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_annotate_cg.h"
 153 #include "pmcpl_calltree.h"
 154
 155 static struct pmc_plugins  {
 156         const char      *pl_name;       /* name */
 157
 158         /* configure */
 159         int (*pl_configure)(char *opt);
 160
 161         /* init and shutdown */
 162         int (*pl_init)(void);
 163         void (*pl_shutdown)(FILE *mf);
 164
 165         /* sample processing */
 166         void (*pl_process)(struct pmcstat_process *pp,
 167             struct pmcstat_pmcrecord *pmcr, uint32_t nsamples,
 168             uintfptr_t *cc, int usermode, uint32_t cpu);
 169
 170         /* image */
 171         void (*pl_initimage)(struct pmcstat_image *pi);
 172         void (*pl_shutdownimage)(struct pmcstat_image *pi);
 173
 174         /* pmc */
 175         void (*pl_newpmc)(pmcstat_interned_string ps,
 176                 struct pmcstat_pmcrecord *pr);
 177
 178         /* top display */
 179         void (*pl_topdisplay)(void);
 180
 181         /* top keypress */
 182         int (*pl_topkeypress)(int c, WINDOW *w);
 183
 184 } plugins[] = {
 185         {
 186                 .pl_name                = "none",
 187         },
 188         {
 189                 .pl_name                = "callgraph",
 190                 .pl_init                = pmcpl_cg_init,
 191                 .pl_shutdown            = pmcpl_cg_shutdown,
 192                 .pl_process             = pmcpl_cg_process,
 193                 .pl_topkeypress         = pmcpl_cg_topkeypress,
 194                 .pl_topdisplay          = pmcpl_cg_topdisplay
 195         },
 196         {
 197                 .pl_name                = "gprof",
 198                 .pl_shutdown            = pmcpl_gmon_shutdown,
 199                 .pl_process             = pmcpl_gmon_process,
 200                 .pl_initimage           = pmcpl_gmon_initimage,
 201                 .pl_shutdownimage       = pmcpl_gmon_shutdownimage,
 202                 .pl_newpmc              = pmcpl_gmon_newpmc
 203         },
 204         {
 205                 .pl_name                = "annotate",
 206                 .pl_process             = pmcpl_annotate_process
 207         },
 208         {
 209                 .pl_name                = "calltree",
 210                 .pl_configure           = pmcpl_ct_configure,
 211                 .pl_init                = pmcpl_ct_init,
 212                 .pl_shutdown            = pmcpl_ct_shutdown,
 213                 .pl_process             = pmcpl_ct_process,
 214                 .pl_topkeypress         = pmcpl_ct_topkeypress,
 215                 .pl_topdisplay          = pmcpl_ct_topdisplay
 216         },
 217         {
 218                 .pl_name                = "annotate_cg",
 219                 .pl_process             = pmcpl_annotate_cg_process
 220         },
 221
 222         {
 223                 .pl_name                = NULL
 224         }
 225 };
 226
 227 static int pmcstat_mergepmc;
 228
 229 int pmcstat_pmcinfilter = 0; /* PMC filter for top mode. */
 230 float pmcstat_threshold = 0.5; /* Cost filter for top mode. */
 231
 232 /*
 233  * Prototypes
 234  */
 235
 236 static struct pmcstat_image *pmcstat_image_from_path(pmcstat_interned_string
 237     _path, int _iskernelmodule);
 238 static void pmcstat_image_get_aout_params(struct pmcstat_image *_image);
 239 static void pmcstat_image_get_elf_params(struct pmcstat_image *_image);
 240 static void     pmcstat_image_link(struct pmcstat_process *_pp,
 241     struct pmcstat_image *_i, uintfptr_t _lpc);
 242
 243 static void     pmcstat_pmcid_add(pmc_id_t _pmcid,
 244     pmcstat_interned_string _name);
 245
 246 static void     pmcstat_process_aout_exec(struct pmcstat_process *_pp,
 247     struct pmcstat_image *_image, uintfptr_t _entryaddr);
 248 static void     pmcstat_process_elf_exec(struct pmcstat_process *_pp,
 249     struct pmcstat_image *_image, uintfptr_t _entryaddr);
 250 static void     pmcstat_process_exec(struct pmcstat_process *_pp,
 251     pmcstat_interned_string _path, uintfptr_t _entryaddr);
 252 static struct pmcstat_process *pmcstat_process_lookup(pid_t _pid,
 253     int _allocate);
 254 static int      pmcstat_string_compute_hash(const char *_string);
 255 static void pmcstat_string_initialize(void);
 256 static int      pmcstat_string_lookup_hash(pmcstat_interned_string _is);
 257 static void pmcstat_string_shutdown(void);
 258 static void pmcstat_stats_reset(int _reset_global);
 259
 260 /*
 261  * A simple implementation of interned strings.  Each interned string
 262  * is assigned a unique address, so that subsequent string compares
 263  * can be done by a simple pointer comparison instead of using
 264  * strcmp().  This speeds up hash table lookups and saves memory if
 265  * duplicate strings are the norm.
 266  */
 267 struct pmcstat_string {
 268         LIST_ENTRY(pmcstat_string)      ps_next;        /* hash link */
 269         int             ps_len;
 270         int             ps_hash;
 271         char            *ps_string;
 272 };
 273
 274 static LIST_HEAD(,pmcstat_string)       pmcstat_string_hash[PMCSTAT_NHASH];
 275
 276 /*
 277  * PMC count.
 278  */
 279 int pmcstat_npmcs;
 280
 281 /*
 282  * PMC Top mode pause state.
 283  */
 284 static int pmcstat_pause;
 285
 286 static void
 287 pmcstat_stats_reset(int reset_global)
 288 {
 289         struct pmcstat_pmcrecord *pr;
 290
 291         /* Flush PMCs stats. */
 292         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
 293                 pr->pr_samples = 0;
 294                 pr->pr_dubious_frames = 0;
 295         }
 296         ps_samples_period = 0;
 297
 298         /* Flush global stats. */
 299         if (reset_global)
 300                 bzero(&pmcstat_stats, sizeof(struct pmcstat_stats));
 301 }
 302
 303 /*
 304  * Compute a 'hash' value for a string.
 305  */
 306
 307 static int
 308 pmcstat_string_compute_hash(const char *s)
 309 {
 310         unsigned hash;
 311
 312         for (hash = 2166136261; *s; s++)
 313                 hash = (hash ^ *s) * 16777619;
 314
 315         return (hash & PMCSTAT_HASH_MASK);
 316 }
 317
 318 /*
 319  * Intern a copy of string 's', and return a pointer to the
 320  * interned structure.
 321  */
 322
 323 pmcstat_interned_string
 324 pmcstat_string_intern(const char *s)
 325 {
 326         struct pmcstat_string *ps;
 327         const struct pmcstat_string *cps;
 328         int hash, len;
 329
 330         if ((cps = pmcstat_string_lookup(s)) != NULL)
 331                 return (cps);
 332
 333         hash = pmcstat_string_compute_hash(s);
 334         len  = strlen(s);
 335
 336         if ((ps = malloc(sizeof(*ps))) == NULL)
 337                 err(EX_OSERR, "ERROR: Could not intern string");
 338         ps->ps_len = len;
 339         ps->ps_hash = hash;
 340         ps->ps_string = strdup(s);
 341         LIST_INSERT_HEAD(&pmcstat_string_hash[hash], ps, ps_next);
 342         return ((pmcstat_interned_string) ps);
 343 }
 344
 345 const char *
 346 pmcstat_string_unintern(pmcstat_interned_string str)
 347 {
 348         const char *s;
 349
 350         s = ((const struct pmcstat_string *) str)->ps_string;
 351         return (s);
 352 }
 353
 354 pmcstat_interned_string
 355 pmcstat_string_lookup(const char *s)
 356 {
 357         struct pmcstat_string *ps;
 358         int hash, len;
 359
 360         hash = pmcstat_string_compute_hash(s);
 361         len = strlen(s);
 362
 363         LIST_FOREACH(ps, &pmcstat_string_hash[hash], ps_next)
 364             if (ps->ps_len == len && ps->ps_hash == hash &&
 365                 strcmp(ps->ps_string, s) == 0)
 366                     return (ps);
 367         return (NULL);
 368 }
 369
 370 static int
 371 pmcstat_string_lookup_hash(pmcstat_interned_string s)
 372 {
 373         const struct pmcstat_string *ps;
 374
 375         ps = (const struct pmcstat_string *) s;
 376         return (ps->ps_hash);
 377 }
 378
 379 /*
 380  * Initialize the string interning facility.
 381  */
 382
 383 static void
 384 pmcstat_string_initialize(void)
 385 {
 386         int i;
 387
 388         for (i = 0; i < PMCSTAT_NHASH; i++)
 389                 LIST_INIT(&pmcstat_string_hash[i]);
 390 }
 391
 392 /*
 393  * Destroy the string table, free'ing up space.
 394  */
 395
 396 static void
 397 pmcstat_string_shutdown(void)
 398 {
 399         int i;
 400         struct pmcstat_string *ps, *pstmp;
 401
 402         for (i = 0; i < PMCSTAT_NHASH; i++)
 403                 LIST_FOREACH_SAFE(ps, &pmcstat_string_hash[i], ps_next,
 404                     pstmp) {
 405                         LIST_REMOVE(ps, ps_next);
 406                         free(ps->ps_string);
 407                         free(ps);
 408                 }
 409 }
 410
 411 /*
 412  * Determine whether a given executable image is an A.OUT object, and
 413  * if so, fill in its parameters from the text file.
 414  * Sets image->pi_type.
