseph Koshy
   3  * Copyright (c) 2007 The FreeBSD Foundation
   4  * All rights reserved.
   5  *
   6  * Portions of this software were developed by A. Joseph Koshy under
   7  * sponsorship from the FreeBSD Foundation and Google, Inc.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  */
  30
  31 /*
  32  * Transform a hwpmc(4) log into human readable form, and into
  33  * gprof(1) compatible profiles.
  34  */
  35
  36 #include <sys/cdefs.h>
  37 __FBSDID("$FreeBSD$");
  38
  39 #include <sys/param.h>
  40 #include <sys/endian.h>
  41 #include <sys/cpuset.h>
  42 #include <sys/gmon.h>
  43 #include <sys/imgact_aout.h>
  44 #include <sys/imgact_elf.h>
  45 #include <sys/mman.h>
  46 #include <sys/pmc.h>
  47 #include <sys/queue.h>
  48 #include <sys/socket.h>
  49 #include <sys/stat.h>
  50 #include <sys/wait.h>
  51
  52 #include <netinet/in.h>
  53
  54 #include <assert.h>
  55 #include <curses.h>
  56 #include <err.h>
  57 #include <errno.h>
  58 #include <fcntl.h>
  59 #include <gelf.h>
  60 #include <libgen.h>
  61 #include <limits.h>
  62 #include <netdb.h>
  63 #include <pmc.h>
  64 #include <pmclog.h>
  65 #include <sysexits.h>
  66 #include <stdint.h>
  67 #include <stdio.h>
  68 #include <stdlib.h>
  69 #include <string.h>
  70 #include <unistd.h>
  71
  72 #include "pmcstat.h"
  73 #include "pmcstat_log.h"
  74 #include "pmcstat_top.h"
  75
  76 #define PMCSTAT_ALLOCATE                1
  77
  78 /*
  79  * PUBLIC INTERFACES
  80  *
  81  * pmcstat_initialize_logging() initialize this module, called first
  82  * pmcstat_shutdown_logging()           orderly shutdown, called last
  83  * pmcstat_open_log()                   open an eventlog for processing
  84  * pmcstat_process_log()                print/convert an event log
  85  * pmcstat_display_log()                top mode display for the log
  86  * pmcstat_close_log()                  finish processing an event log
  87  *
  88  * IMPLEMENTATION NOTES
  89  *
  90  * We correlate each 'callchain' or 'sample' entry seen in the event
  91  * log back to an executable object in the system. Executable objects
  92  * include:
  93  *      - program executables,
  94  *      - shared libraries loaded by the runtime loader,
  95  *      - dlopen()'ed objects loaded by the program,
  96  *      - the runtime loader itself,
  97  *      - the kernel and kernel modules.
  98  *
  99  * Each process that we know about is treated as a set of regions that
 100  * map to executable objects.  Processes are described by
 101  * 'pmcstat_process' structures.  Executable objects are tracked by
 102  * 'pmcstat_image' structures.  The kernel and kernel modules are
 103  * common to all processes (they reside at the same virtual addresses
 104  * for all processes).  Individual processes can have their text
 105  * segments and shared libraries loaded at process-specific locations.
 106  *
 107  * A given executable object can be in use by multiple processes
 108  * (e.g., libc.so) and loaded at a different address in each.
 109  * pmcstat_pcmap structures track per-image mappings.
 110  *
 111  * The sample log could have samples from multiple PMCs; we
 112  * generate one 'gmon.out' profile per PMC.
 113  *
 114  * IMPLEMENTATION OF GMON OUTPUT
 115  *
 116  * Each executable object gets one 'gmon.out' profile, per PMC in
 117  * use.  Creation of 'gmon.out' profiles is done lazily.  The
 118  * 'gmon.out' profiles generated for a given sampling PMC are
 119  * aggregates of all the samples for that particular executable
 120  * object.
 121  *
 122  * IMPLEMENTATION OF SYSTEM-WIDE CALLGRAPH OUTPUT
 123  *
 124  * Each active pmcid has its own callgraph structure, described by a
 125  * 'struct pmcstat_callgraph'.  Given a process id and a list of pc
 126  * values, we map each pc value to a tuple (image, symbol), where
 127  * 'image' denotes an executable object and 'symbol' is the closest
 128  * symbol that precedes the pc value.  Each pc value in the list is
 129  * also given a 'rank' that reflects its depth in the call stack.
 130  */
 131
 132 struct pmcstat_pmcs pmcstat_pmcs = LIST_HEAD_INITIALIZER(pmcstat_pmcs);
 133
 134 /*
 135  * All image descriptors are kept in a hash table.
 136  */
 137 struct pmcstat_image_hash_list pmcstat_image_hash[PMCSTAT_NHASH];
 138
 139 /*
 140  * All process descriptors are kept in a hash table.
 141  */
 142 struct pmcstat_process_hash_list pmcstat_process_hash[PMCSTAT_NHASH];
 143
 144 struct pmcstat_stats pmcstat_stats; /* statistics */
 145 static int ps_samples_period; /* samples count between top refresh. */
 146
 147 struct pmcstat_process *pmcstat_kernproc; /* kernel 'process' */
 148
 149 #include "pmcpl_gprof.h"
 150 #include "pmcpl_callgraph.h"
 151 #include "pmcpl_annotate.h"
 152 #include "pmcpl_calltree.h"
 153
 154 static struct pmc_plugins  {
 155         const char      *pl_name;       /* name */
 156
 157         /* configure */
 158         int (*pl_configure)(char *opt);
 159
 160         /* init and shutdown */
 161         int (*pl_init)(void);
 162         void (*pl_shutdown)(FILE *mf);
 163
 164         /* sample processing */
 165         void (*pl_process)(struct pmcstat_process *pp,
 166             struct pmcstat_pmcrecord *pmcr, uint32_t nsamples,
 167             uintfptr_t *cc, int usermode, uint32_t cpu);
 168
 169         /* image */
 170         void (*pl_initimage)(struct pmcstat_image *pi);
 171         void (*pl_shutdownimage)(struct pmcstat_image *pi);
 172
 173         /* pmc */
 174         void (*pl_newpmc)(pmcstat_interned_string ps,
 175                 struct pmcstat_pmcrecord *pr);
 176
 177         /* top display */
 178         void (*pl_topdisplay)(void);
 179
 180         /* top keypress */
 181         int (*pl_topkeypress)(int c, WINDOW *w);
 182
 183 } plugins[] = {
 184         {
 185                 .pl_name                = "none",
 186         },
 187         {
 188                 .pl_name                = "callgraph",
 189                 .pl_init                = pmcpl_cg_init,
 190                 .pl_shutdown            = pmcpl_cg_shutdown,
 191                 .pl_process             = pmcpl_cg_process,
 192                 .pl_topkeypress         = pmcpl_cg_topkeypress,
 193                 .pl_topdisplay          = pmcpl_cg_topdisplay
 194         },
 195         {
 196                 .pl_name                = "gprof",
 197                 .pl_shutdown            = pmcpl_gmon_shutdown,
 198                 .pl_process             = pmcpl_gmon_process,
 199                 .pl_initimage           = pmcpl_gmon_initimage,
 200                 .pl_shutdownimage       = pmcpl_gmon_shutdownimage,
 201                 .pl_newpmc              = pmcpl_gmon_newpmc
 202         },
 203         {
 204                 .pl_name                = "annotate",
 205                 .pl_process             = pmcpl_annotate_process
 206         },
 207         {
 208                 .pl_name                = "calltree",
 209                 .pl_configure           = pmcpl_ct_configure,
 210                 .pl_init                = pmcpl_ct_init,
 211                 .pl_shutdown            = pmcpl_ct_shutdown,
 212                 .pl_process             = pmcpl_ct_process,
 213                 .pl_topkeypress         = pmcpl_ct_topkeypress,
 214                 .pl_topdisplay          = pmcpl_ct_topdisplay
 215         },
 216         {
 217                 .pl_name                = NULL
 218         }
 219 };
 220
 221 static int pmcstat_mergepmc;
 222
 223 int pmcstat_pmcinfilter = 0; /* PMC filter for top mode. */
 224 float pmcstat_threshold = 0.5; /* Cost filter for top mode. */
 225
 226 /*
 227  * Prototypes
 228  */
 229
 230 static struct pmcstat_image *pmcstat_image_from_path(pmcstat_interned_string
 231     _path, int _iskernelmodule);
 232 static void pmcstat_image_get_aout_params(struct pmcstat_image *_image);
 233 static void pmcstat_image_get_elf_params(struct pmcstat_image *_image);
 234 static void     pmcstat_image_link(struct pmcstat_process *_pp,
 235     struct pmcstat_image *_i, uintfptr_t _lpc);
 236
 237 static void     pmcstat_pmcid_add(pmc_id_t _pmcid,
 238     pmcstat_interned_string _name);
 239
 240 static void     pmcstat_process_aout_exec(struct pmcstat_process *_pp,
 241     struct pmcstat_image *_image, uintfptr_t _entryaddr);
 242 static void     pmcstat_process_elf_exec(struct pmcstat_process *_pp,
 243     struct pmcstat_image *_image, uintfptr_t _entryaddr);
 244 static void     pmcstat_process_exec(struct pmcstat_process *_pp,
 245     pmcstat_interned_string _path, uintfptr_t _entryaddr);
 246 static struct pmcstat_process *pmcstat_process_lookup(pid_t _pid,
 247     int _allocate);
 248 static int      pmcstat_string_compute_hash(const char *_string);
 249 static void pmcstat_string_initialize(void);
 250 static int      pmcstat_string_lookup_hash(pmcstat_interned_string _is);
 251 static void pmcstat_string_shutdown(void);
 252 static void pmcstat_stats_reset(int _reset_global);
 253
 254 /*
 255  * A simple implementation of interned strings.  Each interned string
 256  * is assigned a unique address, so that subsequent string compares
 257  * can be done by a simple pointer comparison instead of using
 258  * strcmp().  This speeds up hash table lookups and saves memory if
 259  * duplicate strings are the norm.
 260  */
 261 struct pmcstat_string {
 262         LIST_ENTRY(pmcstat_string)      ps_next;        /* hash link */
 263         int             ps_len;
 264         int             ps_hash;
 265         char            *ps_string;
 266 };
 267
 268 static LIST_HEAD(,pmcstat_string)       pmcstat_string_hash[PMCSTAT_NHASH];
 269
 270 /*
 271  * PMC count.
