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