 415  */
 416
 417 static void
 418 pmcstat_image_get_aout_params(struct pmcstat_image *image)
 419 {
 420         int fd;
 421         ssize_t nbytes;
 422         struct exec ex;
 423         const char *path;
 424         char buffer[PATH_MAX];
 425
 426         path = pmcstat_string_unintern(image->pi_execpath);
 427         assert(path != NULL);
 428
 429         if (image->pi_iskernelmodule)
 430                 errx(EX_SOFTWARE,
 431                     "ERROR: a.out kernel modules are unsupported \"%s\"", path);
 432
 433         (void) snprintf(buffer, sizeof(buffer), "%s%s",
 434             args.pa_fsroot, path);
 435
 436         if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
 437             (nbytes = read(fd, &ex, sizeof(ex))) < 0) {
 438                 if (args.pa_verbosity >= 2)
 439                         warn("WARNING: Cannot determine type of \"%s\"",
 440                             path);
 441                 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
 442                 if (fd != -1)
 443                         (void) close(fd);
 444                 return;
 445         }
 446
 447         (void) close(fd);
 448
 449         if ((unsigned) nbytes != sizeof(ex) ||
 450             N_BADMAG(ex))
 451                 return;
 452
 453         image->pi_type = PMCSTAT_IMAGE_AOUT;
 454
 455         /* TODO: the rest of a.out processing */
 456
 457         return;
 458 }
 459
 460 /*
 461  * Helper function.
 462  */
 463
 464 static int
 465 pmcstat_symbol_compare(const void *a, const void *b)
 466 {
 467         const struct pmcstat_symbol *sym1, *sym2;
 468
 469         sym1 = (const struct pmcstat_symbol *) a;
 470         sym2 = (const struct pmcstat_symbol *) b;
 471
 472         if (sym1->ps_end <= sym2->ps_start)
 473                 return (-1);
 474         if (sym1->ps_start >= sym2->ps_end)
 475                 return (1);
 476         return (0);
 477 }
 478
 479 /*
 480  * Map an address to a symbol in an image.
 481  */
 482
 483 struct pmcstat_symbol *
 484 pmcstat_symbol_search(struct pmcstat_image *image, uintfptr_t addr)
 485 {
 486         struct pmcstat_symbol sym;
 487
 488         if (image->pi_symbols == NULL)
 489                 return (NULL);
 490
 491         sym.ps_name  = NULL;
 492         sym.ps_start = addr;
 493         sym.ps_end   = addr + 1;
 494
 495         return (bsearch((void *) &sym, image->pi_symbols,
 496                     image->pi_symcount, sizeof(struct pmcstat_symbol),
 497                     pmcstat_symbol_compare));
 498 }
 499
 500 /*
 501  * Add the list of symbols in the given section to the list associated
 502  * with the object.
 503  */
 504 static void
 505 pmcstat_image_add_symbols(struct pmcstat_image *image, Elf *e,
 506     Elf_Scn *scn, GElf_Shdr *sh)
 507 {
 508         int firsttime;
 509         size_t n, newsyms, nshsyms, nfuncsyms;
 510         struct pmcstat_symbol *symptr;
 511         char *fnname;
 512         GElf_Sym sym;
 513         Elf_Data *data;
 514
 515         if ((data = elf_getdata(scn, NULL)) == NULL)
 516                 return;
 517
 518         /*
 519          * Determine the number of functions named in this
 520          * section.
 521          */
 522
 523         nshsyms = sh->sh_size / sh->sh_entsize;
 524         for (n = nfuncsyms = 0; n < nshsyms; n++) {
 525                 if (gelf_getsym(data, (int) n, &sym) != &sym)
 526                         return;
 527                 if (GELF_ST_TYPE(sym.st_info) == STT_FUNC)
 528                         nfuncsyms++;
 529         }
 530
 531         if (nfuncsyms == 0)
 532                 return;
 533
 534         /*
 535          * Allocate space for the new entries.
 536          */
 537         firsttime = image->pi_symbols == NULL;
 538         symptr = reallocarray(image->pi_symbols,
 539             image->pi_symcount + nfuncsyms, sizeof(*symptr));
 540         if (symptr == image->pi_symbols) /* realloc() failed. */
 541                 return;
 542         image->pi_symbols = symptr;
 543
 544         /*
 545          * Append new symbols to the end of the current table.
 546          */
 547         symptr += image->pi_symcount;
 548
 549         for (n = newsyms = 0; n < nshsyms; n++) {
 550                 if (gelf_getsym(data, (int) n, &sym) != &sym)
 551                         return;
 552                 if (GELF_ST_TYPE(sym.st_info) != STT_FUNC)
 553                         continue;
 554                 if (sym.st_shndx == STN_UNDEF)
 555                         continue;
 556
 557                 if (!firsttime && pmcstat_symbol_search(image, sym.st_value))
 558                         continue; /* We've seen this symbol already. */
 559
 560                 if ((fnname = elf_strptr(e, sh->sh_link, sym.st_name))
 561                     == NULL)
 562                         continue;
 563 #ifdef __arm__
 564                 /* Remove spurious ARM function name. */
 565                 if (fnname[0] == '$' &&
 566                     (fnname[1] == 'a' || fnname[1] == 't' ||
 567                     fnname[1] == 'd') &&
 568                     fnname[2] == '\0')
 569                         continue;
 570 #endif
 571
 572                 symptr->ps_name  = pmcstat_string_intern(fnname);
 573                 symptr->ps_start = sym.st_value - image->pi_vaddr;
 574                 symptr->ps_end   = symptr->ps_start + sym.st_size;
 575                 symptr++;
 576
 577                 newsyms++;
 578         }
 579
 580         image->pi_symcount += newsyms;
 581         if (image->pi_symcount == 0)
 582                 return;
 583
 584         assert(newsyms <= nfuncsyms);
 585
 586         /*
 587          * Return space to the system if there were duplicates.
 588          */
 589         if (newsyms < nfuncsyms)
 590                 image->pi_symbols = reallocarray(image->pi_symbols,
 591                     image->pi_symcount, sizeof(*symptr));
 592
 593         /*
 594          * Keep the list of symbols sorted.
 595          */
 596         qsort(image->pi_symbols, image->pi_symcount, sizeof(*symptr),
 597             pmcstat_symbol_compare);
 598
 599         /*
 600          * Deal with function symbols that have a size of 'zero' by
 601          * making them extend to the next higher address.  These
 602          * symbols are usually defined in assembly code.
 603          */
 604         for (symptr = image->pi_symbols;
 605              symptr < image->pi_symbols + (image->pi_symcount - 1);
 606              symptr++)
 607                 if (symptr->ps_start == symptr->ps_end)
 608                         symptr->ps_end = (symptr+1)->ps_start;
 609 }
 610
 611 /*
 612  * Examine an ELF file to determine the size of its text segment.
 613  * Sets image->pi_type if anything conclusive can be determined about
 614  * this image.
 615  */
 616
 617 static void
 618 pmcstat_image_get_elf_params(struct pmcstat_image *image)
 619 {
 620         int fd;
 621         size_t i, nph, nsh;
 622         const char *path, *elfbase;
 623         char *p, *endp;
 624         uintfptr_t minva, maxva;
 625         Elf *e;
 626         Elf_Scn *scn;
 627         GElf_Ehdr eh;
 628         GElf_Phdr ph;
 629         GElf_Shdr sh;
 630         enum pmcstat_image_type image_type;
 631         char buffer[PATH_MAX];
 632
 633         assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
 634
 635         image->pi_start = minva = ~(uintfptr_t) 0;
 636         image->pi_end = maxva = (uintfptr_t) 0;
 637         image->pi_type = image_type = PMCSTAT_IMAGE_INDETERMINABLE;
 638         image->pi_isdynamic = 0;
 639         image->pi_dynlinkerpath = NULL;
 640         image->pi_vaddr = 0;
 641
 642         path = pmcstat_string_unintern(image->pi_execpath);
 643         assert(path != NULL);
 644
 645         /*
 646          * Look for kernel modules under FSROOT/KERNELPATH/NAME,
 647          * and user mode executable objects under FSROOT/PATHNAME.
 648          */
 649         if (image->pi_iskernelmodule)
 650                 (void) snprintf(buffer, sizeof(buffer), "%s%s/%s",
 651                     args.pa_fsroot, args.pa_kernel, path);
 652         else
 653                 (void) snprintf(buffer, sizeof(buffer), "%s%s",
 654                     args.pa_fsroot, path);
 655
 656         e = NULL;
 657         if ((fd = open(buffer, O_RDONLY, 0)) < 0 ||
 658             (e = elf_begin(fd, ELF_C_READ, NULL)) == NULL ||
 659             (elf_kind(e) != ELF_K_ELF)) {
 660                 if (args.pa_verbosity >= 2)
 661                         warnx("WARNING: Cannot determine the type of \"%s\".",
 662                             buffer);
 663                 goto done;
 664         }
 665
 666         if (gelf_getehdr(e, &eh) != &eh) {
 667                 warnx(
 668                     "WARNING: Cannot retrieve the ELF Header for \"%s\": %s.",
 669                     buffer, elf_errmsg(-1));
 670                 goto done;
 671         }
 672
 673         if (eh.e_type != ET_EXEC && eh.e_type != ET_DYN &&
 674             !(image->pi_iskernelmodule && eh.e_type == ET_REL)) {
 675                 warnx("WARNING: \"%s\" is of an unsupported ELF type.",
 676                     buffer);
 677                 goto done;
 678         }
 679
 680         image_type = eh.e_ident[EI_CLASS] == ELFCLASS32 ?
 681             PMCSTAT_IMAGE_ELF32 : PMCSTAT_IMAGE_ELF64;
 682
 683         /*
 684          * Determine the virtual address where an executable would be
 685          * loaded.  Additionally, for dynamically linked executables,
 686          * save the pathname to the runtime linker.
