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