granicus.if.org Git - postgresql/blob - src/bin/pg_resetxlog/pg_resetxlog.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * pg_resetxlog.c
   4  *        A utility to "zero out" the xlog when it's corrupt beyond recovery.
   5  *        Can also rebuild pg_control if needed.
   6  *
   7  * The theory of operation is fairly simple:
   8  *        1. Read the existing pg_control (which will include the last
   9  *               checkpoint record).  If it is an old format then update to
  10  *               current format.
  11  *        2. If pg_control is corrupt, attempt to intuit reasonable values,
  12  *               by scanning the old xlog if necessary.
  13  *        3. Modify pg_control to reflect a "shutdown" state with a checkpoint
  14  *               record at the start of xlog.
  15  *        4. Flush the existing xlog files and write a new segment with
  16  *               just a checkpoint record in it.  The new segment is positioned
  17  *               just past the end of the old xlog, so that existing LSNs in
  18  *               data pages will appear to be "in the past".
  19  * This is all pretty straightforward except for the intuition part of
  20  * step 2 ...
  21  *
  22  *
  23  * Portions Copyright (c) 1996-2013, PostgreSQL Global Development Group
  24  * Portions Copyright (c) 1994, Regents of the University of California
  25  *
  26  * src/bin/pg_resetxlog/pg_resetxlog.c
  27  *
  28  *-------------------------------------------------------------------------
  29  */
  30
  31 /*
  32  * We have to use postgres.h not postgres_fe.h here, because there's so much
  33  * backend-only stuff in the XLOG include files we need.  But we need a
  34  * frontend-ish environment otherwise.  Hence this ugly hack.
  35  */
  36 #define FRONTEND 1
  37
  38 #include "postgres.h"
  39
  40 #include <dirent.h>
  41 #include <fcntl.h>
  42 #include <locale.h>
  43 #include <sys/stat.h>
  44 #include <sys/time.h>
  45 #include <time.h>
  46 #include <unistd.h>
  47 #ifdef HAVE_GETOPT_H
  48 #include <getopt.h>
  49 #endif
  50
  51 #include "access/transam.h"
  52 #include "access/tuptoaster.h"
  53 #include "access/multixact.h"
  54 #include "access/xlog_internal.h"
  55 #include "catalog/catversion.h"
  56 #include "catalog/pg_control.h"
  57 #include "common/fe_memutils.h"
  58
  59 extern int      optind;
  60 extern char *optarg;
  61
  62
  63 static ControlFileData ControlFile;             /* pg_control values */
  64 static XLogSegNo newXlogSegNo;  /* new XLOG segment # */
  65 static bool guessed = false;    /* T if we had to guess at any values */
  66 static const char *progname;
  67
  68 static bool ReadControlFile(void);
  69 static void GuessControlValues(void);
  70 static void PrintControlValues(bool guessed);
  71 static void RewriteControlFile(void);
  72 static void FindEndOfXLOG(void);
  73 static void KillExistingXLOG(void);
  74 static void KillExistingArchiveStatus(void);
  75 static void WriteEmptyXLOG(void);
  76 static void usage(void);
  77
  78
  79 int
  80 main(int argc, char *argv[])
  81 {
  82         int                     c;
  83         bool            force = false;
  84         bool            noupdate = false;
  85         uint32          set_xid_epoch = (uint32) -1;
  86         TransactionId set_xid = 0;
  87         Oid                     set_oid = 0;
  88         MultiXactId set_mxid = 0;
  89         MultiXactId set_oldestmxid = 0;
  90         MultiXactOffset set_mxoff = (MultiXactOffset) -1;
  91         uint32          minXlogTli = 0;
  92         XLogSegNo       minXlogSegNo = 0;
  93         char       *endptr;
  94         char       *endptr2;
  95         char       *DataDir;
  96         int                     fd;
  97
  98         set_pglocale_pgservice(argv[0], PG_TEXTDOMAIN("pg_resetxlog"));
  99
 100         progname = get_progname(argv[0]);
 101
 102         if (argc > 1)
 103         {
 104                 if (strcmp(argv[1], "--help") == 0 || strcmp(argv[1], "-?") == 0)
 105                 {
 106                         usage();
 107                         exit(0);
 108                 }
 109                 if (strcmp(argv[1], "--version") == 0 || strcmp(argv[1], "-V") == 0)
 110                 {
 111                         puts("pg_resetxlog (PostgreSQL) " PG_VERSION);
 112                         exit(0);
 113                 }
 114         }
 115
 116
 117         while ((c = getopt(argc, argv, "fl:m:no:O:x:e:")) != -1)
 118         {