 687          */
 688         if (eh.e_type == ET_EXEC) {
 689                 if (elf_getphnum(e, &nph) == 0) {
 690                         warnx(
 691 "WARNING: Could not determine the number of program headers in \"%s\": %s.",
 692                             buffer,
 693                             elf_errmsg(-1));
 694                         goto done;
 695                 }
 696                 for (i = 0; i < eh.e_phnum; i++) {
 697                         if (gelf_getphdr(e, i, &ph) != &ph) {
 698                                 warnx(
 699 "WARNING: Retrieval of PHDR entry #%ju in \"%s\" failed: %s.",
 700                                     (uintmax_t) i, buffer, elf_errmsg(-1));
 701                                 goto done;
 702                         }
 703                         switch (ph.p_type) {
 704                         case PT_DYNAMIC:
 705                                 image->pi_isdynamic = 1;
 706                                 break;
 707                         case PT_INTERP:
 708                                 if ((elfbase = elf_rawfile(e, NULL)) == NULL) {
 709                                         warnx(
 710 "WARNING: Cannot retrieve the interpreter for \"%s\": %s.",
 711                                             buffer, elf_errmsg(-1));
 712                                         goto done;
 713                                 }
 714                                 image->pi_dynlinkerpath =
 715                                     pmcstat_string_intern(elfbase +
 716                                         ph.p_offset);
 717                                 break;
 718                         case PT_LOAD:
 719                                 if ((ph.p_flags & PF_X) != 0 &&
 720                                     (ph.p_offset & (-ph.p_align)) == 0)
 721                                         image->pi_vaddr = ph.p_vaddr & (-ph.p_align);
 722                                 break;
 723                         }
 724                 }
 725         }
 726
 727         /*
 728          * Get the min and max VA associated with this ELF object.
 729          */
 730         if (elf_getshnum(e, &nsh) == 0) {
 731                 warnx(
 732 "WARNING: Could not determine the number of sections for \"%s\": %s.",
 733                     buffer, elf_errmsg(-1));
 734                 goto done;
 735         }
 736
 737         for (i = 0; i < nsh; i++) {
 738                 if ((scn = elf_getscn(e, i)) == NULL ||
 739                     gelf_getshdr(scn, &sh) != &sh) {
 740                         warnx(
 741 "WARNING: Could not retrieve section header #%ju in \"%s\": %s.",
 742                             (uintmax_t) i, buffer, elf_errmsg(-1));
 743                         goto done;
 744                 }
 745                 if (sh.sh_flags & SHF_EXECINSTR) {
 746                         minva = min(minva, sh.sh_addr);
 747                         maxva = max(maxva, sh.sh_addr + sh.sh_size);
 748                 }
 749                 if (sh.sh_type == SHT_SYMTAB || sh.sh_type == SHT_DYNSYM)
 750                         pmcstat_image_add_symbols(image, e, scn, &sh);
 751         }
 752
 753         image->pi_start = minva;
 754         image->pi_end   = maxva;
 755         image->pi_type  = image_type;
 756         image->pi_fullpath = pmcstat_string_intern(buffer);
 757
 758         /* Build display name
 759          */
 760         endp = buffer;
 761         for (p = buffer; *p; p++)
 762                 if (*p == '/')
 763                         endp = p+1;
 764         image->pi_name = pmcstat_string_intern(endp);
 765
 766  done:
 767         (void) elf_end(e);
 768         if (fd >= 0)
 769                 (void) close(fd);
 770         return;
 771 }
 772
 773 /*
 774  * Given an image descriptor, determine whether it is an ELF, or AOUT.
 775  * If no handler claims the image, set its type to 'INDETERMINABLE'.
 776  */
 777
 778 void
 779 pmcstat_image_determine_type(struct pmcstat_image *image)
 780 {
 781         assert(image->pi_type == PMCSTAT_IMAGE_UNKNOWN);
 782
 783         /* Try each kind of handler in turn */
 784         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
 785                 pmcstat_image_get_elf_params(image);
 786         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
 787                 pmcstat_image_get_aout_params(image);
 788
 789         /*
 790          * Otherwise, remember that we tried to determine
 791          * the object's type and had failed.
 792          */
 793         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
 794                 image->pi_type = PMCSTAT_IMAGE_INDETERMINABLE;
 795 }
 796
 797 /*
 798  * Locate an image descriptor given an interned path, adding a fresh
 799  * descriptor to the cache if necessary.  This function also finds a
 800  * suitable name for this image's sample file.
 801  *
 802  * We defer filling in the file format specific parts of the image
 803  * structure till the time we actually see a sample that would fall
 804  * into this image.
 805  */
 806
 807 static struct pmcstat_image *
 808 pmcstat_image_from_path(pmcstat_interned_string internedpath,
 809     int iskernelmodule)
 810 {
 811         int hash;
 812         struct pmcstat_image *pi;
 813
 814         hash = pmcstat_string_lookup_hash(internedpath);
 815
 816         /* First, look for an existing entry. */
 817         LIST_FOREACH(pi, &pmcstat_image_hash[hash], pi_next)
 818             if (pi->pi_execpath == internedpath &&
 819                   pi->pi_iskernelmodule == iskernelmodule)
 820                     return (pi);
 821
 822         /*
 823          * Allocate a new entry and place it at the head of the hash
 824          * and LRU lists.
 825          */
 826         pi = malloc(sizeof(*pi));
 827         if (pi == NULL)
 828                 return (NULL);
 829
 830         pi->pi_type = PMCSTAT_IMAGE_UNKNOWN;
 831         pi->pi_execpath = internedpath;
 832         pi->pi_start = ~0;
 833         pi->pi_end = 0;
 834         pi->pi_entry = 0;
 835         pi->pi_vaddr = 0;
 836         pi->pi_isdynamic = 0;
 837         pi->pi_iskernelmodule = iskernelmodule;
 838         pi->pi_dynlinkerpath = NULL;
 839         pi->pi_symbols = NULL;
 840         pi->pi_symcount = 0;
 841         pi->pi_addr2line = NULL;
 842
 843         if (plugins[args.pa_pplugin].pl_initimage != NULL)
 844                 plugins[args.pa_pplugin].pl_initimage(pi);
 845         if (plugins[args.pa_plugin].pl_initimage != NULL)
 846                 plugins[args.pa_plugin].pl_initimage(pi);
 847
 848         LIST_INSERT_HEAD(&pmcstat_image_hash[hash], pi, pi_next);
 849
 850         return (pi);
 851 }
 852
 853 /*
 854  * Record the fact that PC values from 'start' to 'end' come from
 855  * image 'image'.
 856  */
 857
 858 static void
 859 pmcstat_image_link(struct pmcstat_process *pp, struct pmcstat_image *image,
 860     uintfptr_t start)
 861 {
 862         struct pmcstat_pcmap *pcm, *pcmnew;
 863         uintfptr_t offset;
 864
 865         assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN &&
 866             image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE);
 867
 868         if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
 869                 err(EX_OSERR, "ERROR: Cannot create a map entry");
 870
 871         /*
 872          * Adjust the map entry to only cover the text portion
 873          * of the object.
 874          */
 875
 876         offset = start - image->pi_vaddr;
 877         pcmnew->ppm_lowpc  = image->pi_start + offset;
 878         pcmnew->ppm_highpc = image->pi_end + offset;
 879         pcmnew->ppm_image  = image;
 880
 881         assert(pcmnew->ppm_lowpc < pcmnew->ppm_highpc);
 882
 883         /* Overlapped mmap()'s are assumed to never occur. */
 884         TAILQ_FOREACH(pcm, &pp->pp_map, ppm_next)
 885             if (pcm->ppm_lowpc >= pcmnew->ppm_highpc)
 886                     break;
 887
 888         if (pcm == NULL)
 889                 TAILQ_INSERT_TAIL(&pp->pp_map, pcmnew, ppm_next);
 890         else
 891                 TAILQ_INSERT_BEFORE(pcm, pcmnew, ppm_next);
 892 }
 893
 894 /*
 895  * Unmap images in the range [start..end) associated with process
 896  * 'pp'.
 897  */
 898
 899 static void
 900 pmcstat_image_unmap(struct pmcstat_process *pp, uintfptr_t start,
 901     uintfptr_t end)
 902 {
 903         struct pmcstat_pcmap *pcm, *pcmtmp, *pcmnew;
 904
 905         assert(pp != NULL);
 906         assert(start < end);
 907
 908         /*
 909          * Cases:
 910          * - we could have the range completely in the middle of an
 911          *   existing pcmap; in this case we have to split the pcmap
 912          *   structure into two (i.e., generate a 'hole').
 913          * - we could have the range covering multiple pcmaps; these
 914          *   will have to be removed.
 915          * - we could have either 'start' or 'end' falling in the
 916          *   middle of a pcmap; in this case shorten the entry.
 917          */
 918         TAILQ_FOREACH_SAFE(pcm, &pp->pp_map, ppm_next, pcmtmp) {
 919                 assert(pcm->ppm_lowpc < pcm->ppm_highpc);
 920                 if (pcm->ppm_highpc <= start)
 921                         continue;
 922                 if (pcm->ppm_lowpc >= end)
 923                         return;
 924                 if (pcm->ppm_lowpc >= start && pcm->ppm_highpc <= end) {
 925                         /*
 926                          * The current pcmap is completely inside the
 927                          * unmapped range: remove it entirely.
 928                          */
 929                         TAILQ_REMOVE(&pp->pp_map, pcm, ppm_next);
 930                         free(pcm);
 931                 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc > end) {
 932                         /*
 933                          * Split this pcmap into two; curtail the
 934                          * current map to end at [start-1], and start
 935                          * the new one at [end].
 936                          */
 937                         if ((pcmnew = malloc(sizeof(*pcmnew))) == NULL)
 938                                 err(EX_OSERR,
 939                                     "ERROR: Cannot split a map entry");
 940
 941                         pcmnew->ppm_image = pcm->ppm_image;
 942
 943                         pcmnew->ppm_lowpc = end;
 944                         pcmnew->ppm_highpc = pcm->ppm_highpc;
 945
 946                         pcm->ppm_highpc = start;
 947
 948                         TAILQ_INSERT_AFTER(&pp->pp_map, pcm, pcmnew, ppm_next);
 949
 950                         return;
 951                 } else if (pcm->ppm_lowpc < start && pcm->ppm_highpc <= end)
 952                         pcm->ppm_highpc = start;
 953                 else if (pcm->ppm_lowpc >= start && pcm->ppm_highpc > end)
 954                         pcm->ppm_lowpc = end;
 955                 else
 956                         assert(0);
 957         }
 958 }
 959
 960 /*
 961  * Resolve file name and line number for the given address.