 272  */
 273 int pmcstat_npmcs;
 274
 275 /*
 276  * PMC Top mode pause state.
 277  */
 278 static int pmcstat_pause;
 279
 280 static void
 281 pmcstat_stats_reset(int reset_global)
 282 {
 283         struct pmcstat_pmcrecord *pr;
 284
 285         /* Flush PMCs stats. */
 286         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
 287                 pr->pr_samples = 0;
 288                 pr->pr_dubious_frames = 0;
 289         }
 290         ps_samples_period = 0;
 291
 292         /* Flush global stats. */
 293         if (reset_global)
 294                 bzero(&pmcstat_stats, sizeof(struct pmcstat_stats));
 295 }
 296
 297 /*
 298  * Compute a 'hash' value for a string.
 299  */
 300
 301 static int
 302 pmcstat_string_compute_hash(const char *s)
 303 {
 304         int hash;
 305
 306         for (hash = 0; *s; s++)
 307                 hash ^= *s;
 308
 309         return (hash & PMCSTAT_HASH_MASK);
 310 }
 311
 312 /*
 313  * Intern a copy of string 's', and return a pointer to the
 314  * interned structure.
 315  */
 316
 317 pmcstat_interned_string
 318 pmcstat_string_intern(const char *s)
 319 {
 320         struct pmcstat_string *ps;
 321         const struct pmcstat_string *cps;
 322         int hash, len;
 323
 324         if ((cps = pmcstat_string_lookup(s)) != NULL)
 325                 return (cps);
 326
 327         hash = pmcstat_string_compute_hash(s);
 328         len  = strlen(s);
 329
 330         if ((ps = malloc(sizeof(*ps))) == NULL)
 331                 err(EX_OSERR, "ERROR: Could not intern string");
 332         ps->ps_len = len;
 333         ps->ps_hash = hash;
 334         ps->ps_string = strdup(s);
 335         LIST_INSERT_HEAD(&pmcstat_string_hash[hash], ps, ps_next);
 336         return ((pmcstat_interned_string) ps);
 337 }
 338
 339 const char *
 340 pmcstat_string_unintern(pmcstat_interned_string str)
 341 {
 342         const char *s;
 343
 344         s = ((const struct pmcstat_string *) str)->ps_string;
 345         return (s);
 346 }
 347
 348 pmcstat_interned_string
 349 pmcstat_string_lookup(const char *s)
 350 {
 351         struct pmcstat_string *ps;
 352         int hash, len;
 353
 354         hash = pmcstat_string_compute_hash(s);
 355         len = strlen(s);
 356
 357         LIST_FOREACH(ps, &pmcstat_string_hash[hash], ps_next)
 358             if (ps->ps_len == len && ps->ps_hash == hash &&
 359                 strcmp(ps->ps_string, s) == 0)
 360                     return (ps);
 361         return (NULL);
 362 }
 363
 364 static int
 365 pmcstat_string_lookup_hash(pmcstat_interned_string s)
 366 {
 367         const struct pmcstat_string *ps;
 368
 369         ps = (const struct pmcstat_string *) s;
 370         return (ps->ps_hash);
 371 }
 372
 373 /*
 374  * Initialize the string interning facility.
 375  */
 376
 377 static void
 378 pmcstat_string_initialize(void)
 379 {
 380         int i;
 381
 382         for (i = 0; i < PMCSTAT_NHASH; i++)
 383                 LIST_INIT(&pmcstat_string_hash[i]);
 384 }
 385
 386 /*
 387  * Destroy the string table, free'ing up space.
 388  */
 389
 390 static void
 391 pmcstat_string_shutdown(void)
 392 {
 393         int i;
 394         struct pmcstat_string *ps, *pstmp;
 395
 396         for (i = 0; i < PMCSTAT_NHASH; i++)
 397                 LIST_FOREACH_SAFE(ps, &pmcstat_string_hash[i], ps_next,
 398                     pstmp) {
 399                         LIST_REMOVE(ps, ps_next);
 400                         free(ps->ps_string);
 401                         free(ps);
 402                 }
 403 }
 404
 405 /*
 406  * Determine whether a given executable image is an A.OUT object, and
 407  * if so, fill in its parameters from the text file.
 408  * Sets image->pi_type.
 409  */
 410
 411 static void
 412 pmcstat_image_get_aout_params(struct pmcstat_image *image)
 413 {
 414         int fd;
 415         ssize_t nbytes;
 416         struct exec ex;
 417         const char *path;
 418         char buffer[PATH_MAX];
 419
 420         path = pmcstat_string_unintern(image->pi_execpath);
 421         assert(path != NULL);
 422
 423         if (image->pi_iskernelmodule)
 424                 errx(EX_SOFTWARE,
 425                     "ERROR: a.out kernel modules are unsupported \"%s\"", path);
 426
 427         (void) snprintf(buffer, sizeof(buffer), "%s%s",
 428             args.pa_fsroot, path);
 429
 430         if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
 431             (nbytes = read(fd, &ex, sizeof(ex))) < 0) {
 432                 if (args.pa_verbosity >= 2)
 433                         warn("WARNING: Cannot determine type of \"%s\"",
 434                             path);
 435                 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
 436                 if (fd != -1)
 437                         (void) close(fd);
 438                 return;
 439         }
 440
 441         (void) close(fd);
 442
 443         if ((unsigned) nbytes != sizeof(ex) ||
 444             N_BADMAG(ex))
 445                 return;
 446
 447         image->pi_type = PMCSTAT_IMAGE_AOUT;
 448
 449         /* TODO: the rest of a.out processing */
 450
 451         return;
 452 }
 453
 454 /*
 455  * Helper function.
 456  */
 457
 458 static int
 459 pmcstat_symbol_compare(const void *a, const void *b)
 460 {
 461         const struct pmcstat_symbol *sym1, *sym2;
 462
 463         sym1 = (const struct pmcstat_symbol *) a;
 464         sym2 = (const struct pmcstat_symbol *) b;
 465
 466         if (sym1->ps_end <= sym2->ps_start)
 467                 return (-1);
 468         if (sym1->ps_start >= sym2->ps_end)
 469                 return (1);
 470         return (0);
 471 }
 472
 473 /*
 474  * Map an address to a symbol in an image.
 475  */
 476
 477 struct pmcstat_symbol *
 478 pmcstat_symbol_search(struct pmcstat_image *image, uintfptr_t addr)
 479 {
 480         struct pmcstat_symbol sym;
 481
 482         if (image->pi_symbols == NULL)
 483                 return (NULL);
 484
 485         sym.ps_name  = NULL;
 486         sym.ps_start = addr;
 487         sym.ps_end   = addr + 1;
 488
 489         return (bsearch((void *) &sym, image->pi_symbols,
 490                     image->pi_symcount, sizeof(struct pmcstat_symbol),
 491                     pmcstat_symbol_compare));
 492 }
 493
 494 /*
 495  * Add the list of symbols in the given section to the list associated
 496  * with the object.
 497  */
 498 static void
 499 pmcstat_image_add_symbols(struct pmcstat_image *image, Elf *e,
 500     Elf_Scn *scn, GElf_Shdr *sh)
 501 {
 502         int firsttime;
 503         size_t n, newsyms, nshsyms, nfuncsyms;
 504         struct pmcstat_symbol *symptr;
 505         char *fnname;
 506         GElf_Sym sym;
 507         Elf_Data *data;
 508
 509         if ((data = elf_getdata(scn, NULL)) == NULL)
 510                 return;
 511
 512         /*
 513          * Determine the number of functions named in this
 514          * section.
 515          */
 516
 517         nshsyms = sh->sh_size / sh->sh_entsize;
 518         for (n = nfuncsyms = 0; n < nshsyms; n++) {
 519                 if (gelf_getsym(data, (int) n, &sym) != &sym)
 520                         return;
 521                 if (GELF_ST_TYPE(sym.st_info) == STT_FUNC)
 522                         nfuncsyms++;
 523         }
 524
 525         if (nfuncsyms == 0)
 526                 return;
 527
 528         /*
 529          * Allocate space for the new entries.
 530          */
 531         firsttime = image->pi_symbols == NULL;
 532         symptr = realloc(image->pi_symbols,
 533             sizeof(*symptr) * (image->pi_symcount + nfuncsyms));
 534         if (symptr == image->pi_symbols) /* realloc() failed. */
 535                 return;
 536         image->pi_symbols = symptr;
 537
 538         /*
 539          * Append new symbols to the end of the current table.
 540          */
 541         symptr += image->pi_symcount;
 542
 543         for (n = newsyms = 0; n < nshsyms; n++) {
 544                 if (gelf_getsym(data, (int) n, &sym) != &sym)
 545                         return;
 546                 if (GELF_ST_TYPE(sym.st_info) != STT_FUNC)
 547                         continue;
 548                 if (sym.st_shndx == STN_UNDEF)
 549                         continue;
 550
 551                 if (!firsttime && pmcstat_symbol_search(image, sym.st_value))
 552                         continue; /* We've seen this symbol already. */
 553
 554                 if ((fnname = elf_strptr(e, sh->sh_link, sym.st_name))
 555                     == NULL)
 556                         continue;
 557 #ifdef __arm__
 558                 /* Remove spurious ARM function name. */
 559                 if (fnname[0] == '$' &&
 560                     (fnname[1] == 'a' || fnname[1] == 't' ||
 561                     fnname[1] == 'd') &&
 562                     fnname[2] == '\0')
 563                         continue;
 564 #endif
 565
 566                 symptr->ps_name  = pmcstat_string_intern(fnname);
 567                 symptr->ps_start = sym.st_value - image->pi_vaddr;
 568                 symptr->ps_end   = symptr->ps_start + sym.st_size;
 569                 symptr++;
 570
 571                 newsyms++;
 572         }
 573
 574         image->pi_symcount += newsyms;
 575         if (image->pi_symcount == 0)
 576                 return;
 577
 578         assert(newsyms <= nfuncsyms);
 579
 580         /*
 581          * Return space to the system if there were duplicates.
 582          */
 583         if (newsyms < nfuncsyms)
 584                 image->pi_symbols = realloc(image->pi_symbols,
 585                     sizeof(*symptr) * image->pi_symcount);
 586
 587         /*
 588          * Keep the list of symbols sorted.
 589          */
 590         qsort(image->pi_symbols, image->pi_symcount, sizeof(*symptr),
 591             pmcstat_symbol_compare);
 592
 593         /*
 594          * Deal with function symbols that have a size of 'zero' by
 595          * making them extend to the next higher address.  These
 596          * symbols are usually defined in assembly code.