 119                 switch (c)
 120                 {
 121                         case 'f':
 122                                 force = true;
 123                                 break;
 124
 125                         case 'n':
 126                                 noupdate = true;
 127                                 break;
 128
 129                         case 'e':
 130                                 set_xid_epoch = strtoul(optarg, &endptr, 0);
 131                                 if (endptr == optarg || *endptr != '\0')
 132                                 {
 133                                         fprintf(stderr, _("%s: invalid argument for option -e\n"), progname);
 134                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 135                                         exit(1);
 136                                 }
 137                                 if (set_xid_epoch == -1)
 138                                 {
 139                                         fprintf(stderr, _("%s: transaction ID epoch (-e) must not be -1\n"), progname);
 140                                         exit(1);
 141                                 }
 142                                 break;
 143
 144                         case 'x':
 145                                 set_xid = strtoul(optarg, &endptr, 0);
 146                                 if (endptr == optarg || *endptr != '\0')
 147                                 {
 148                                         fprintf(stderr, _("%s: invalid argument for option -x\n"), progname);
 149                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 150                                         exit(1);
 151                                 }
 152                                 if (set_xid == 0)
 153                                 {
 154                                         fprintf(stderr, _("%s: transaction ID (-x) must not be 0\n"), progname);
 155                                         exit(1);
 156                                 }
 157                                 break;
 158
 159                         case 'o':
 160                                 set_oid = strtoul(optarg, &endptr, 0);
 161                                 if (endptr == optarg || *endptr != '\0')
 162                                 {
 163                                         fprintf(stderr, _("%s: invalid argument for option -o\n"), progname);
 164                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 165                                         exit(1);
 166                                 }
 167                                 if (set_oid == 0)
 168                                 {
 169                                         fprintf(stderr, _("%s: OID (-o) must not be 0\n"), progname);
 170                                         exit(1);
 171                                 }
 172                                 break;
 173
 174                         case 'm':
 175                                 set_mxid = strtoul(optarg, &endptr, 0);
 176                                 if (endptr == optarg || *endptr != ',')
 177                                 {
 178                                         fprintf(stderr, _("%s: invalid argument for option -m\n"), progname);
 179                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 180                                         exit(1);
 181                                 }
 182
 183                                 set_oldestmxid = strtoul(endptr + 1, &endptr2, 0);
 184                                 if (endptr2 == endptr + 1 || *endptr2 != '\0')
 185                                 {
 186                                         fprintf(stderr, _("%s: invalid argument for option -m\n"), progname);
 187                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 188                                         exit(1);
 189                                 }
 190                                 if (set_mxid == 0)
 191                                 {
 192                                         fprintf(stderr, _("%s: multitransaction ID (-m) must not be 0\n"), progname);
 193                                         exit(1);
 194                                 }
 195
 196                                 /*
 197                                  * XXX It'd be nice to have more sanity checks here, e.g. so
 198                                  * that oldest is not wrapped around w.r.t. nextMulti.