 962  */
 963 int
 964 pmcstat_image_addr2line(struct pmcstat_image *image, uintfptr_t addr,
 965     char *sourcefile, size_t sourcefile_len, unsigned *sourceline,
 966     char *funcname, size_t funcname_len)
 967 {
 968         static int addr2line_warn = 0;
 969
 970         char *sep, cmdline[PATH_MAX], imagepath[PATH_MAX];
 971         unsigned l;
 972         int fd;
 973
 974         if (image->pi_addr2line == NULL) {
 975                 /* Try default debug file location. */
 976                 snprintf(imagepath, sizeof(imagepath),
 977                     "/usr/lib/debug/%s%s.debug",
 978                     args.pa_fsroot,
 979                     pmcstat_string_unintern(image->pi_fullpath));
 980                 fd = open(imagepath, O_RDONLY);
 981                 if (fd < 0) {
 982                         /* Old kernel symbol path. */
 983                         snprintf(imagepath, sizeof(imagepath), "%s%s.symbols",
 984                             args.pa_fsroot,
 985                             pmcstat_string_unintern(image->pi_fullpath));
 986                         fd = open(imagepath, O_RDONLY);
 987                         if (fd < 0) {
 988                                 snprintf(imagepath, sizeof(imagepath), "%s%s",
 989                                     args.pa_fsroot,
 990                                     pmcstat_string_unintern(
 991                                         image->pi_fullpath));
 992                         }
 993                 }
 994                 if (fd >= 0)
 995                         close(fd);
 996                 /*
 997                  * New addr2line support recursive inline function with -i
 998                  * but the format does not add a marker when no more entries
 999                  * are available.
1000                  */
1001                 snprintf(cmdline, sizeof(cmdline), "addr2line -Cfe \"%s\"",
1002                     imagepath);
1003                 image->pi_addr2line = popen(cmdline, "r+");
1004                 if (image->pi_addr2line == NULL) {
1005                         if (!addr2line_warn) {
1006                                 addr2line_warn = 1;
1007                                 warnx(
1008 "WARNING: addr2line is needed for source code information."
1009                                     );
1010                         }
1011                         return (0);
1012                 }
1013         }
1014
1015         if (feof(image->pi_addr2line) || ferror(image->pi_addr2line)) {
1016                 warnx("WARNING: addr2line pipe error");
1017                 pclose(image->pi_addr2line);
1018                 image->pi_addr2line = NULL;
1019                 return (0);
1020         }
1021
1022         fprintf(image->pi_addr2line, "%p\n", (void *)addr);
1023
1024         if (fgets(funcname, funcname_len, image->pi_addr2line) == NULL) {
1025                 warnx("WARNING: addr2line function name read error");
1026                 return (0);
1027         }
1028         sep = strchr(funcname, '\n');
1029         if (sep != NULL)
1030                 *sep = '\0';
1031
1032         if (fgets(sourcefile, sourcefile_len, image->pi_addr2line) == NULL) {
1033                 warnx("WARNING: addr2line source file read error");
1034                 return (0);
1035         }
1036         sep = strchr(sourcefile, ':');
1037         if (sep == NULL) {
1038                 warnx("WARNING: addr2line source line separator missing");
1039                 return (0);
1040         }
1041         *sep = '\0';
1042         l = atoi(sep+1);
1043         if (l == 0)
1044                 return (0);
1045         *sourceline = l;
1046         return (1);
1047 }
1048
1049 /*
1050  * Add a {pmcid,name} mapping.
1051  */
1052
1053 static void
1054 pmcstat_pmcid_add(pmc_id_t pmcid, pmcstat_interned_string ps)
1055 {
1056         struct pmcstat_pmcrecord *pr, *prm;
1057
1058         /* Replace an existing name for the PMC. */
1059         prm = NULL;
1060         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1061                 if (pr->pr_pmcid == pmcid) {
1062                         pr->pr_pmcname = ps;
1063                         return;
1064                 } else if (pr->pr_pmcname == ps)
1065                         prm = pr;
1066
1067         /*
1068          * Otherwise, allocate a new descriptor and call the
1069          * plugins hook.
1070          */
1071         if ((pr = malloc(sizeof(*pr))) == NULL)
1072                 err(EX_OSERR, "ERROR: Cannot allocate pmc record");
1073
1074         pr->pr_pmcid = pmcid;
1075         pr->pr_pmcname = ps;
1076         pr->pr_pmcin = pmcstat_npmcs++;
1077         pr->pr_samples = 0;
1078         pr->pr_dubious_frames = 0;
1079         pr->pr_merge = prm == NULL ? pr : prm;
1080
1081         LIST_INSERT_HEAD(&pmcstat_pmcs, pr, pr_next);
1082
1083         if (plugins[args.pa_pplugin].pl_newpmc != NULL)
1084                 plugins[args.pa_pplugin].pl_newpmc(ps, pr);
1085         if (plugins[args.pa_plugin].pl_newpmc != NULL)
1086                 plugins[args.pa_plugin].pl_newpmc(ps, pr);
1087 }
1088
1089 /*
1090  * Given a pmcid in use, find its human-readable name.
1091  */
1092
1093 const char *
1094 pmcstat_pmcid_to_name(pmc_id_t pmcid)
1095 {
1096         struct pmcstat_pmcrecord *pr;
1097
1098         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1099             if (pr->pr_pmcid == pmcid)
1100                     return (pmcstat_string_unintern(pr->pr_pmcname));
1101
1102         return NULL;
1103 }
1104
1105 /*
1106  * Convert PMC index to name.
1107  */
1108
1109 const char *
1110 pmcstat_pmcindex_to_name(int pmcin)
1111 {
1112         struct pmcstat_pmcrecord *pr;
1113
1114         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1115                 if (pr->pr_pmcin == pmcin)
1116                         return pmcstat_string_unintern(pr->pr_pmcname);
1117
1118         return NULL;
1119 }
1120
1121 /*
1122  * Return PMC record with given index.
1123  */
1124
1125 struct pmcstat_pmcrecord *
1126 pmcstat_pmcindex_to_pmcr(int pmcin)
1127 {
1128         struct pmcstat_pmcrecord *pr;
1129
1130         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next)
1131                 if (pr->pr_pmcin == pmcin)
1132                         return pr;
1133
1134         return NULL;
1135 }
1136
1137 /*
1138  * Get PMC record by id, apply merge policy.
1139  */
1140
1141 static struct pmcstat_pmcrecord *
1142 pmcstat_lookup_pmcid(pmc_id_t pmcid)
1143 {
1144         struct pmcstat_pmcrecord *pr;
1145
1146         LIST_FOREACH(pr, &pmcstat_pmcs, pr_next) {
1147                 if (pr->pr_pmcid == pmcid) {
1148                         if (pmcstat_mergepmc)
1149                                 return pr->pr_merge;
1150                         return pr;
1151                 }
1152         }
1153
1154         return NULL;
1155 }
1156
1157 /*
1158  * Associate an AOUT image with a process.
1159  */
1160
1161 static void
1162 pmcstat_process_aout_exec(struct pmcstat_process *pp,
1163     struct pmcstat_image *image, uintfptr_t entryaddr)
1164 {
1165         (void) pp;
1166         (void) image;
1167         (void) entryaddr;
1168         /* TODO Implement a.out handling */
1169 }
1170
1171 /*
1172  * Associate an ELF image with a process.
1173  */
1174
1175 static void
1176 pmcstat_process_elf_exec(struct pmcstat_process *pp,
1177     struct pmcstat_image *image, uintfptr_t entryaddr)
1178 {
1179         uintmax_t libstart;
1180         struct pmcstat_image *rtldimage;
1181
1182         assert(image->pi_type == PMCSTAT_IMAGE_ELF32 ||
1183             image->pi_type == PMCSTAT_IMAGE_ELF64);
1184
1185         /* Create a map entry for the base executable. */
1186         pmcstat_image_link(pp, image, image->pi_vaddr);
1187
1188         /*
1189          * For dynamically linked executables we need to determine
1190          * where the dynamic linker was mapped to for this process,
1191          * Subsequent executable objects that are mapped in by the
1192          * dynamic linker will be tracked by log events of type
1193          * PMCLOG_TYPE_MAP_IN.
1194          */
1195
1196         if (image->pi_isdynamic) {
1197
1198                 /*
1199                  * The runtime loader gets loaded just after the maximum
1200                  * possible heap address.  Like so:
1201                  *
1202                  * [  TEXT DATA BSS HEAP -->*RTLD  SHLIBS   <--STACK]
1203                  * ^                                                ^
1204                  * 0                               VM_MAXUSER_ADDRESS
1205
1206                  *
1207                  * The exact address where the loader gets mapped in
1208                  * will vary according to the size of the executable
1209                  * and the limits on the size of the process'es data
1210                  * segment at the time of exec().  The entry address
1211                  * recorded at process exec time corresponds to the
1212                  * 'start' address inside the dynamic linker.  From
1213                  * this we can figure out the address where the
1214                  * runtime loader's file object had been mapped to.
1215                  */
1216                 rtldimage = pmcstat_image_from_path(image->pi_dynlinkerpath, 0);
1217                 if (rtldimage == NULL) {
1218                         warnx("WARNING: Cannot find image for \"%s\".",
1219                             pmcstat_string_unintern(image->pi_dynlinkerpath));
1220                         pmcstat_stats.ps_exec_errors++;
1221                         return;
1222                 }
1223
1224                 if (rtldimage->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1225                         pmcstat_image_get_elf_params(rtldimage);
1226
1227                 if (rtldimage->pi_type != PMCSTAT_IMAGE_ELF32 &&
1228                     rtldimage->pi_type != PMCSTAT_IMAGE_ELF64) {
1229                         warnx("WARNING: rtld not an ELF object \"%s\".",
1230                             pmcstat_string_unintern(image->pi_dynlinkerpath));
1231                         return;
1232                 }
1233
1234                 libstart = entryaddr - rtldimage->pi_entry;
1235                 pmcstat_image_link(pp, rtldimage, libstart);
1236         }
1237 }
1238
1239 /*
1240  * Find the process descriptor corresponding to a PID.  If 'allocate'
1241  * is zero, we return a NULL if a pid descriptor could not be found or
1242  * a process descriptor process.  If 'allocate' is non-zero, then we
1243  * will attempt to allocate a fresh process descriptor.  Zombie
1244  * process descriptors are only removed if a fresh allocation for the
1245  * same PID is requested.