 597          */
 598         for (symptr = image->pi_symbols;
 599              symptr < image->pi_symbols + (image->pi_symcount - 1);
 600              symptr++)
 601                 if (symptr->ps_start == symptr->ps_end)
 602                         symptr->ps_end = (symptr+1)->ps_start;
 603 }
 604
 605 /*
 606  * Examine an ELF file to determine the size of its text segment.
 607  * Sets image->pi_type if anything conclusive can be determined about
 608  * this image.
 609  */
 610
 611 static void
 612 pmcstat_image_get_elf_params(struct pmcstat_image *image)
 613 {
 614         int fd;
 615         size_t i, nph, nsh;
 616         const char *path, *elfbase;
 617         char *p, *endp;
 618         uintfptr_t minva, maxva;
 619         Elf *e;
 620         Elf_Scn *scn;
 621         GElf_Ehdr eh;
 622         GElf_Phdr ph;
 623         GElf_Shdr sh;
 624         enum pmcstat_image_type image_type;
 625         char buffer[PATH_MAX];
 626
 627         assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
 628
 629         image->pi_start = minva = ~(uintfptr_t) 0;
 630         image->pi_end = maxva = (uintfptr_t) 0;
 631         image->pi_type = image_type = PMCSTAT_IMAGE_INDETERMINABLE;
 632         image->pi_isdynamic = 0;
 633         image->pi_dynlinkerpath = NULL;
 634         image->pi_vaddr = 0;
 635
 636         path = pmcstat_string_unintern(image->pi_execpath);
 637         assert(path != NULL);
 638
 639         /*
 640          * Look for kernel modules under FSROOT/KERNELPATH/NAME,
 641          * and user mode executable objects under FSROOT/PATHNAME.
 642          */
 643         if (image->pi_iskernelmodule)
 644                 (void) snprintf(buffer, sizeof(buffer), "%s%s/%s",
 645                     args.pa_fsroot, args.pa_kernel, path);
 646         else
 647                 (void) snprintf(buffer, sizeof(buffer), "%s%s",
 648                     args.pa_fsroot, path);
 649
 650         e = NULL;
 651         if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
 652             (e = elf_begin(fd, ELF_C_READ, NULL)) == NULL ||
 653             (elf_kind(e) != ELF_K_ELF)) {
 654                 if (args.pa_verbosity >= 2)
 655                         warnx("WARNING: Cannot determine the type of \"%s\".",
 656                             buffer);
 657                 goto done;
 658         }
 659
 660         if (gelf_getehdr(e, &eh) != &eh) {
 661                 warnx(
 662                     "WARNING: Cannot retrieve the ELF Header for \"%s\": %s.",
 663                     buffer, elf_errmsg(-1));
 664                 goto done;
 665         }
 666
 667         if (eh.e_type != ET_EXEC && eh.e_type != ET_DYN &&
 668             !(image->pi_iskernelmodule && eh.e_type == ET_REL)) {
 669                 warnx("WARNING: \"%s\" is of an unsupported ELF type.",
 670                     buffer);
 671                 goto done;
 672         }
 673
 674         image_type = eh.e_ident[EI_CLASS] == ELFCLASS32 ?
 675             PMCSTAT_IMAGE_ELF32 : PMCSTAT_IMAGE_ELF64;
 676
 677         /*
 678          * Determine the virtual address where an executable would be
 679          * loaded.  Additionally, for dynamically linked executables,
 680          * save the pathname to the runtime linker.
 681          */
 682         if (eh.e_type == ET_EXEC) {
 683                 if (elf_getphnum(e, &nph) == 0) {
 684                         warnx(
 685 "WARNING: Could not determine the number of program headers in \"%s\": %s.",
 686                             buffer,
 687                             elf_errmsg(-1));
 688                         goto done;
 689                 }
 690                 for (i = 0; i < eh.e_phnum; i++) {
 691                         if (gelf_getphdr(e, i, &ph) != &ph) {
 692                                 warnx(
 693 "WARNING: Retrieval of PHDR entry #%ju in \"%s\" failed: %s.",
 694                                     (uintmax_t) i, buffer, elf_errmsg(-1));
 695                                 goto done;
 696                         }
 697                         switch (ph.p_type) {
 698                         case PT_DYNAMIC:
 699                                 image->pi_isdynamic = 1;
 700                                 break;
 701                         case PT_INTERP:
 702                                 if ((elfbase = elf_rawfile(e, NULL)) == NULL) {
 703                                         warnx(
 704 "WARNING: Cannot retrieve the interpreter for \"%s\": %s.",
 705                                             buffer, elf_errmsg(-1));
 706                                         goto done;
 707                                 }
 708                                 image->pi_dynlinkerpath =
 709                                     pmcstat_string_intern(elfbase +
 710                                         ph.p_offset);
 711                                 break;
 712                         case PT_LOAD:
 713                                 if ((ph.p_offset & (-ph.p_align)) == 0)
 714                                         image->pi_vaddr = ph.p_vaddr & (-ph.p_align);
 715                                 break;
 716                         }
 717                 }
 718         }
 719
 720         /*
 721          * Get the min and max VA associated with this ELF object.
 722          */
 723         if (elf_getshnum(e, &nsh) == 0) {
 724                 warnx(
 725 "WARNING: Could not determine the number of sections for \"%s\": %s.",
 726                     buffer, elf_errmsg(-1));
 727                 goto done;
 728         }
 729
 730         for (i = 0; i < nsh; i++) {
 731                 if ((scn = elf_getscn(e, i)) == NULL ||
 732                     gelf_getshdr(scn, &sh) != &sh) {
 733                         warnx(
 734 "WARNING: Could not retrieve section header #%ju in \"%s\": %s.",
 735                             (uintmax_t) i, buffer, elf_errmsg(-1));
 736                         goto done;
 737                 }
 738                 if (sh.sh_flags & SHF_EXECINSTR) {
 739                         minva = min(minva, sh.sh_addr);
 740                         maxva = max(maxva, sh.sh_addr + sh.sh_size);
 741                 }
 742                 if (sh.sh_type == SHT_SYMTAB || sh.sh_type == SHT_DYNSYM)
 743                         pmcstat_image_add_symbols(image, e, scn, &sh);
 744         }
 745
 746         image->pi_start = minva;
 747         image->pi_end   = maxva;
 748         image->pi_type  = image_type;
 749         image->pi_fullpath = pmcstat_string_intern(buffer);
 750
 751         /* Build display name
 752          */
 753         endp = buffer;
 754         for (p = buffer; *p; p++)
 755                 if (*p == '/')
 756                         endp = p+1;
 757         image->pi_name = pmcstat_string_intern(endp);
 758
 759  done:
 760         (void) elf_end(e);
 761         if (fd >= 0)
 762                 (void) close(fd);
 763         return;
 764 }
 765
 766 /*
 767  * Given an image descriptor, determine whether it is an ELF, or AOUT.
 768  * If no handler claims the image, set its type to 'INDETERMINABLE'.
 769  */
 770
 771 void
 772 pmcstat_image_determine_type(struct pmcstat_image *image)
 773 {
 774         assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
 775
 776         /* Try each kind of handler in turn */
 777         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
 778                 pmcstat_image_get_elf_params(image);
 779         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
 780                 pmcstat_image_get_aout_params(image);
 781
 782         /*
 783          * Otherwise, remember that we tried to determine
 784          * the object's type and had failed.
 785          */
 786         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
 787                 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
 788 }
 789
 790 /*
 791  * Locate an image descriptor given an interned path, adding a fresh
 792  * descriptor to the cache if necessary.  This function also finds a
 793  * suitable name for this image's sample file.
 794  *
 795  * We defer filling in the file format specific parts of the image
 796  * structure till the time we actually see a sample that would fall
 797  * into this image.
 798  */
 799
 800 static struct pmcstat_image *
 801 pmcstat_image_from_path(pmcstat_interned_string internedpath,
 802     int iskernelmodule)
 803 {
 804         int hash;
 805         struct pmcstat_image *pi;
 806
 807         hash = pmcstat_string_lookup_hash(internedpath);
 808
 809         /* First, look for an existing entry. */
 810         LIST_FOREACH(pi, &pmcstat_image_hash[hash], pi_next)
 811             if (pi->pi_execpath == internedpath &&
 812                   pi->pi_iskernelmodule == iskernelmodule)
 813                     return (pi);
 814
 815         /*
 816          * Allocate a new entry and place it at the head of the hash
 817          * and LRU lists.
 818          */
 819         pi = malloc(sizeof(*pi));
 820         if (pi == NULL)
 821                 return (NULL);
 822
 823         pi->pi_type = PMCSTAT_IMAGE_UNKNOWN;
 824         pi->pi_execpath = internedpath;
 825         pi->pi_start = ~0;
 826         pi->pi_end = 0;
 827         pi->pi_entry = 0;
 828         pi->pi_vaddr = 0;
 829         pi->pi_isdynamic = 0;
 830         pi->pi_iskernelmodule = iskernelmodule;
 831         pi->pi_dynlinkerpath = NULL;
 832         pi->pi_symbols = NULL;
 833         pi->pi_symcount = 0;
 834         pi->pi_addr2line = NULL;
 835
 836         if (plugins[args.pa_pplugin].pl_initimage != NULL)
 837                 plugins[args.pa_pplugin].pl_initimage(pi);
 838         if (plugins[args.pa_plugin].pl_initimage != NULL)
 839                 plugins[args.pa_plugin].pl_initimage(pi);
 840
 841         LIST_INSERT_HEAD(&pmcstat_image_hash[hash], pi, pi_next);
 842
 843         return (pi);
 844 }
 845
 846 /*
 847  * Record the fact that PC values from 'start' to 'end' come from
 848  * image 'image'.
 849  */
 850
 851 static void
 852 pmcstat_image_link(struct pmcstat_process *pp, struct pmcstat_image *image,
 853     uintfptr_t start)
 854 {
 855         struct pmcstat_pcmap *pcm, *pcmnew;
 856         uintfptr_t offset;
 857
 858         assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN &&
 859             image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE);
 860
 861         if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
 862                 err(EX_OSERR, "ERROR: Cannot create a map entry");
 863
 864         /*
 865          * Adjust the map entry to only cover the text portion
 866          * of the object.