 199                                  */
 200                                 if (set_oldestmxid == 0)
 201                                 {
 202                                         fprintf(stderr, _("%s: oldest multitransaction ID (-m) must not be 0\n"),
 203                                                         progname);
 204                                         exit(1);
 205                                 }
 206                                 break;
 207
 208                         case 'O':
 209                                 set_mxoff = strtoul(optarg, &endptr, 0);
 210                                 if (endptr == optarg || *endptr != '\0')
 211                                 {
 212                                         fprintf(stderr, _("%s: invalid argument for option -O\n"), progname);
 213                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 214                                         exit(1);
 215                                 }
 216                                 if (set_mxoff == -1)
 217                                 {
 218                                         fprintf(stderr, _("%s: multitransaction offset (-O) must not be -1\n"), progname);
 219                                         exit(1);
 220                                 }
 221                                 break;
 222
 223                         case 'l':
 224                                 if (strspn(optarg, "01234567890ABCDEFabcdef") != 24)
 225                                 {
 226                                         fprintf(stderr, _("%s: invalid argument for option -l\n"), progname);
 227                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 228                                         exit(1);
 229                                 }
 230                                 XLogFromFileName(optarg, &minXlogTli, &minXlogSegNo);
 231                                 break;
 232
 233                         default:
 234                                 fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 235                                 exit(1);
 236                 }
 237         }
 238
 239         if (optind == argc)
 240         {
 241                 fprintf(stderr, _("%s: no data directory specified\n"), progname);
 242                 fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 243                 exit(1);
 244         }
 245
 246         /*
 247          * Don't allow pg_resetxlog to be run as root, to avoid overwriting the
 248          * ownership of files in the data directory. We need only check for root
 249          * -- any other user won't have sufficient permissions to modify files in
 250          * the data directory.
 251          */
 252 #ifndef WIN32
 253         if (geteuid() == 0)
 254         {
 255                 fprintf(stderr, _("%s: cannot be executed by \"root\"\n"),
 256                                 progname);
 257                 fprintf(stderr, _("You must run %s as the PostgreSQL superuser.\n"),
 258                                 progname);
 259                 exit(1);
 260         }
 261 #endif
 262
 263         DataDir = argv[optind];
 264
 265         if (chdir(DataDir) < 0)
 266         {
 267                 fprintf(stderr, _("%s: could not change directory to \"%s\": %s\n"),
 268                                 progname, DataDir, strerror(errno));
 269                 exit(1);
 270         }
 271
 272         /*
 273          * Check for a postmaster lock file --- if there is one, refuse to
 274          * proceed, on grounds we might be interfering with a live installation.
 275          */
 276         if ((fd = open("postmaster.pid", O_RDONLY, 0)) < 0)
 277         {
 278                 if (errno != ENOENT)
 279                 {
 280                         fprintf(stderr, _("%s: could not open file \"%s\" for reading: %s\n"),
 281                                         progname, "postmaster.pid", strerror(errno));
 282                         exit(1);
 283                 }
 284         }
 285         else
 286         {
 287                 fprintf(stderr, _("%s: lock file \"%s\" exists\n"
 288                                                   "Is a server running?  If not, delete the lock file and try again.\n"),
 289                                 progname, "postmaster.pid");
 290                 exit(1);
 291         }
 292
 293         /*
 294          * Attempt to read the existing pg_control file
 295          */
 296         if (!ReadControlFile())
 297                 GuessControlValues();
 298
 299         /*
 300          * Also look at existing segment files to set up newXlogSegNo
 301          */
 302         FindEndOfXLOG();
 303
 304         /*
 305          * Adjust fields if required by switches.  (Do this now so that printout,
 306          * if any, includes these values.)
 307          */
 308         if (set_xid_epoch != -1)
 309                 ControlFile.checkPointCopy.nextXidEpoch = set_xid_epoch;
 310
 311         if (set_xid != 0)
 312         {
 313                 ControlFile.checkPointCopy.nextXid = set_xid;
 314
 315                 /*
 316                  * For the moment, just set oldestXid to a value that will force
 317                  * immediate autovacuum-for-wraparound.  It's not clear whether adding
 318                  * user control of this is useful, so let's just do something that's
 319                  * reasonably safe.  The magic constant here corresponds to the
 320                  * maximum allowed value of autovacuum_freeze_max_age.