1246  */
1247
1248 static struct pmcstat_process *
1249 pmcstat_process_lookup(pid_t pid, int allocate)
1250 {
1251         uint32_t hash;
1252         struct pmcstat_pcmap *ppm, *ppmtmp;
1253         struct pmcstat_process *pp, *pptmp;
1254
1255         hash = (uint32_t) pid & PMCSTAT_HASH_MASK;      /* simplicity wins */
1256
1257         LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[hash], pp_next, pptmp)
1258                 if (pp->pp_pid == pid) {
1259                         /* Found a descriptor, check and process zombies */
1260                         if (allocate && pp->pp_isactive == 0) {
1261                                 /* remove maps */
1262                                 TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next,
1263                                     ppmtmp) {
1264                                         TAILQ_REMOVE(&pp->pp_map, ppm,
1265                                             ppm_next);
1266                                         free(ppm);
1267                                 }
1268                                 /* remove process entry */
1269                                 LIST_REMOVE(pp, pp_next);
1270                                 free(pp);
1271                                 break;
1272                         }
1273                         return (pp);
1274                 }
1275
1276         if (!allocate)
1277                 return (NULL);
1278
1279         if ((pp = malloc(sizeof(*pp))) == NULL)
1280                 err(EX_OSERR, "ERROR: Cannot allocate pid descriptor");
1281
1282         pp->pp_pid = pid;
1283         pp->pp_isactive = 1;
1284
1285         TAILQ_INIT(&pp->pp_map);
1286
1287         LIST_INSERT_HEAD(&pmcstat_process_hash[hash], pp, pp_next);
1288         return (pp);
1289 }
1290
1291 /*
1292  * Associate an image and a process.
1293  */
1294
1295 static void
1296 pmcstat_process_exec(struct pmcstat_process *pp,
1297     pmcstat_interned_string path, uintfptr_t entryaddr)
1298 {
1299         struct pmcstat_image *image;
1300
1301         if ((image = pmcstat_image_from_path(path, 0)) == NULL) {
1302                 pmcstat_stats.ps_exec_errors++;
1303                 return;
1304         }
1305
1306         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1307                 pmcstat_image_determine_type(image);
1308
1309         assert(image->pi_type != PMCSTAT_IMAGE_UNKNOWN);
1310
1311         switch (image->pi_type) {
1312         case PMCSTAT_IMAGE_ELF32:
1313         case PMCSTAT_IMAGE_ELF64:
1314                 pmcstat_stats.ps_exec_elf++;
1315                 pmcstat_process_elf_exec(pp, image, entryaddr);
1316                 break;
1317
1318         case PMCSTAT_IMAGE_AOUT:
1319                 pmcstat_stats.ps_exec_aout++;
1320                 pmcstat_process_aout_exec(pp, image, entryaddr);
1321                 break;
1322
1323         case PMCSTAT_IMAGE_INDETERMINABLE:
1324                 pmcstat_stats.ps_exec_indeterminable++;
1325                 break;
1326
1327         default:
1328                 err(EX_SOFTWARE,
1329                     "ERROR: Unsupported executable type for \"%s\"",
1330                     pmcstat_string_unintern(path));
1331         }
1332 }
1333
1334
1335 /*
1336  * Find the map entry associated with process 'p' at PC value 'pc'.
1337  */
1338
1339 struct pmcstat_pcmap *
1340 pmcstat_process_find_map(struct pmcstat_process *p, uintfptr_t pc)
1341 {
1342         struct pmcstat_pcmap *ppm;
1343
1344         TAILQ_FOREACH(ppm, &p->pp_map, ppm_next) {
1345                 if (pc >= ppm->ppm_lowpc && pc < ppm->ppm_highpc)
1346                         return (ppm);
1347                 if (pc < ppm->ppm_lowpc)
1348                         return (NULL);
1349         }
1350
1351         return (NULL);
1352 }
1353
1354 /*
1355  * Convert a hwpmc(4) log to profile information.  A system-wide
1356  * callgraph is generated if FLAG_DO_CALLGRAPHS is set.  gmon.out
1357  * files usable by gprof(1) are created if FLAG_DO_GPROF is set.
1358  */
1359 static int
1360 pmcstat_analyze_log(void)
1361 {
1362         uint32_t cpu, cpuflags;
1363         uintfptr_t pc;
1364         pid_t pid;
1365         struct pmcstat_image *image;
1366         struct pmcstat_process *pp, *ppnew;
1367         struct pmcstat_pcmap *ppm, *ppmtmp;
1368         struct pmclog_ev ev;
1369         struct pmcstat_pmcrecord *pmcr;
1370         pmcstat_interned_string image_path;
1371
1372         assert(args.pa_flags & FLAG_DO_ANALYSIS);
1373
1374         if (elf_version(EV_CURRENT) == EV_NONE)
1375                 err(EX_UNAVAILABLE, "Elf library initialization failed");
1376
1377         while (pmclog_read(args.pa_logparser, &ev) == 0) {
1378                 assert(ev.pl_state == PMCLOG_OK);
1379
1380                 switch (ev.pl_type) {
1381                 case PMCLOG_TYPE_INITIALIZE:
1382                         if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1383                             PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1384                                 warnx(
1385 "WARNING: Log version 0x%x does not match compiled version 0x%x.",
1386                                     ev.pl_u.pl_i.pl_version, PMC_VERSION_MAJOR);
1387                         break;
1388
1389                 case PMCLOG_TYPE_MAP_IN:
1390                         /*
1391                          * Introduce an address range mapping for a
1392                          * userland process or the kernel (pid == -1).
1393                          *
1394                          * We always allocate a process descriptor so
1395                          * that subsequent samples seen for this
1396                          * address range are mapped to the current
1397                          * object being mapped in.
1398                          */
1399                         pid = ev.pl_u.pl_mi.pl_pid;
1400                         if (pid == -1)
1401                                 pp = pmcstat_kernproc;
1402                         else
1403                                 pp = pmcstat_process_lookup(pid,
1404                                     PMCSTAT_ALLOCATE);
1405
1406                         assert(pp != NULL);
1407
1408                         image_path = pmcstat_string_intern(ev.pl_u.pl_mi.
1409                             pl_pathname);
1410                         image = pmcstat_image_from_path(image_path, pid == -1);
1411                         if (image->pi_type == PMCSTAT_IMAGE_UNKNOWN)
1412                                 pmcstat_image_determine_type(image);
1413                         if (image->pi_type != PMCSTAT_IMAGE_INDETERMINABLE)
1414                                 pmcstat_image_link(pp, image,
1415                                     ev.pl_u.pl_mi.pl_start);
1416                         break;
1417
1418                 case PMCLOG_TYPE_MAP_OUT:
1419                         /*
1420                          * Remove an address map.
1421                          */
1422                         pid = ev.pl_u.pl_mo.pl_pid;
1423                         if (pid == -1)
1424                                 pp = pmcstat_kernproc;
1425                         else
1426                                 pp = pmcstat_process_lookup(pid, 0);
1427
1428                         if (pp == NULL) /* unknown process */
1429                                 break;
1430
1431                         pmcstat_image_unmap(pp, ev.pl_u.pl_mo.pl_start,
1432                             ev.pl_u.pl_mo.pl_end);
1433                         break;
1434
1435                 case PMCLOG_TYPE_PCSAMPLE:
1436                         /*
1437                          * Note: the `PCSAMPLE' log entry is not
1438                          * generated by hpwmc(4) after version 2.
1439                          */
1440
1441                         /*
1442                          * We bring in the gmon file for the image
1443                          * currently associated with the PMC & pid
1444                          * pair and increment the appropriate entry
1445                          * bin inside this.
1446                          */
1447                         pmcstat_stats.ps_samples_total++;
1448                         ps_samples_period++;
1449
1450                         pc = ev.pl_u.pl_s.pl_pc;
1451                         pp = pmcstat_process_lookup(ev.pl_u.pl_s.pl_pid,
1452                             PMCSTAT_ALLOCATE);
1453
1454                         /* Get PMC record. */
1455                         pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_s.pl_pmcid);
1456                         assert(pmcr != NULL);
1457                         pmcr->pr_samples++;
1458
1459                         /*
1460                          * Call the plugins processing
1461                          * TODO: move pmcstat_process_find_map inside plugins
1462                          */
1463
1464                         if (plugins[args.pa_pplugin].pl_process != NULL)
1465                                 plugins[args.pa_pplugin].pl_process(
1466                                     pp, pmcr, 1, &pc,
1467                                     pmcstat_process_find_map(pp, pc) != NULL, 0);
1468                         plugins[args.pa_plugin].pl_process(
1469                             pp, pmcr, 1, &pc,
1470                             pmcstat_process_find_map(pp, pc) != NULL, 0);
1471                         break;
1472
1473                 case PMCLOG_TYPE_CALLCHAIN:
1474                         pmcstat_stats.ps_samples_total++;
1475                         ps_samples_period++;
1476
1477                         cpuflags = ev.pl_u.pl_cc.pl_cpuflags;
1478                         cpu = PMC_CALLCHAIN_CPUFLAGS_TO_CPU(cpuflags);
1479
1480                         /* Filter on the CPU id. */
1481                         if (!CPU_ISSET(cpu, &(args.pa_cpumask))) {
1482                                 pmcstat_stats.ps_samples_skipped++;
1483                                 break;
1484                         }
1485
1486                         pp = pmcstat_process_lookup(ev.pl_u.pl_cc.pl_pid,
1487                             PMCSTAT_ALLOCATE);
1488
1489                         /* Get PMC record. */
1490                         pmcr = pmcstat_lookup_pmcid(ev.pl_u.pl_cc.pl_pmcid);
1491                         assert(pmcr != NULL);
1492                         pmcr->pr_samples++;
1493
1494                         /*
1495                          * Call the plugins processing
1496                          */
1497
1498                         if (plugins[args.pa_pplugin].pl_process != NULL)
1499                                 plugins[args.pa_pplugin].pl_process(
1500                                     pp, pmcr,
1501                                     ev.pl_u.pl_cc.pl_npc,
1502                                     ev.pl_u.pl_cc.pl_pc,
1503                                     PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1504                                     cpu);
1505                         plugins[args.pa_plugin].pl_process(
1506                             pp, pmcr,
1507                             ev.pl_u.pl_cc.pl_npc,
1508                             ev.pl_u.pl_cc.pl_pc,
1509                             PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(cpuflags),
1510                             cpu);
1511                         break;
1512
1513                 case PMCLOG_TYPE_PMCALLOCATE:
1514                         /*
1515                          * Record the association pmc id between this
1516                          * PMC and its name.