 867          */
 868
 869         offset = start - image->pi_vaddr;
 870         pcmnew->ppm_lowpc  = image->pi_start + offset;
 871         pcmnew->ppm_highpc = image->pi_end + offset;
 872         pcmnew->ppm_image  = image;
 873
 874         assert(pcmnew->ppm_lowpc < pcmnew->ppm_highpc);
 875
 876         /* Overlapped mmap()'s are assumed to never occur. */
 877         TAILQ_FOREACH(pcm, &pp->pp_map, ppm_next)
 878             if (pcm->ppm_lowpc >= pcmnew->ppm_highpc)
 879                     break;
 880
 881         if (pcm == NULL)
 882                 TAILQ_INSERT_TAIL(&pp->pp_map, pcmnew, ppm_next);
 883         else
 884                 TAILQ_INSERT_BEFORE(pcm, pcmnew, ppm_next);
 885 }
 886
 887 /*
 888  * Unmap images in the range [start..end) associated with process
 889  * 'pp'.
 890  */
 891
 892 static void
 893 pmcstat_image_unmap(struct pmcstat_process *pp, uintfptr_t start,
 894     uintfptr_t end)
 895 {
 896         struct pmcstat_pcmap *pcm, *pcmtmp, *pcmnew;
 897
 898         assert(pp != NULL);
 899         assert(start < end);
 900
 901         /*
 902          * Cases:
 903          * - we could have the range completely in the middle of an
 904          *   existing pcmap; in this case we have to split the pcmap
 905          *   structure into two (i.e., generate a 'hole').
 906          * - we could have the range covering multiple pcmaps; these
 907          *   will have to be removed.
 908          * - we could have either 'start' or 'end' falling in the
 909          *   middle of a pcmap; in this case shorten the entry.
 910          */
 911         TAILQ_FOREACH_SAFE(pcm, &pp->pp_map, ppm_next, pcmtmp) {
 912                 assert(pcm->ppm_lowpc < pcm->ppm_highpc);
 913                 if (pcm->ppm_highpc <= start)
 914                         continue;
 915                 if (pcm->ppm_lowpc >= end)
 916                         return;
 917                 if (pcm->ppm_lowpc >= start && pcm->ppm_highpc <= end) {
 918                         /*
 919                          * The current pcmap is completely inside the
 920                          * unmapped range: remove it entirely.
 921                          */
 922                         TAILQ_REMOVE(&pp->pp_map, pcm, ppm_next);
 923                         free(pcm);
 924                 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc > end) {
 925                         /*
 926                          * Split this pcmap into two; curtail the
 927                          * current map to end at [start-1], and start
 928                          * the new one at [end].
 929                          */
 930                         if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
 931                                 err(EX_OSERR,
 932                                     "ERROR: Cannot split a map entry");
 933
 934                         pcmnew->ppm_image = pcm->ppm_image;
 935
 936                         pcmnew->ppm_lowpc = end;
 937                         pcmnew->ppm_highpc = pcm->ppm_highpc;
 938
 939                         pcm->ppm_highpc = start;
 940
 941                         TAILQ_INSERT_AFTER(&pp->pp_map, pcm, pcmnew, ppm_next);
 942
 943                         return;
 944                 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc <= end)
 945                         pcm->ppm_highpc = start;
 946                 else if (pcm->ppm_lowpc >= start && pcm->ppm_highpc > end)
 947                         pcm->ppm_lowpc = end;
 948                 else
 949                         assert(0);
 950         }
 951 }
 952
 953 /*
 954  * Resolve file name and line number for the given address.
 955  */
 956 int
 957 pmcstat_image_addr2line(struct pmcstat_image *image, uintfptr_t addr,
 958     char *sourcefile, size_t sourcefile_len, unsigned *sourceline,
 959     char *funcname, size_t funcname_len)
 960 {
 961         static int addr2line_warn = 0;
 962         unsigned l;
 963
 964         char *sep, cmdline[PATH_MAX], imagepath[PATH_MAX];
 965         int fd;
 966
 967         if (image->pi_addr2line == NULL) {
 968                 snprintf(imagepath, sizeof(imagepath), "%s%s.symbols",
 969                     args.pa_fsroot,
 970                     pmcstat_string_unintern(image->pi_fullpath));
 971                 fd = open(imagepath, O_RDONLY);
 972                 if (fd < 0) {
 973                         snprintf(imagepath, sizeof(imagepath), "%s%s",
 974                             args.pa_fsroot,
 975                             pmcstat_string_unintern(image->pi_fullpath));
 976                 } else
 977                         close(fd);
 978                 /*
 979                  * New addr2line support recursive inline function with -i
 980                  * but the format does not add a marker when no more entries
 981                  * are available.
 982                  */
 983                 snprintf(cmdline, sizeof(cmdline), "addr2line -Cfe \"%s\"",
 984                     imagepath);
 985                 image->pi_addr2line = popen(cmdline, "r+");
 986                 if (image->pi_addr2line == NULL) {
 987                         if (!addr2line_warn) {
 988                                 addr2line_warn = 1;
 989                                 warnx(
 990 "WARNING: addr2line is needed for source code information."
 991                                     );
 992                         }
 993                         return (0);
 994                 }
 995         }
 996
 997         if (feof(image->pi_addr2line) || ferror(image->pi_addr2line)) {
 998                 warnx("WARNING: addr2line pipe error");
 999                 pclose(image->pi_addr2line);
1000                 image->pi_addr2line = NULL;
1001                 return (0);
1002         }
1003
1004         fprintf(image->pi_addr2line, "%p\n", (void *)addr);
1005
1006         if (fgets(funcname, funcname_len, image->pi_addr2line) == NULL) {
1007                 warnx("WARNING: addr2line function name read error");
1008                 return (0);
1009         }
1010         sep = strchr(funcname, '\n');
1011         if (sep != NULL)
1012                 *sep = '\0';
1013
1014         if (fgets(sourcefile, sourcefile_len, image->pi_addr2line) == NULL) {
1015                 warnx("WARNING: addr2line source file read error");
1016                 return (0);
1017         }
1018         sep = strchr(sourcefile, ':');
1019         if (sep == NULL) {
1020                 warnx("WARNING: addr2line source line separator missing");
1021                 return (0);
1022         }
1023         *sep = '\0';
1024         l = atoi(sep+1);
1025         if (l == 0)
1026                 return (0);
1027         *sourceline = l;
1028         return (1);
1029 }
1030
1031 /*
1032  * Add a {pmcid,name} mapping.
1033  */
1034
1035 static void
1036 pmcstat_pmcid_add(pmc_id_t pmcid, pmcstat_interned_string ps)
1037 {
1038         struct pmcstat_pmcrecord *pr, *prm;
1039
1040         /* Replace an existing name for the PMC. */
1041         prm = NULL;
1042         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1043                 if (pr->pr_pmcid == pmcid) {
1044                         pr->pr_pmcname = ps;
1045                         return;
1046                 } else if (pr->pr_pmcname == ps)
1047                         prm = pr;
1048
1049         /*
1050          * Otherwise, allocate a new descriptor and call the
1051          * plugins hook.
1052          */
1053         if ((pr = malloc(sizeof(*pr))) == NULL)
1054                 err(EX_OSERR, "ERROR: Cannot allocate pmc record");
1055
1056         pr->pr_pmcid = pmcid;
1057         pr->pr_pmcname = ps;
1058         pr->pr_pmcin = pmcstat_npmcs++;
1059         pr->pr_samples = 0;
1060         pr->pr_dubious_frames = 0;
1061         pr->pr_merge = prm == NULL ? pr : prm;
1062
1063         LIST_INSERT_HEAD(&pmcstat_pmcs, pr, pr_next);
1064
1065         if (plugins[args.pa_pplugin].pl_newpmc != NULL)
1066                 plugins[args.pa_pplugin].pl_newpmc(ps, pr);
1067         if (plugins[args.pa_plugin].pl_newpmc != NULL)
1068                 plugins[args.pa_plugin].pl_newpmc(ps, pr);
1069 }
1070
1071 /*
1072  * Given a pmcid in use, find its human-readable name.
1073  */
1074
1075 const char *
1076 pmcstat_pmcid_to_name(pmc_id_t pmcid)
1077 {
1078         struct pmcstat_pmcrecord *pr;
1079
1080         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1081             if (pr->pr_pmcid == pmcid)
1082                     return (pmcstat_string_unintern(pr->pr_pmcname));
1083
1084         return NULL;
1085 }
1086
1087 /*
1088  * Convert PMC index to name.
1089  */
1090
1091 const char *
1092 pmcstat_pmcindex_to_name(int pmcin)
1093 {
1094         struct pmcstat_pmcrecord *pr;
1095
1096         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1097                 if (pr->pr_pmcin == pmcin)
1098                         return pmcstat_string_unintern(pr->pr_pmcname);
1099
1100         return NULL;
1101 }
1102
1103 /*
1104  * Return PMC record with given index.
1105  */
1106
1107 struct pmcstat_pmcrecord *
1108 pmcstat_pmcindex_to_pmcr(int pmcin)
1109 {
1110         struct pmcstat_pmcrecord *pr;
1111
1112         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1113                 if (pr->pr_pmcin == pmcin)
1114                         return pr;
1115
1116         return NULL;
1117 }
1118
1119 /*
1120  * Get PMC record by id, apply merge policy.
1121  */
1122
1123 static struct pmcstat_pmcrecord *
1124 pmcstat_lookup_pmcid(pmc_id_t pmcid)
1125 {
1126         struct pmcstat_pmcrecord *pr;
1127
1128         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
1129                 if (pr->pr_pmcid == pmcid) {
1130                         if (pmcstat_mergepmc)
1131                                 return pr->pr_merge;
1132                         return pr;
1133                 }
1134         }
1135
1136         return NULL;
1137 }
1138
1139 /*
1140  * Associate an AOUT image with a process.
1141  */
1142
1143 static void
1144 pmcstat_process_aout_exec(struct pmcstat_process *pp,
1145     struct pmcstat_image *image, uintfptr_t entryaddr)
1146 {
1147         (void) pp;
1148         (void) image;
1149         (void) entryaddr;
1150         /* TODO Implement a.out handling */
1151 }
1152
1153 /*
1154  * Associate an ELF image with a process.
1155  */
1156
1157 static void
1158 pmcstat_process_elf_exec(struct pmcstat_process *pp,
1159     struct pmcstat_image *image, uintfptr_t entryaddr)
1160 {
1161         uintmax_t libstart;
1162         struct pmcstat_image *rtldimage;
1163
1164         assert(image->pi_type == PMCSTAT_IMAGE_ELF32 ||
1165             image->pi_type == PMCSTAT_IMAGE_ELF64);
1166
1167         /* Create a map entry for the base executable. */
1168         pmcstat_image_link(pp, image, image->pi_vaddr);
1169
1170         /*
1171          * For dynamically linked executables we need to determine
1172          * where the dynamic linker was mapped to for this process,
1173          * Subsequent executable objects that are mapped in by the
1174          * dynamic linker will be tracked by log events of type
1175          * PMCLOG_TYPE_MAP_IN.
1176          */
1177
1178         if (image->pi_isdynamic) {
1179
1180                 /*
1181                  * The runtime loader gets loaded just after the maximum
1182                  * possible heap address.  Like so:
1183                  *
1184                  * [  TEXT DATA BSS HEAP -->*RTLD  SHLIBS   <--STACK]
1185                  * ^                                                ^
1186                  * 0                               VM_MAXUSER_ADDRESS
1187
1188                  *
1189                  * The exact address where the loader gets mapped in
1190                  * will vary according to the size of the executable
1191                  * and the limits on the size of the process'es data
1192                  * segment at the time of exec().  The entry address
1193                  * recorded at process exec time corresponds to the
1194                  * 'start' address inside the dynamic linker.  From
1195                  * this we can figure out the address where the
1196                  * runtime loader's file object had been mapped to.
1197                  */
1198                 rtldimage = pmcstat_image_from_path(image->pi_dynlinkerpath, 0);
1199                 if (rtldimage == NULL) {
1200                         warnx("WARNING: Cannot find image for \"%s\".",
1201                             pmcstat_string_unintern(image->pi_dynlinkerpath));
1202                         pmcstat_stats.ps_exec_errors++;
1203                         return;
1204                 }
1205
1206                 if (rtldimage->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1207                         pmcstat_image_get_elf_params(rtldimage);
1208
1209                 if (rtldimage->pi_type != PMCSTAT_IMAGE_ELF32 &&
1210                     rtldimage->pi_type != PMCSTAT_IMAGE_ELF64) {
1211                         warnx("WARNING: rtld not an ELF object \"%s\".",
1212                             pmcstat_string_unintern(image->pi_dynlinkerpath));
1213                         return;
1214                 }
1215
1216                 libstart = entryaddr - rtldimage->pi_entry;
1217                 pmcstat_image_link(pp, rtldimage, libstart);
1218         }
1219 }
1220
1221 /*
1222  * Find the process descriptor corresponding to a PID.  If 'allocate'
1223  * is zero, we return a NULL if a pid descriptor could not be found or
1224  * a process descriptor process.  If 'allocate' is non-zero, then we
1225  * will attempt to allocate a fresh process descriptor.  Zombie
1226  * process descriptors are only removed if a fresh allocation for the
1227  * same PID is requested.
1228  */
1229
1230 static struct pmcstat_process *
1231 pmcstat_process_lookup(pid_t pid, int allocate)
1232 {
1233         uint32_t hash;
1234         struct pmcstat_pcmap *ppm, *ppmtmp;
1235         struct pmcstat_process *pp, *pptmp;
1236
1237         hash = (uint32_t) pid & PMCSTAT_HASH_MASK;      /* simplicity wins */
1238
1239         LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[hash], pp_next, pptmp)
1240                 if (pp->pp_pid == pid) {
1241                         /* Found a descriptor, check and process zombies */
1242                         if (allocate && pp->pp_isactive == 0) {
1243                                 /* remove maps */
1244                                 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next,
1245                                     ppmtmp) {
1246                                         TAILQ_REMOVE(&pp->pp_map, ppm,
1247                                             ppm_next);
1248                                         free(ppm);
1249                                 }
1250                                 /* remove process entry */
1251                                 LIST_REMOVE(pp, pp_next);
1252                                 free(pp);
1253                                 break;
1254                         }
1255                         return (pp);
1256                 }
1257
1258         if (!allocate)
1259                 return (NULL);
1260
1261         if ((pp = malloc(sizeof(*pp))) == NULL)
1262                 err(EX_OSERR, "ERROR: Cannot allocate pid descriptor");
1263
1264         pp->pp_pid = pid;
1265         pp->pp_isactive = 1;
1266
1267         TAILQ_INIT(&pp->pp_map);
1268
1269         LIST_INSERT_HEAD(&pmcstat_process_hash[hash], pp, pp_next);
1270         return (pp);
1271 }
1272
1273 /*
1274  * Associate an image and a process.
1275  */
1276
1277 static void
1278 pmcstat_process_exec(struct pmcstat_process *pp,
1279     pmcstat_interned_string path, uintfptr_t entryaddr)
1280 {
1281         struct pmcstat_image *image;
1282
1283         if ((image = pmcstat_image_from_path(path, 0)) == NULL) {
1284                 pmcstat_stats.ps_exec_errors++;
1285                 return;
1286         }
1287
1288         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1289                 pmcstat_image_determine_type(image);
1290
1291         assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN);
1292
1293         switch (image->pi_type) {
1294         case PMCSTAT_IMAGE_ELF32:
1295         case PMCSTAT_IMAGE_ELF64:
1296                 pmcstat_stats.ps_exec_elf++;
1297                 pmcstat_process_elf_exec(pp, image, entryaddr);
1298                 break;
1299
1300         case PMCSTAT_IMAGE_AOUT:
1301                 pmcstat_stats.ps_exec_aout++;
1302                 pmcstat_process_aout_exec(pp, image, entryaddr);
1303                 break;
1304
1305         case PMCSTAT_IMAGE_INDETERMINABLE:
1306                 pmcstat_stats.ps_exec_indeterminable++;
1307                 break;
1308
1309         default:
1310                 err(EX_SOFTWARE,
1311                     "ERROR: Unsupported executable type for \"%s\"",
1312                     pmcstat_string_unintern(path));
1313         }
1314 }
1315
1316
1317 /*
1318  * Find the map entry associated with process 'p' at PC value 'pc'.
1319  */
1320
1321 struct pmcstat_pcmap *
1322 pmcstat_process_find_map(struct pmcstat_process *p, uintfptr_t pc)
1323 {
1324         struct pmcstat_pcmap *ppm;
1325
1326         TAILQ_FOREACH(ppm, &p->pp_map, ppm_next) {
1327                 if (pc >= ppm->ppm_lowpc && pc < ppm->ppm_highpc)
1328                         return (ppm);
1329                 if (pc < ppm->ppm_lowpc)
1330                         return (NULL);
1331         }
1332
1333         return (NULL);
1334 }
1335
1336 /*
1337  * Convert a hwpmc(4) log to profile information.  A system-wide
1338  * callgraph is generated if FLAG_DO_CALLGRAPHS is set.  gmon.out
1339  * files usable by gprof(1) are created if FLAG_DO_GPROF is set.
1340  */
1341 static int
1342 pmcstat_analyze_log(void)
1343 {
1344         uint32_t cpu, cpuflags;
1345         uintfptr_t pc;
1346         pid_t pid;
1347         struct pmcstat_image *image;
1348         struct pmcstat_process *pp, *ppnew;
1349         struct pmcstat_pcmap *ppm, *ppmtmp;
1350         struct pmclog_ev ev;
1351         struct pmcstat_pmcrecord *pmcr;
1352         pmcstat_interned_string image_path;
1353
1354         assert(args.pa_flags & FLAG_DO_ANALYSIS);
1355
1356         if (elf_version(EV_CURRENT) == EV_NONE)
1357                 err(EX_UNAVAILABLE, "Elf library intialization failed");
1358
1359         while (pmclog_read(args.pa_logparser, &ev) == 0) {
1360                 assert(ev.pl_state == PMCLOG_OK);
1361
1362                 switch (ev.pl_type) {
1363                 case PMCLOG_TYPE_INITIALIZE:
1364                         if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1365                             PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1366                                 warnx(
1367 "WARNING: Log version 0x%x does not match compiled version 0x%x.",
1368                                     ev.pl_u.pl_i.pl_version, PMC_VERSION_MAJOR);
1369                         break;
1370
1371                 case PMCLOG_TYPE_MAP_IN:
1372                         /*
1373                          * Introduce an address range mapping for a
1374                          * userland process or the kernel (pid == -1).
1375                          *
1376                          * We always allocate a process descriptor so
1377                          * that subsequent samples seen for this
1378                          * address range are mapped to the current
1379                          * object being mapped in.
1380                          */
1381                         pid = ev.pl_u.pl_mi.pl_pid;
1382                         if (pid == -1)
1383                                 pp = pmcstat_kernproc;
1384                         else
1385                                 pp = pmcstat_process_lookup(pid,
1386                                     PMCSTAT_ALLOCATE);
1387
1388                         assert(pp != NULL);
1389
1390                         image_path = pmcstat_string_intern(ev.pl_u.pl_mi.
1391                             pl_pathname);
1392                         image = pmcstat_image_from_path(image_path, pid == -1);
1393                         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1394                                 pmcstat_image_determine_type(image);
1395                         if (image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE)
1396                                 pmcstat_image_link(pp, image,
1397                                     ev.pl_u.pl_mi.pl_start);
1398                         break;
1399
1400                 case PMCLOG_TYPE_MAP_OUT:
1401                         /*
1402                          * Remove an address map.
1403                          */
1404                         pid = ev.pl_u.pl_mo.pl_pid;
1405                         if (pid == -1)
1406                                 pp = pmcstat_kernproc;
1407                         else
1408                                 pp = pmcstat_process_lookup(pid, 0);
1409
1410                         if (pp == NULL) /* unknown process */
1411                                 break;
1412
1413                         pmcstat_image_unmap(pp, ev.pl_u.pl_mo.pl_start,
1414                             ev.pl_u.pl_mo.pl_end);
1415                         break;
1416
1417                 case PMCLOG_TYPE_PCSAMPLE:
1418                         /*
1419                          * Note: the `PCSAMPLE' log entry is not
1420                          * generated by hpwmc(4) after version 2.
1421                          */
1422
1423                         /*
1424                          * We bring in the gmon file for the image
1425                          * currently associated with the PMC & pid
1426                          * pair and increment the appropriate entry
1427                          * bin inside this.