1517                          */
1518                         pmcstat_pmcid_add(ev.pl_u.pl_a.pl_pmcid,
1519                             pmcstat_string_intern(ev.pl_u.pl_a.pl_evname));
1520                         break;
1521
1522                 case PMCLOG_TYPE_PMCALLOCATEDYN:
1523                         /*
1524                          * Record the association pmc id between this
1525                          * PMC and its name.
1526                          */
1527                         pmcstat_pmcid_add(ev.pl_u.pl_ad.pl_pmcid,
1528                             pmcstat_string_intern(ev.pl_u.pl_ad.pl_evname));
1529                         break;
1530
1531                 case PMCLOG_TYPE_PROCEXEC:
1532
1533                         /*
1534                          * Change the executable image associated with
1535                          * a process.
1536                          */
1537                         pp = pmcstat_process_lookup(ev.pl_u.pl_x.pl_pid,
1538                             PMCSTAT_ALLOCATE);
1539
1540                         /* delete the current process map */
1541                         TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
1542                                 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
1543                                 free(ppm);
1544                         }
1545
1546                         /*
1547                          * Associate this process image.
1548                          */
1549                         image_path = pmcstat_string_intern(
1550                                 ev.pl_u.pl_x.pl_pathname);
1551                         assert(image_path != NULL);
1552                         pmcstat_process_exec(pp, image_path,
1553                             ev.pl_u.pl_x.pl_entryaddr);
1554                         break;
1555
1556                 case PMCLOG_TYPE_PROCEXIT:
1557
1558                         /*
1559                          * Due to the way the log is generated, the
1560                          * last few samples corresponding to a process
1561                          * may appear in the log after the process
1562                          * exit event is recorded.  Thus we keep the
1563                          * process' descriptor and associated data
1564                          * structures around, but mark the process as
1565                          * having exited.
1566                          */
1567                         pp = pmcstat_process_lookup(ev.pl_u.pl_e.pl_pid, 0);
1568                         if (pp == NULL)
1569                                 break;
1570                         pp->pp_isactive = 0;    /* mark as a zombie */
1571                         break;
1572
1573                 case PMCLOG_TYPE_SYSEXIT:
1574                         pp = pmcstat_process_lookup(ev.pl_u.pl_se.pl_pid, 0);
1575                         if (pp == NULL)
1576                                 break;
1577                         pp->pp_isactive = 0;    /* make a zombie */
1578                         break;
1579
1580                 case PMCLOG_TYPE_PROCFORK:
1581
1582                         /*
1583                          * Allocate a process descriptor for the new
1584                          * (child) process.
1585                          */
1586                         ppnew =
1587                             pmcstat_process_lookup(ev.pl_u.pl_f.pl_newpid,
1588                                 PMCSTAT_ALLOCATE);
1589
1590                         /*
1591                          * If we had been tracking the parent, clone
1592                          * its address maps.
1593                          */
1594                         pp = pmcstat_process_lookup(ev.pl_u.pl_f.pl_oldpid, 0);
1595                         if (pp == NULL)
1596                                 break;
1597                         TAILQ_FOREACH(ppm, &pp->pp_map, ppm_next)
1598                             pmcstat_image_link(ppnew, ppm->ppm_image,
1599                                 ppm->ppm_lowpc);
1600                         break;
1601
1602                 default:        /* other types of entries are not relevant */
1603                         break;
1604                 }
1605         }
1606
1607         if (ev.pl_state == PMCLOG_EOF)
1608                 return (PMCSTAT_FINISHED);
1609         else if (ev.pl_state == PMCLOG_REQUIRE_DATA)
1610                 return (PMCSTAT_RUNNING);
1611
1612         err(EX_DATAERR,
1613             "ERROR: event parsing failed (record %jd, offset 0x%jx)",
1614             (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1615 }
1616
1617 /*
1618  * Print log entries as text.
1619  */
1620
1621 static int
1622 pmcstat_print_log(void)
1623 {
1624         struct pmclog_ev ev;
1625         uint32_t npc;
1626
1627         while (pmclog_read(args.pa_logparser, &ev) == 0) {
1628                 assert(ev.pl_state == PMCLOG_OK);
1629                 switch (ev.pl_type) {
1630                 case PMCLOG_TYPE_CALLCHAIN:
1631                         PMCSTAT_PRINT_ENTRY("callchain",
1632                             "%d 0x%x %d %d %c", ev.pl_u.pl_cc.pl_pid,
1633                             ev.pl_u.pl_cc.pl_pmcid,
1634                             PMC_CALLCHAIN_CPUFLAGS_TO_CPU(ev.pl_u.pl_cc. \
1635                                 pl_cpuflags), ev.pl_u.pl_cc.pl_npc,
1636                             PMC_CALLCHAIN_CPUFLAGS_TO_USERMODE(ev.pl_u.pl_cc.\
1637                                 pl_cpuflags) ? 'u' : 's');
1638                         for (npc = 0; npc < ev.pl_u.pl_cc.pl_npc; npc++)
1639                                 PMCSTAT_PRINT_ENTRY("...", "%p",
1640                                     (void *) ev.pl_u.pl_cc.pl_pc[npc]);
1641                         break;
1642                 case PMCLOG_TYPE_CLOSELOG:
1643                         PMCSTAT_PRINT_ENTRY("closelog",);
1644                         break;
1645                 case PMCLOG_TYPE_DROPNOTIFY:
1646                         PMCSTAT_PRINT_ENTRY("drop",);
1647                         break;
1648                 case PMCLOG_TYPE_INITIALIZE:
1649                         PMCSTAT_PRINT_ENTRY("initlog","0x%x \"%s\"",
1650                             ev.pl_u.pl_i.pl_version,
1651                             pmc_name_of_cputype(ev.pl_u.pl_i.pl_arch));
1652                         if ((ev.pl_u.pl_i.pl_version & 0xFF000000) !=
1653                             PMC_VERSION_MAJOR << 24 && args.pa_verbosity > 0)
1654                                 warnx(
1655 "WARNING: Log version 0x%x != expected version 0x%x.",
1656                                     ev.pl_u.pl_i.pl_version, PMC_VERSION);
1657                         break;
1658                 case PMCLOG_TYPE_MAP_IN:
1659                         PMCSTAT_PRINT_ENTRY("map-in","%d %p \"%s\"",
1660                             ev.pl_u.pl_mi.pl_pid,
1661                             (void *) ev.pl_u.pl_mi.pl_start,
1662                             ev.pl_u.pl_mi.pl_pathname);
1663                         break;
1664                 case PMCLOG_TYPE_MAP_OUT:
1665                         PMCSTAT_PRINT_ENTRY("map-out","%d %p %p",
1666                             ev.pl_u.pl_mo.pl_pid,
1667                             (void *) ev.pl_u.pl_mo.pl_start,
1668                             (void *) ev.pl_u.pl_mo.pl_end);
1669                         break;
1670                 case PMCLOG_TYPE_PCSAMPLE:
1671                         PMCSTAT_PRINT_ENTRY("sample","0x%x %d %p %c",
1672                             ev.pl_u.pl_s.pl_pmcid,
1673                             ev.pl_u.pl_s.pl_pid,
1674                             (void *) ev.pl_u.pl_s.pl_pc,
1675                             ev.pl_u.pl_s.pl_usermode ? 'u' : 's');
1676                         break;
1677                 case PMCLOG_TYPE_PMCALLOCATE:
1678                         PMCSTAT_PRINT_ENTRY("allocate","0x%x \"%s\" 0x%x",
1679                             ev.pl_u.pl_a.pl_pmcid,
1680                             ev.pl_u.pl_a.pl_evname,
1681                             ev.pl_u.pl_a.pl_flags);
1682                         break;
1683                 case PMCLOG_TYPE_PMCALLOCATEDYN:
1684                         PMCSTAT_PRINT_ENTRY("allocatedyn","0x%x \"%s\" 0x%x",
1685                             ev.pl_u.pl_ad.pl_pmcid,
1686                             ev.pl_u.pl_ad.pl_evname,
1687                             ev.pl_u.pl_ad.pl_flags);
1688                         break;
1689                 case PMCLOG_TYPE_PMCATTACH:
1690                         PMCSTAT_PRINT_ENTRY("attach","0x%x %d \"%s\"",
1691                             ev.pl_u.pl_t.pl_pmcid,
1692                             ev.pl_u.pl_t.pl_pid,
1693                             ev.pl_u.pl_t.pl_pathname);
1694                         break;
1695                 case PMCLOG_TYPE_PMCDETACH:
1696                         PMCSTAT_PRINT_ENTRY("detach","0x%x %d",
1697                             ev.pl_u.pl_d.pl_pmcid,
1698                             ev.pl_u.pl_d.pl_pid);
1699                         break;
1700                 case PMCLOG_TYPE_PROCCSW:
1701                         PMCSTAT_PRINT_ENTRY("cswval","0x%x %d %jd",
1702                             ev.pl_u.pl_c.pl_pmcid,
1703                             ev.pl_u.pl_c.pl_pid,
1704                             ev.pl_u.pl_c.pl_value);
1705                         break;
1706                 case PMCLOG_TYPE_PROCEXEC:
1707                         PMCSTAT_PRINT_ENTRY("exec","0x%x %d %p \"%s\"",
1708                             ev.pl_u.pl_x.pl_pmcid,
1709                             ev.pl_u.pl_x.pl_pid,
1710                             (void *) ev.pl_u.pl_x.pl_entryaddr,
1711                             ev.pl_u.pl_x.pl_pathname);
1712                         break;
1713                 case PMCLOG_TYPE_PROCEXIT:
1714                         PMCSTAT_PRINT_ENTRY("exitval","0x%x %d %jd",
1715                             ev.pl_u.pl_e.pl_pmcid,
1716                             ev.pl_u.pl_e.pl_pid,
1717                             ev.pl_u.pl_e.pl_value);
1718                         break;
1719                 case PMCLOG_TYPE_PROCFORK:
1720                         PMCSTAT_PRINT_ENTRY("fork","%d %d",
1721                             ev.pl_u.pl_f.pl_oldpid,
1722                             ev.pl_u.pl_f.pl_newpid);
1723                         break;
1724                 case PMCLOG_TYPE_USERDATA:
1725                         PMCSTAT_PRINT_ENTRY("userdata","0x%x",
1726                             ev.pl_u.pl_u.pl_userdata);
1727                         break;
1728                 case PMCLOG_TYPE_SYSEXIT:
1729                         PMCSTAT_PRINT_ENTRY("exit","%d",
1730                             ev.pl_u.pl_se.pl_pid);
1731                         break;
1732                 default:
1733                         fprintf(args.pa_printfile, "unknown event (type %d).\n",
1734                             ev.pl_type);
1735                 }
1736         }
1737
1738         if (ev.pl_state == PMCLOG_EOF)
1739                 return (PMCSTAT_FINISHED);
1740         else if (ev.pl_state ==  PMCLOG_REQUIRE_DATA)
1741                 return (PMCSTAT_RUNNING);
1742
1743         errx(EX_DATAERR,
1744             "ERROR: event parsing failed (record %jd, offset 0x%jx).",
1745             (uintmax_t) ev.pl_count + 1, ev.pl_offset);
1746         /*NOTREACHED*/
1747 }
1748
1749 /*
1750  * Public Interfaces.