1428                          */
1429                         pmcstat_stats.ps_samples_total++;
1430                         ps_samples_period++;
1431
1432                         pc = ev.pl_u.pl_s.pl_pc;
1433                         pp = pmcstat_process_lookup(ev.pl_u.pl_s.pl_pid,
1434                             PMCSTAT_ALLOCATE);
1435
1436                         /* Get PMC record. */
1437                         pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_s.pl_pmcid);
1438                         assert(pmcr != NULL);
1439                         pmcr->pr_samples++;
1440
1441                         /*
1442                          * Call the plugins processing
1443                          * TODO: move pmcstat_process_find_map inside plugins
1444                          */
1445
1446                         if (plugins[args.pa_pplugin].pl_process != NULL)
1447                                 plugins[args.pa_pplugin].pl_process(
1448                                     pp, pmcr, 1, &pc,
1449                                     pmcstat_process_find_map(pp, pc) != NULL, 0);
1450                         plugins[args.pa_plugin].pl_process(
1451                             pp, pmcr, 1, &pc,
1452                             pmcstat_process_find_map(pp, pc) != NULL, 0);
1453                         break;
1454
1455                 case PMCLOG_TYPE_CALLCHAIN:
1456                         pmcstat_stats.ps_samples_total++;
1457                         ps_samples_period++;
1458
1459                         cpuflags = ev.pl_u.pl_cc.pl_cpuflags;
1460                         cpu = PMC_CALLCHAIN_CPUFLAGS_TO_CPU(cpuflags);
1461
1462                         /* Filter on the CPU id. */
1463                         if (!CPU_ISSET(cpu, &(args.pa_cpumask))) {
1464                                 pmcstat_stats.ps_samples_skipped++;
1465                                 break;
1466                         }
1467
1468                         pp = pmcstat_process_lookup(ev.pl_u.pl_cc.pl_pid,
1469                             PMCSTAT_ALLOCATE);
1470
1471                         /* Get PMC record. */
1472                         pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_cc.pl_pmcid);
1473                         assert(pmcr != NULL);
1474                         pmcr->pr_samples++;
1475
1476                         /*
1477                          * Call the plugins processing
1478                          */
1479
1480                         if (plugins[args.pa_pplugin].pl_process != NULL)
1481                                 plugins[args.pa_pplugin].pl_process(
1482                                     pp, pmcr,
1483                                     ev.pl_u.pl_cc.pl_npc,
1484                                     ev.pl_u.pl_cc.pl_pc,
1485                                     PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1486                                     cpu);
1487                         plugins[args.pa_plugin].pl_process(
1488                             pp, pmcr,
1489                             ev.pl_u.pl_cc.pl_npc,
1490                             ev.pl_u.pl_cc.pl_pc,
1491                             PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1492                             cpu);
1493                         break;
1494
1495                 case PMCLOG_TYPE_PMCALLOCATE:
1496                         /*
1497                          * Record the association pmc id between this
1498                          * PMC and its name.
1499                          */
1500                         pmcstat_pmcid_add(ev.pl_u.pl_a.pl_pmcid,
1501                             pmcstat_string_intern(ev.pl_u.pl_a.pl_evname));
1502                         break;
1503
1504                 case PMCLOG_TYPE_PMCALLOCATEDYN:
1505                         /*
1506                          * Record the association pmc id between this
1507                          * PMC and its name.
1508                          */
1509                         pmcstat_pmcid_add(ev.pl_u.pl_ad.pl_pmcid,
1510                             pmcstat_string_intern(ev.pl_u.pl_ad.pl_evname));
1511                         break;
1512
1513                 case PMCLOG_TYPE_PROCEXEC:
1514
1515                         /*
1516                          * Change the executable image associated with
1517                          * a process.
1518                          */
1519                         pp = pmcstat_process_lookup(ev.pl_u.pl_x.pl_pid,
1520                             PMCSTAT_ALLOCATE);
1521
1522                         /* delete the current process map */
1523                         TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
1524                                 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
1525                                 free(ppm);
1526                         }
1527
1528                         /* associate this process  image */
1529                         image_path = pmcstat_string_intern(
1530                                 ev.pl_u.pl_x.pl_pathname);
1531                         assert(image_path != NULL);
1532                         pmcstat_process_exec(pp, image_path,
1533                             ev.pl_u.pl_x.pl_entryaddr);
1534                         break;
1535
1536                 case PMCLOG_TYPE_PROCEXIT:
1537
1538                         /*
1539                          * Due to the way the log is generated, the
1540                          * last few samples corresponding to a process
1541                          * may appear in the log after the process
1542                          * exit event is recorded.  Thus we keep the
1543                          * process' descriptor and associated data
1544                          * structures around, but mark the process as
1545                          * having exited.
1546                          */
1547                         pp = pmcstat_process_lookup(ev.pl_u.pl_e.pl_pid, 0);
1548                         if (pp == NULL)
1549                                 break;
1550                         pp->pp_isactive = 0;    /* mark as a zombie */
1551                         break;
1552
1553                 case PMCLOG_TYPE_SYSEXIT:
1554                         pp = pmcstat_process_lookup(ev.pl_u.pl_se.pl_pid, 0);
1555                         if (pp == NULL)
1556                                 break;
1557                         pp->pp_isactive = 0;    /* make a zombie */
1558                         break;
1559
1560                 case PMCLOG_TYPE_PROCFORK:
1561
1562                         /*
1563                          * Allocate a process descriptor for the new
1564                          * (child) process.
1565                          */
1566                         ppnew =
1567                             pmcstat_process_lookup(ev.pl_u.pl_f.pl_newpid,
1568                                 PMCSTAT_ALLOCATE);
1569
1570                         /*
1571                          * If we had been tracking the parent, clone
1572                          * its address maps.
1573                          */
1574                         pp = pmcstat_process_lookup(ev.pl_u.pl_f.pl_oldpid, 0);
1575                         if (pp == NULL)
1576                                 break;
1577                         TAILQ_FOREACH(ppm, &pp->pp_map, ppm_next)
1578                             pmcstat_image_link(ppnew, ppm->ppm_image,
1579                                 ppm->ppm_lowpc);
1580                         break;
1581
1582                 default:        /* other types of entries are not relevant */
1583                         break;
1584                 }
1585         }
1586
1587         if (ev.pl_state == PMCLOG_EOF)
1588                 return (PMCSTAT_FINISHED);
1589         else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1590                 return (PMCSTAT_RUNNING);
1591
1592         err(EX_DATAERR,
1593             "ERROR: event parsing failed (record %jd, offset 0x%jx)",
1594             (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1595 }
1596
1597 /*
1598  * Print log entries as text.
1599  */
1600
1601 static int
1602 pmcstat_print_log(void)
1603 {
1604         struct pmclog_ev ev;
1605         uint32_t npc;
1606
1607         while (pmclog_read(args.pa_logparser, &ev) == 0) {
1608                 assert(ev.pl_state == PMCLOG_OK);
1609                 switch (ev.pl_type) {
1610                 case PMCLOG_TYPE_CALLCHAIN:
1611                         PMCSTAT_PRINT_ENTRY("callchain",
1612                             "%d 0x%x %d %d %c", ev.pl_u.pl_cc.pl_pid,
1613                             ev.pl_u.pl_cc.pl_pmcid,
1614                             PMC_CALLCHAIN_CPUFLAGS_TO_CPU(ev.pl_u.pl_cc. \
1615                                 pl_cpuflags), ev.pl_u.pl_cc.pl_npc,
1616                             PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(ev.pl_u.pl_cc.\
1617                                 pl_cpuflags) ? 'u' : 's');
1618                         for (npc = 0; npc < ev.pl_u.pl_cc.pl_npc; npc++)
1619                                 PMCSTAT_PRINT_ENTRY("...", "%p",
1620                                     (void *) ev.pl_u.pl_cc.pl_pc[npc]);
1621                         break;
1622                 case PMCLOG_TYPE_CLOSELOG:
1623                         PMCSTAT_PRINT_ENTRY("closelog",);
1624                         break;
1625                 case PMCLOG_TYPE_DROPNOTIFY:
1626                         PMCSTAT_PRINT_ENTRY("drop",);
1627                         break;
1628                 case PMCLOG_TYPE_INITIALIZE:
1629                         PMCSTAT_PRINT_ENTRY("initlog","0x%x \"%s\"",
1630                             ev.pl_u.pl_i.pl_version,
1631                             pmc_name_of_cputype(ev.pl_u.pl_i.pl_arch));
1632                         if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1633                             PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1634                                 warnx(
1635 "WARNING: Log version 0x%x != expected version 0x%x.",
1636                                     ev.pl_u.pl_i.pl_version, PMC_VERSION);
1637                         break;
1638                 case PMCLOG_TYPE_MAP_IN:
1639                         PMCSTAT_PRINT_ENTRY("map-in","%d %p \"%s\"",
1640                             ev.pl_u.pl_mi.pl_pid,
1641                             (void *) ev.pl_u.pl_mi.pl_start,
1642                             ev.pl_u.pl_mi.pl_pathname);
1643                         break;
1644                 case PMCLOG_TYPE_MAP_OUT:
1645                         PMCSTAT_PRINT_ENTRY("map-out","%d %p %p",
1646                             ev.pl_u.pl_mo.pl_pid,
1647                             (void *) ev.pl_u.pl_mo.pl_start,
1648                             (void *) ev.pl_u.pl_mo.pl_end);
1649                         break;
1650                 case PMCLOG_TYPE_PCSAMPLE:
1651                         PMCSTAT_PRINT_ENTRY("sample","0x%x %d %p %c",
1652                             ev.pl_u.pl_s.pl_pmcid,
1653                             ev.pl_u.pl_s.pl_pid,
1654                             (void *) ev.pl_u.pl_s.pl_pc,
1655                             ev.pl_u.pl_s.pl_usermode ? 'u' : 's');
1656                         break;
1657                 case PMCLOG_TYPE_PMCALLOCATE:
1658                         PMCSTAT_PRINT_ENTRY("allocate","0x%x \"%s\" 0x%x",
1659                             ev.pl_u.pl_a.pl_pmcid,
1660                             ev.pl_u.pl_a.pl_evname,
1661                             ev.pl_u.pl_a.pl_flags);
1662                         break;
1663                 case PMCLOG_TYPE_PMCALLOCATEDYN:
1664                         PMCSTAT_PRINT_ENTRY("allocatedyn","0x%x \"%s\" 0x%x",
1665                             ev.pl_u.pl_ad.pl_pmcid,
1666                             ev.pl_u.pl_ad.pl_evname,
1667                             ev.pl_u.pl_ad.pl_flags);
1668                         break;
1669                 case PMCLOG_TYPE_PMCATTACH:
1670                         PMCSTAT_PRINT_ENTRY("attach","0x%x %d \"%s\"",
1671                             ev.pl_u.pl_t.pl_pmcid,
1672                             ev.pl_u.pl_t.pl_pid,
1673                             ev.pl_u.pl_t.pl_pathname);
1674                         break;
1675                 case PMCLOG_TYPE_PMCDETACH:
1676                         PMCSTAT_PRINT_ENTRY("detach","0x%x %d",
1677                             ev.pl_u.pl_d.pl_pmcid,
1678                             ev.pl_u.pl_d.pl_pid);
1679                         break;
1680                 case PMCLOG_TYPE_PROCCSW:
1681                         PMCSTAT_PRINT_ENTRY("cswval","0x%x %d %jd",
1682                             ev.pl_u.pl_c.pl_pmcid,
1683                             ev.pl_u.pl_c.pl_pid,
1684                             ev.pl_u.pl_c.pl_value);
1685                         break;
1686                 case PMCLOG_TYPE_PROCEXEC:
1687                         PMCSTAT_PRINT_ENTRY("exec","0x%x %d %p \"%s\"",
1688                             ev.pl_u.pl_x.pl_pmcid,
1689                             ev.pl_u.pl_x.pl_pid,
1690                             (void *) ev.pl_u.pl_x.pl_entryaddr,
1691                             ev.pl_u.pl_x.pl_pathname);
1692                         break;
1693                 case PMCLOG_TYPE_PROCEXIT:
1694                         PMCSTAT_PRINT_ENTRY("exitval","0x%x %d %jd",
1695                             ev.pl_u.pl_e.pl_pmcid,
1696                             ev.pl_u.pl_e.pl_pid,
1697                             ev.pl_u.pl_e.pl_value);
1698                         break;
1699                 case PMCLOG_TYPE_PROCFORK:
1700                         PMCSTAT_PRINT_ENTRY("fork","%d %d",
1701                             ev.pl_u.pl_f.pl_oldpid,
1702                             ev.pl_u.pl_f.pl_newpid);
1703                         break;
1704                 case PMCLOG_TYPE_USERDATA:
1705                         PMCSTAT_PRINT_ENTRY("userdata","0x%x",
1706                             ev.pl_u.pl_u.pl_userdata);
1707                         break;
1708                 case PMCLOG_TYPE_SYSEXIT:
1709                         PMCSTAT_PRINT_ENTRY("exit","%d",
1710                             ev.pl_u.pl_se.pl_pid);
1711                         break;
1712                 default:
1713                         fprintf(args.pa_printfile, "unknown event (type %d).\n",
1714                             ev.pl_type);
1715                 }
1716         }
1717
1718         if (ev.pl_state == PMCLOG_EOF)
1719                 return (PMCSTAT_FINISHED);
1720         else if (ev.pl_state ==  PMCLOG_REQUIRE_DATA)
1721                 return (PMCSTAT_RUNNING);
1722
1723         errx(EX_DATAERR,
1724             "ERROR: event parsing failed (record %jd, offset 0x%jx).",
1725             (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1726         /*NOTREACHED*/
1727 }
1728
1729 /*
1730  * Public Interfaces.