1751  */
1752
1753 /*
1754  * Close a logfile, after first flushing all in-module queued data.
1755  */
1756
1757 int
1758 pmcstat_close_log(void)
1759 {
1760         /* If a local logfile is configured ask the kernel to stop
1761          * and flush data. Kernel will close the file when data is flushed
1762          * so keep the status to EXITING.
1763          */
1764         if (args.pa_logfd != -1) {
1765                 if (pmc_close_logfile() < 0)
1766                         err(EX_OSERR, "ERROR: logging failed");
1767         }
1768
1769         return (args.pa_flags & FLAG_HAS_PIPE ? PMCSTAT_EXITING :
1770             PMCSTAT_FINISHED);
1771 }
1772
1773
1774
1775 /*
1776  * Open a log file, for reading or writing.
1777  *
1778  * The function returns the fd of a successfully opened log or -1 in
1779  * case of failure.
1780  */
1781
1782 int
1783 pmcstat_open_log(const char *path, int mode)
1784 {
1785         int error, fd, cfd;
1786         size_t hlen;
1787         const char *p, *errstr;
1788         struct addrinfo hints, *res, *res0;
1789         char hostname[MAXHOSTNAMELEN];
1790
1791         errstr = NULL;
1792         fd = -1;
1793
1794         /*
1795          * If 'path' is "-" then open one of stdin or stdout depending
1796          * on the value of 'mode'.
1797          *
1798          * If 'path' contains a ':' and does not start with a '/' or '.',
1799          * and is being opened for writing, treat it as a "host:port"
1800          * specification and open a network socket.
1801          *
1802          * Otherwise, treat 'path' as a file name and open that.
1803          */
1804         if (path[0] == '-' && path[1] == '\0')
1805                 fd = (mode == PMCSTAT_OPEN_FOR_READ) ? 0 : 1;
1806         else if (path[0] != '/' &&
1807             path[0] != '.' && strchr(path, ':') != NULL) {
1808
1809                 p = strrchr(path, ':');
1810                 hlen = p - path;
1811                 if (p == path || hlen >= sizeof(hostname)) {
1812                         errstr = strerror(EINVAL);
1813                         goto done;
1814                 }
1815
1816                 assert(hlen < sizeof(hostname));
1817                 (void) strncpy(hostname, path, hlen);
1818                 hostname[hlen] = '\0';
1819
1820                 (void) memset(&hints, 0, sizeof(hints));
1821                 hints.ai_family = AF_UNSPEC;
1822                 hints.ai_socktype = SOCK_STREAM;
1823                 if ((error = getaddrinfo(hostname, p+1, &hints, &res0)) != 0) {
1824                         errstr = gai_strerror(error);
1825                         goto done;
1826                 }
1827
1828                 fd = -1;
1829                 for (res = res0; res; res = res->ai_next) {
1830                         if ((fd = socket(res->ai_family, res->ai_socktype,
1831                             res->ai_protocol)) < 0) {
1832                                 errstr = strerror(errno);
1833                                 continue;
1834                         }
1835                         if (mode == PMCSTAT_OPEN_FOR_READ) {
1836                                 if (bind(fd, res->ai_addr, res->ai_addrlen) < 0) {
1837                                         errstr = strerror(errno);
1838                                         (void) close(fd);
1839                                         fd = -1;
1840                                         continue;
1841                                 }
1842                                 listen(fd, 1);
1843                                 cfd = accept(fd, NULL, NULL);
1844                                 (void) close(fd);
1845                                 if (cfd < 0) {
1846                                         errstr = strerror(errno);
1847                                         fd = -1;
1848                                         break;
1849                                 }
1850                                 fd = cfd;
1851                         } else {
1852                                 if (connect(fd, res->ai_addr, res->ai_addrlen) < 0) {
1853                                         errstr = strerror(errno);
1854                                         (void) close(fd);
1855                                         fd = -1;
1856                                         continue;
1857                                 }
1858                         }
1859                         errstr = NULL;
1860                         break;
1861                 }
1862                 freeaddrinfo(res0);
1863
1864         } else if ((fd = open(path, mode == PMCSTAT_OPEN_FOR_READ ?
1865                     O_RDONLY : (O_WRONLY|O_CREAT|O_TRUNC),
1866                     S_IRUSR|S_IWUSR|S_IRGRP|S_IROTH)) < 0)
1867                         errstr = strerror(errno);
1868
1869   done:
1870         if (errstr)
1871                 errx(EX_OSERR, "ERROR: Cannot open \"%s\" for %s: %s.", path,
1872                     (mode == PMCSTAT_OPEN_FOR_READ ? "reading" : "writing"),
1873                     errstr);
1874
1875         return (fd);
1876 }
1877
1878 /*
1879  * Process a log file in offline analysis mode.
1880  */
1881
1882 int
1883 pmcstat_process_log(void)
1884 {
1885
1886         /*
1887          * If analysis has not been asked for, just print the log to
1888          * the current output file.
1889          */
1890         if (args.pa_flags & FLAG_DO_PRINT)
1891                 return (pmcstat_print_log());
1892         else
1893                 return (pmcstat_analyze_log());
1894 }
1895
1896 /*
1897  * Refresh top display.
1898  */
1899
1900 static void
1901 pmcstat_refresh_top(void)
1902 {
1903         int v_attrs;
1904         float v;
1905         char pmcname[40];
1906         struct pmcstat_pmcrecord *pmcpr;
1907
1908         /* If in pause mode do not refresh display. */
1909         if (pmcstat_pause)
1910                 return;
1911
1912         /* Wait until PMC pop in the log. */
1913         pmcpr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1914         if (pmcpr == NULL)
1915                 return;
1916
1917         /* Format PMC name. */
1918         if (pmcstat_mergepmc)
1919                 snprintf(pmcname, sizeof(pmcname), "[%s]",
1920                     pmcstat_string_unintern(pmcpr->pr_pmcname));
1921         else
1922                 snprintf(pmcname, sizeof(pmcname), "%s.%d",
1923                     pmcstat_string_unintern(pmcpr->pr_pmcname),
1924                     pmcstat_pmcinfilter);
1925
1926         /* Format samples count. */
1927         if (ps_samples_period > 0)
1928                 v = (pmcpr->pr_samples * 100.0) / ps_samples_period;
1929         else
1930                 v = 0.;
1931         v_attrs = PMCSTAT_ATTRPERCENT(v);
1932
1933         PMCSTAT_PRINTBEGIN();
1934         PMCSTAT_PRINTW("PMC: %s Samples: %u ",
1935             pmcname,
1936             pmcpr->pr_samples);
1937         PMCSTAT_ATTRON(v_attrs);
1938         PMCSTAT_PRINTW("(%.1f%%) ", v);
1939         PMCSTAT_ATTROFF(v_attrs);
1940         PMCSTAT_PRINTW(", %u unresolved\n\n",
1941             pmcpr->pr_dubious_frames);
1942         if (plugins[args.pa_plugin].pl_topdisplay != NULL)
1943                 plugins[args.pa_plugin].pl_topdisplay();
1944         PMCSTAT_PRINTEND();
1945 }
1946
1947 /*
1948  * Find the next pmc index to display.
1949  */
1950
1951 static void
1952 pmcstat_changefilter(void)
1953 {
1954         int pmcin;
1955         struct pmcstat_pmcrecord *pmcr;
1956
1957         /*
1958          * Find the next merge target.
1959          */
1960         if (pmcstat_mergepmc) {
1961                 pmcin = pmcstat_pmcinfilter;
1962
1963                 do {
1964                         pmcr = pmcstat_pmcindex_to_pmcr(pmcstat_pmcinfilter);
1965                         if (pmcr == NULL || pmcr == pmcr->pr_merge)
1966                                 break;
1967
1968                         pmcstat_pmcinfilter++;
1969                         if (pmcstat_pmcinfilter >= pmcstat_npmcs)
1970                                 pmcstat_pmcinfilter = 0;
1971
1972                 } while (pmcstat_pmcinfilter != pmcin);
1973         }
1974 }
1975
1976 /*
1977  * Top mode keypress.