1731  */
1732
1733 /*
1734  * Close a logfile, after first flushing all in-module queued data.
1735  */
1736
1737 int
1738 pmcstat_close_log(void)
1739 {
1740         /* If a local logfile is configured ask the kernel to stop
1741          * and flush data. Kernel will close the file when data is flushed
1742          * so keep the status to EXITING.
1743          */
1744         if (args.pa_logfd != -1) {
1745                 if (pmc_close_logfile() < 0)
1746                         err(EX_OSERR, "ERROR: logging failed");
1747         }
1748
1749         return (args.pa_flags & FLAG_HAS_PIPE ? PMCSTAT_EXITING :
1750             PMCSTAT_FINISHED);
1751 }
1752
1753
1754
1755 /*
1756  * Open a log file, for reading or writing.
1757  *
1758  * The function returns the fd of a successfully opened log or -1 in
1759  * case of failure.
1760  */
1761
1762 int
1763 pmcstat_open_log(const char *path, int mode)
1764 {
1765         int error, fd, cfd;
1766         size_t hlen;
1767         const char *p, *errstr;
1768         struct addrinfo hints, *res, *res0;
1769         char hostname[MAXHOSTNAMELEN];
1770
1771         errstr = NULL;
1772         fd = -1;
1773
1774         /*
1775          * If 'path' is "-" then open one of stdin or stdout depending
1776          * on the value of 'mode'.
1777          *
1778          * If 'path' contains a ':' and does not start with a '/' or '.',
1779          * and is being opened for writing, treat it as a "host:port"
1780          * specification and open a network socket.
1781          *
1782          * Otherwise, treat 'path' as a file name and open that.
1783          */
1784         if (path[0] == '-' && path[1] == '\0')
1785                 fd = (mode == PMCSTAT_OPEN_FOR_READ) ? 0 : 1;
1786         else if (path[0] != '/' &&
1787             path[0] != '.' && strchr(path, ':') != NULL) {
1788
1789                 p = strrchr(path, ':');
1790                 hlen = p - path;
1791                 if (p == path || hlen >= sizeof(hostname)) {
1792                         errstr = strerror(EINVAL);
1793                         goto done;
1794                 }
1795
1796                 assert(hlen < sizeof(hostname));
1797                 (void) strncpy(hostname, path, hlen);
1798                 hostname[hlen] = '\0';
1799
1800                 (void) memset(&hints, 0, sizeof(hints));
1801                 hints.ai_family = AF_UNSPEC;
1802                 hints.ai_socktype = SOCK_STREAM;
1803                 if ((error = getaddrinfo(hostname, p+1, &hints, &res0)) != 0) {
1804                         errstr = gai_strerror(error);
1805                         goto done;
1806                 }
1807
1808                 fd = -1;
1809                 for (res = res0; res; res = res->ai_next) {
1810                         if ((fd = socket(res->ai_family, res->ai_socktype,
1811                             res->ai_protocol)) < 0) {
1812                                 errstr = strerror(errno);
1813                                 continue;
1814                         }
1815                         if (mode == PMCSTAT_OPEN_FOR_READ) {
1816                                 if (bind(fd, res->ai_addr, res->ai_addrlen) < 0) {
1817                                         errstr = strerror(errno);
1818                                         (void) close(fd);
1819                                         fd = -1;
1820                                         continue;
1821                                 }
1822                                 listen(fd, 1);
1823                                 cfd = accept(fd, NULL, NULL);
1824                                 (void) close(fd);
1825                                 if (cfd < 0) {
1826                                         errstr = strerror(errno);
1827                                         fd = -1;
1828                                         break;
1829                                 }
1830                                 fd = cfd;
1831                         } else {
1832                                 if (connect(fd, res->ai_addr, res->ai_addrlen) < 0) {
1833                                         errstr = strerror(errno);
1834                                         (void) close(fd);
1835                                         fd = -1;
1836                                         continue;
1837                                 }
1838                         }
1839                         errstr = NULL;
1840                         break;
1841                 }
1842                 freeaddrinfo(res0);
1843
1844         } else if ((fd = open(path, mode == PMCSTAT_OPEN_FOR_READ ?
1845                     O_RDONLY : (O_WRONLY|O_CREAT|O_TRUNC),
1846                     S_IRUSR|S_IWUSR|S_IRGRP|S_IROTH)) < 0)
1847                         errstr = strerror(errno);
1848
1849   done:
1850         if (errstr)
1851                 errx(EX_OSERR, "ERROR: Cannot open \"%s\" for %s: %s.", path,
1852                     (mode == PMCSTAT_OPEN_FOR_READ ? "reading" : "writing"),
1853                     errstr);
1854
1855         return (fd);
1856 }
1857
1858 /*
1859  * Process a log file in offline analysis mode.
1860  */
1861
1862 int
1863 pmcstat_process_log(void)
1864 {
1865
1866         /*
1867          * If analysis has not been asked for, just print the log to
1868          * the current output file.
1869          */
1870         if (args.pa_flags & FLAG_DO_PRINT)
1871                 return (pmcstat_print_log());
1872         else
1873                 return (pmcstat_analyze_log());
1874 }
1875
1876 /*
1877  * Refresh top display.
1878  */
1879
1880 static void
1881 pmcstat_refresh_top(void)
1882 {
1883         int v_attrs;
1884         float v;
1885         char pmcname[40];
1886         struct pmcstat_pmcrecord *pmcpr;
1887
1888         /* If in pause mode do not refresh display. */
1889         if (pmcstat_pause)
1890                 return;
1891
1892         /* Wait until PMC pop in the log. */
1893         pmcpr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1894         if (pmcpr == NULL)
1895                 return;
1896
1897         /* Format PMC name. */
1898         if (pmcstat_mergepmc)
1899                 snprintf(pmcname, sizeof(pmcname), "[%s]",
1900                     pmcstat_string_unintern(pmcpr->pr_pmcname));
1901         else
1902                 snprintf(pmcname, sizeof(pmcname), "%s.%d",
1903                     pmcstat_string_unintern(pmcpr->pr_pmcname),
1904                     pmcstat_pmcinfilter);
1905
1906         /* Format samples count. */
1907         if (ps_samples_period > 0)
1908                 v = (pmcpr->pr_samples * 100.0) / ps_samples_period;
1909         else
1910                 v = 0.;
1911         v_attrs = PMCSTAT_ATTRPERCENT(v);
1912
1913         PMCSTAT_PRINTBEGIN();
1914         PMCSTAT_PRINTW("PMC: %s Samples: %u ",
1915             pmcname,
1916             pmcpr->pr_samples);
1917         PMCSTAT_ATTRON(v_attrs);
1918         PMCSTAT_PRINTW("(%.1f%%) ", v);
1919         PMCSTAT_ATTROFF(v_attrs);
1920         PMCSTAT_PRINTW(", %u unresolved\n\n",
1921             pmcpr->pr_dubious_frames);
1922         if (plugins[args.pa_plugin].pl_topdisplay != NULL)
1923                 plugins[args.pa_plugin].pl_topdisplay();
1924         PMCSTAT_PRINTEND();
1925 }
1926
1927 /*
1928  * Find the next pmc index to display.
1929  */
1930
1931 static void
1932 pmcstat_changefilter(void)
1933 {
1934         int pmcin;
1935         struct pmcstat_pmcrecord *pmcr;
1936
1937         /*
1938          * Find the next merge target.