1978  */
1979
1980 int
1981 pmcstat_keypress_log(void)
1982 {
1983         int c, ret = 0;
1984         WINDOW *w;
1985
1986         w = newwin(1, 0, 1, 0);
1987         c = wgetch(w);
1988         wprintw(w, "Key: %c => ", c);
1989         switch (c) {
1990         case 'c':
1991                 wprintw(w, "enter mode 'd' or 'a' => ");
1992                 c = wgetch(w);
1993                 if (c == 'd') {
1994                         args.pa_topmode = PMCSTAT_TOP_DELTA;
1995                         wprintw(w, "switching to delta mode");
1996                 } else {
1997                         args.pa_topmode = PMCSTAT_TOP_ACCUM;
1998                         wprintw(w, "switching to accumulation mode");
1999                 }
2000                 break;
2001         case 'm':
2002                 pmcstat_mergepmc = !pmcstat_mergepmc;
2003                 /*
2004                  * Changing merge state require data reset.
2005                  */
2006                 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2007                         plugins[args.pa_plugin].pl_shutdown(NULL);
2008                 pmcstat_stats_reset(0);
2009                 if (plugins[args.pa_plugin].pl_init != NULL)
2010                         plugins[args.pa_plugin].pl_init();
2011
2012                 /* Update filter to be on a merge target. */
2013                 pmcstat_changefilter();
2014                 wprintw(w, "merge PMC %s", pmcstat_mergepmc ? "on" : "off");
2015                 break;
2016         case 'n':
2017                 /* Close current plugin. */
2018                 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2019                         plugins[args.pa_plugin].pl_shutdown(NULL);
2020
2021                 /* Find next top display available. */
2022                 do {
2023                         args.pa_plugin++;
2024                         if (plugins[args.pa_plugin].pl_name == NULL)
2025                                 args.pa_plugin = 0;
2026                 } while (plugins[args.pa_plugin].pl_topdisplay == NULL);
2027
2028                 /* Open new plugin. */
2029                 pmcstat_stats_reset(0);
2030                 if (plugins[args.pa_plugin].pl_init != NULL)
2031                         plugins[args.pa_plugin].pl_init();
2032                 wprintw(w, "switching to plugin %s",
2033                     plugins[args.pa_plugin].pl_name);
2034                 break;
2035         case 'p':
2036                 pmcstat_pmcinfilter++;
2037                 if (pmcstat_pmcinfilter >= pmcstat_npmcs)
2038                         pmcstat_pmcinfilter = 0;
2039                 pmcstat_changefilter();
2040                 wprintw(w, "switching to PMC %s.%d",
2041                     pmcstat_pmcindex_to_name(pmcstat_pmcinfilter),
2042                     pmcstat_pmcinfilter);
2043                 break;
2044         case ' ':
2045                 pmcstat_pause = !pmcstat_pause;
2046                 if (pmcstat_pause)
2047                         wprintw(w, "pause => press space again to continue");
2048                 break;
2049         case 'q':
2050                 wprintw(w, "exiting...");
2051                 ret = 1;
2052                 break;
2053         default:
2054                 if (plugins[args.pa_plugin].pl_topkeypress != NULL)
2055                         if (plugins[args.pa_plugin].pl_topkeypress(c, w))
2056                                 ret = 1;
2057         }
2058
2059         wrefresh(w);
2060         delwin(w);
2061         return ret;
2062 }
2063
2064
2065 /*
2066  * Top mode display.
2067  */
2068
2069 void
2070 pmcstat_display_log(void)
2071 {
2072
2073         pmcstat_refresh_top();
2074
2075         /* Reset everythings if delta mode. */
2076         if (args.pa_topmode == PMCSTAT_TOP_DELTA) {
2077                 if (plugins[args.pa_plugin].pl_shutdown != NULL)
2078                         plugins[args.pa_plugin].pl_shutdown(NULL);
2079                 pmcstat_stats_reset(0);
2080                 if (plugins[args.pa_plugin].pl_init != NULL)
2081                         plugins[args.pa_plugin].pl_init();
2082         }
2083
2084 }
2085
2086 /*
2087  * Configure a plugins.
2088  */
2089
2090 void
2091 pmcstat_pluginconfigure_log(char *opt)
2092 {
2093
2094         if (strncmp(opt, "threshold=", 10) == 0) {
2095                 pmcstat_threshold = atof(opt+10);
2096         } else {
2097                 if (plugins[args.pa_plugin].pl_configure != NULL) {
2098                         if (!plugins[args.pa_plugin].pl_configure(opt))
2099                                 err(EX_USAGE,
2100                                     "ERROR: unknown option <%s>.", opt);
2101                 }
2102         }
2103 }
2104
2105 /*
2106  * Initialize module.
2107  */
2108
2109 void
2110 pmcstat_initialize_logging(void)
2111 {
2112         int i;
2113
2114         /* use a convenient format for 'ldd' output */
2115         if (setenv("LD_TRACE_LOADED_OBJECTS_FMT1","%o \"%p\" %x\n",1) != 0)
2116                 err(EX_OSERR, "ERROR: Cannot setenv");
2117
2118         /* Initialize hash tables */
2119         pmcstat_string_initialize();
2120         for (i = 0; i < PMCSTAT_NHASH; i++) {
2121                 LIST_INIT(&pmcstat_image_hash[i]);
2122                 LIST_INIT(&pmcstat_process_hash[i]);
2123         }
2124
2125         /*
2126          * Create a fake 'process' entry for the kernel with pid -1.
2127          * hwpmc(4) will subsequently inform us about where the kernel
2128          * and any loaded kernel modules are mapped.
2129          */
2130         if ((pmcstat_kernproc = pmcstat_process_lookup((pid_t) -1,
2131                  PMCSTAT_ALLOCATE)) == NULL)
2132                 err(EX_OSERR, "ERROR: Cannot initialize logging");
2133
2134         /* PMC count. */
2135         pmcstat_npmcs = 0;
2136
2137         /* Merge PMC with same name. */
2138         pmcstat_mergepmc = args.pa_mergepmc;
2139
2140         /*
2141          * Initialize plugins
2142          */
2143
2144         if (plugins[args.pa_pplugin].pl_init != NULL)
2145                 plugins[args.pa_pplugin].pl_init();
2146         if (plugins[args.pa_plugin].pl_init != NULL)
2147                 plugins[args.pa_plugin].pl_init();
2148 }
2149
2150 /*
2151  * Shutdown module.
2152  */
2153
2154 void
2155 pmcstat_shutdown_logging(void)
2156 {
2157         int i;
2158         FILE *mf;
2159         struct pmcstat_image *pi, *pitmp;
2160         struct pmcstat_process *pp, *pptmp;
2161         struct pmcstat_pcmap *ppm, *ppmtmp;
2162
2163         /* determine where to send the map file */
2164         mf = NULL;
2165         if (args.pa_mapfilename != NULL)
2166                 mf = (strcmp(args.pa_mapfilename, "-") == 0) ?
2167                     args.pa_printfile : fopen(args.pa_mapfilename, "w");
2168
2169         if (mf == NULL && args.pa_flags & FLAG_DO_GPROF &&
2170             args.pa_verbosity >= 2)
2171                 mf = args.pa_printfile;
2172
2173         if (mf)
2174                 (void) fprintf(mf, "MAP:\n");
2175
2176         /*
2177          * Shutdown the plugins
2178          */
2179
2180         if (plugins[args.pa_plugin].pl_shutdown != NULL)
2181                 plugins[args.pa_plugin].pl_shutdown(mf);
2182         if (plugins[args.pa_pplugin].pl_shutdown != NULL)
2183                 plugins[args.pa_pplugin].pl_shutdown(mf);
2184
2185         for (i = 0; i < PMCSTAT_NHASH; i++) {
2186                 LIST_FOREACH_SAFE(pi, &pmcstat_image_hash[i], pi_next,
2187                     pitmp) {
2188                         if (plugins[args.pa_plugin].pl_shutdownimage != NULL)
2189                                 plugins[args.pa_plugin].pl_shutdownimage(pi);
2190                         if (plugins[args.pa_pplugin].pl_shutdownimage != NULL)
2191                                 plugins[args.pa_pplugin].pl_shutdownimage(pi);
2192
2193                         free(pi->pi_symbols);
2194                         if (pi->pi_addr2line != NULL)
2195                                 pclose(pi->pi_addr2line);
2196                         LIST_REMOVE(pi, pi_next);
2197                         free(pi);
2198                 }
2199
2200                 LIST_FOREACH_SAFE(pp, &pmcstat_process_hash[i], pp_next,
2201                     pptmp) {
2202                         TAILQ_FOREACH_SAFE(ppm, &pp->pp_map, ppm_next, ppmtmp) {
2203                                 TAILQ_REMOVE(&pp->pp_map, ppm, ppm_next);
2204                                 free(ppm);
2205                         }
2206                         LIST_REMOVE(pp, pp_next);
2207                         free(pp);
2208                 }
2209         }
2210
2211         pmcstat_string_shutdown();
2212
2213         /*
2214          * Print errors unless -q was specified.  Print all statistics
2215          * if verbosity > 1.
2216          */
2217 #define PRINT(N,V) do {                                                 \
2218                 if (pmcstat_stats.ps_##V || args.pa_verbosity >= 2)     \
2219                         (void) fprintf(args.pa_printfile, " %-40s %d\n",\
2220                             N, pmcstat_stats.ps_##V);                   \
2221         } while (0)
2222
2223         if (args.pa_verbosity >= 1 && (args.pa_flags & FLAG_DO_ANALYSIS)) {
2224                 (void) fprintf(args.pa_printfile, "CONVERSION STATISTICS:\n");
2225                 PRINT("#exec/a.out", exec_aout);
2226                 PRINT("#exec/elf", exec_elf);
2227                 PRINT("#exec/unknown", exec_indeterminable);
2228                 PRINT("#exec handling errors", exec_errors);
2229                 PRINT("#samples/total", samples_total);
2230                 PRINT("#samples/unclaimed", samples_unknown_offset);
2231                 PRINT("#samples/unknown-object", samples_indeterminable);
2232                 PRINT("#samples/unknown-function", samples_unknown_function);
2233                 PRINT("#callchain/dubious-frames", callchain_dubious_frames);
2234         }
2235
2236         if (mf)
2237                 (void) fclose(mf);
2238 }