1939          */
1940         if (pmcstat_mergepmc) {
1941                 pmcin = pmcstat_pmcinfilter;
1942
1943                 do {
1944                         pmcr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1945                         if (pmcr == NULL || pmcr == pmcr->pr_merge)
1946                                 break;
1947
1948                         pmcstat_pmcinfilter++;
1949                         if (pmcstat_pmcinfilter >= pmcstat_npmcs)
1950                                 pmcstat_pmcinfilter = 0;
1951
1952                 } while (pmcstat_pmcinfilter != pmcin);
1953         }
1954 }
1955
1956 /*
1957  * Top mode keypress.
1958  */
1959
1960 int
1961 pmcstat_keypress_log(void)
1962 {
1963         int c, ret = 0;
1964         WINDOW *w;
1965
1966         w = newwin(1, 0, 1, 0);
1967         c = wgetch(w);
1968         wprintw(w, "Key: %c => ", c);
1969         switch (c) {
1970         case 'c':
1971                 wprintw(w, "enter mode 'd' or 'a' => ");
1972                 c = wgetch(w);
1973                 if (c == 'd') {
1974                         args.pa_topmode = PMCSTAT_TOP_DELTA;
1975                         wprintw(w, "switching to delta mode");
1976                 } else {
1977                         args.pa_topmode = PMCSTAT_TOP_ACCUM;
1978                         wprintw(w, "switching to accumulation mode");
1979                 }
1980                 break;
1981         case 'm':
1982                 pmcstat_mergepmc = !pmcstat_mergepmc;
1983                 /*
1984                  * Changing merge state require data reset.
1985                  */
1986                 if (plugins[args.pa_plugin].pl_shutdown != NULL)
1987                         plugins[args.pa_plugin].pl_shutdown(NULL);
1988                 pmcstat_stats_reset(0);
1989                 if (plugins[args.pa_plugin].pl_init != NULL)
1990                         plugins[args.pa_plugin].pl_init();
1991
1992                 /* Update filter to be on a merge target. */
1993                 pmcstat_changefilter();
1994                 wprintw(w, "merge PMC %s", pmcstat_mergepmc ? "on" : "off");
1995                 break;
1996         case 'n':
1997                 /* Close current plugin. */
1998                 if (plugins[args.pa_plugin].pl_shutdown != NULL)
1999                         plugins[args.pa_plugin].pl_shutdown(NULL);
2000
2001                 /* Find next top display available. */
2002                 do {
2003                         args.pa_plugin++;
2004                         if (plugins[args.pa_plugin].pl_name == NULL)
2005                                 args.pa_plugin = 0;
2006                 } while (plugins[args.pa_plugin].pl_topdisplay == NULL);
2007
2008                 /* Open new plugin. */
2009                 pmcstat_stats_reset(0);
2010                 if (plugins[args.pa_plugin].pl_init != NULL)
2011                         plugins[args.pa_plugin].pl_init();
2012                 wprintw(w, "switching to plugin %s",
2013                     plugins[args.pa_plugin].pl_name);
2014                 break;
2015         case 'p':
2016                 pmcstat_pmcinfilter++;
2017                 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
2018                         pmcstat_pmcinfilter = 0;
2019                 pmcstat_changefilter();
2020                 wprintw(w, "switching to PMC %s.%d",
2021                     pmcstat_pmcindex_to_name(pmcstat_pmcinfilter),
2022                     pmcstat_pmcinfilter);
2023                 break;
2024         case ' ':
2025                 pmcstat_pause = !pmcstat_pause;
2026                 if (pmcstat_pause)
2027                         wprintw(w, "pause => press space again to continue");
2028                 break;
2029         case 'q':
2030                 wprintw(w, "exiting...");
2031                 ret = 1;
2032                 break;
2033         default:
2034                 if (plugins[args.pa_plugin].pl_topkeypress != NULL)
2035                         if (plugins[args.pa_plugin].pl_topkeypress(c, w))
2036                                 ret = 1;
2037         }
2038
2039         wrefresh(w);
2040         delwin(w);
2041         return ret;
2042 }
2043
2044
2045 /*
2046  * Top mode display.
2047  */
2048
2049 void
2050 pmcstat_display_log(void)
2051 {
2052
2053         pmcstat_refresh_top();
2054
2055         /* Reset everythings if delta mode. */
2056         if (args.pa_topmode == PMCSTAT_TOP_DELTA) {
2057                 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2058                         plugins[args.pa_plugin].pl_shutdown(NULL);
2059                 pmcstat_stats_reset(0);
2060                 if (plugins[args.pa_plugin].pl_init != NULL)
2061                         plugins[args.pa_plugin].pl_init();
2062         }
2063
2064 }
2065
2066 /*
2067  * Configure a plugins.
2068  */
2069
2070 void
2071 pmcstat_pluginconfigure_log(char *opt)
2072 {
2073
2074         if (strncmp(opt, "threshold=", 10) == 0) {
2075                 pmcstat_threshold = atof(opt+10);
2076         } else {
2077                 if (plugins[args.pa_plugin].pl_configure != NULL) {
2078                         if (!plugins[args.pa_plugin].pl_configure(opt))
2079                                 err(EX_USAGE,
2080                                     "ERROR: unknown option <%s>.", opt);
2081                 }
2082         }
2083 }
2084
2085 /*
2086  * Initialize module.
2087  */
2088
2089 void
2090 pmcstat_initialize_logging(void)
2091 {
2092         int i;
2093
2094         /* use a convenient format for 'ldd' output */
2095         if (setenv("LD_TRACE_LOADED_OBJECTS_FMT1","%o \"%p\" %x\n",1) != 0)
2096                 err(EX_OSERR, "ERROR: Cannot setenv");
2097
2098         /* Initialize hash tables */
2099         pmcstat_string_initialize();
2100         for (i = 0; i < PMCSTAT_NHASH; i++) {
2101                 LIST_INIT(&pmcstat_image_hash[i]);
2102                 LIST_INIT(&pmcstat_process_hash[i]);
2103         }
2104
2105         /*
2106          * Create a fake 'process' entry for the kernel with pid -1.
2107          * hwpmc(4) will subsequently inform us about where the kernel
2108          * and any loaded kernel modules are mapped.
2109          */
2110         if ((pmcstat_kernproc = pmcstat_process_lookup((pid_t) -1,
2111                  PMCSTAT_ALLOCATE)) == NULL)
2112                 err(EX_OSERR, "ERROR: Cannot initialize logging");
2113
2114         /* PMC count. */
2115         pmcstat_npmcs = 0;
2116
2117         /* Merge PMC with same name. */
2118         pmcstat_mergepmc = args.pa_mergepmc;
2119
2120         /*
2121          * Initialize plugins
2122          */
2123
2124         if (plugins[args.pa_pplugin].pl_init != NULL)
2125                 plugins[args.pa_pplugin].pl_init();
2126         if (plugins[args.pa_plugin].pl_init != NULL)
2127                 plugins[args.pa_plugin].pl_init();
2128 }
2129
2130 /*
2131  * Shutdown module.
2132  */
2133
2134 void
2135 pmcstat_shutdown_logging(void)
2136 {
2137         int i;
2138         FILE *mf;
2139         struct pmcstat_image *pi, *pitmp;
2140         struct pmcstat_process *pp, *pptmp;
2141         struct pmcstat_pcmap *ppm, *ppmtmp;
2142
2143         /* determine where to send the map file */
2144         mf = NULL;
2145         if (args.pa_mapfilename != NULL)
2146                 mf = (strcmp(args.pa_mapfilename, "-") == 0) ?
2147                     args.pa_printfile : fopen(args.pa_mapfilename, "w");
2148
2149         if (mf == NULL && args.pa_flags & FLAG_DO_GPROF &&
2150             args.pa_verbosity >= 2)
2151                 mf = args.pa_printfile;
2152
2153         if (mf)
2154                 (void) fprintf(mf, "MAP:\n");
2155
2156         /*
2157          * Shutdown the plugins
2158          */
2159
2160         if (plugins[args.pa_plugin].pl_shutdown != NULL)
2161                 plugins[args.pa_plugin].pl_shutdown(mf);
2162         if (plugins[args.pa_pplugin].pl_shutdown != NULL)
2163                 plugins[args.pa_pplugin].pl_shutdown(mf);
2164
2165         for (i = 0; i < PMCSTAT_NHASH; i++) {
2166                 LIST_FOREACH_SAFE(pi, &pmcstat_image_hash[i], pi_next,
2167                     pitmp) {
2168                         if (plugins[args.pa_plugin].pl_shutdownimage != NULL)
2169                                 plugins[args.pa_plugin].pl_shutdownimage(pi);
2170                         if (plugins[args.pa_pplugin].pl_shutdownimage != NULL)
2171                                 plugins[args.pa_pplugin].pl_shutdownimage(pi);
2172
2173                         free(pi->pi_symbols);
2174                         if (pi->pi_addr2line != NULL)
2175                                 pclose(pi->pi_addr2line);
2176                         LIST_REMOVE(pi, pi_next);
2177                         free(pi);
2178                 }
2179
2180                 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[i], pp_next,
2181                     pptmp) {
2182                         TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
2183                                 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
2184                                 free(ppm);
2185                         }
2186                         LIST_REMOVE(pp, pp_next);
2187                         free(pp);
2188                 }
2189         }
2190
2191         pmcstat_string_shutdown();
2192
2193         /*
2194          * Print errors unless -q was specified.  Print all statistics
2195          * if verbosity > 1.
2196          */
2197 #define PRINT(N,V) do {                                                 \
2198                 if (pmcstat_stats.ps_##V || args.pa_verbosity >= 2)     \
2199                         (void) fprintf(args.pa_printfile, " %-40s %d\n",\
2200                             N, pmcstat_stats.ps_##V);                   \
2201         } while (0)
2202
2203         if (args.pa_verbosity >= 1 && (args.pa_flags & FLAG_DO_ANALYSIS)) {
2204                 (void) fprintf(args.pa_printfile, "CONVERSION STATISTICS:\n");
2205                 PRINT("#exec/a.out", exec_aout);
2206                 PRINT("#exec/elf", exec_elf);
2207                 PRINT("#exec/unknown", exec_indeterminable);
2208                 PRINT("#exec handling errors", exec_errors);
2209                 PRINT("#samples/total", samples_total);
2210                 PRINT("#samples/unclaimed", samples_unknown_offset);
2211                 PRINT("#samples/unknown-object", samples_indeterminable);
2212                 PRINT("#samples/unknown-function", samples_unknown_function);
2213                 PRINT("#callchain/dubious-frames", callchain_dubious_frames);
2214         }
2215
2216         if (mf)
2217                 (void) fclose(mf);
2218 }