 321                  */
 322                 ControlFile.checkPointCopy.oldestXid = set_xid - 2000000000;
 323                 if (ControlFile.checkPointCopy.oldestXid < FirstNormalTransactionId)
 324                         ControlFile.checkPointCopy.oldestXid += FirstNormalTransactionId;
 325                 ControlFile.checkPointCopy.oldestXidDB = InvalidOid;
 326         }
 327
 328         if (set_oid != 0)
 329                 ControlFile.checkPointCopy.nextOid = set_oid;
 330
 331         if (set_mxid != 0)
 332         {
 333                 ControlFile.checkPointCopy.nextMulti = set_mxid;
 334
 335                 ControlFile.checkPointCopy.oldestMulti = set_oldestmxid;
 336                 if (ControlFile.checkPointCopy.oldestMulti < FirstMultiXactId)
 337                         ControlFile.checkPointCopy.oldestMulti += FirstMultiXactId;
 338                 ControlFile.checkPointCopy.oldestMultiDB = InvalidOid;
 339         }
 340
 341         if (set_mxoff != -1)
 342                 ControlFile.checkPointCopy.nextMultiOffset = set_mxoff;
 343
 344         if (minXlogTli > ControlFile.checkPointCopy.ThisTimeLineID)
 345         {
 346                 ControlFile.checkPointCopy.ThisTimeLineID = minXlogTli;
 347                 ControlFile.checkPointCopy.PrevTimeLineID = minXlogTli;
 348         }
 349
 350         if (minXlogSegNo > newXlogSegNo)
 351                 newXlogSegNo = minXlogSegNo;
 352
 353         /*
 354          * If we had to guess anything, and -f was not given, just print the
 355          * guessed values and exit.  Also print if -n is given.
 356          */
 357         if ((guessed && !force) || noupdate)
 358         {
 359                 PrintControlValues(guessed);
 360                 if (!noupdate)
 361                 {
 362                         printf(_("\nIf these values seem acceptable, use -f to force reset.\n"));
 363                         exit(1);
 364                 }
 365                 else
 366                         exit(0);
 367         }
 368
 369         /*
 370          * Don't reset from a dirty pg_control without -f, either.
 371          */
 372         if (ControlFile.state != DB_SHUTDOWNED && !force)
 373         {
 374                 printf(_("The database server was not shut down cleanly.\n"
 375                            "Resetting the transaction log might cause data to be lost.\n"
 376                                  "If you want to proceed anyway, use -f to force reset.\n"));
 377                 exit(1);
 378         }
 379
 380         /*
 381          * Else, do the dirty deed.
 382          */
 383         RewriteControlFile();
 384         KillExistingXLOG();
 385         KillExistingArchiveStatus();
 386         WriteEmptyXLOG();
 387
 388         printf(_("Transaction log reset\n"));
 389         return 0;
 390 }
 391
 392
 393 /*
 394  * Try to read the existing pg_control file.
 395  *
 396  * This routine is also responsible for updating old pg_control versions
 397  * to the current format.  (Currently we don't do anything of the sort.)
 398  */
 399 static bool
 400 ReadControlFile(void)
 401 {
 402         int                     fd;
 403         int                     len;
 404         char       *buffer;
 405         pg_crc32        crc;
 406
 407         if ((fd = open(XLOG_CONTROL_FILE, O_RDONLY | PG_BINARY, 0)) < 0)
 408         {
 409                 /*
 410                  * If pg_control is not there at all, or we can't read it, the odds
 411                  * are we've been handed a bad DataDir path, so give up. User can do
 412                  * "touch pg_control" to force us to proceed.
 413                  */
 414                 fprintf(stderr, _("%s: could not open file \"%s\" for reading: %s\n"),
 415                                 progname, XLOG_CONTROL_FILE, strerror(errno));
 416                 if (errno == ENOENT)
 417                         fprintf(stderr, _("If you are sure the data directory path is correct, execute\n"
 418                                                           "  touch %s\n"
 419                                                           "and try again.\n"),
 420                                         XLOG_CONTROL_FILE);
 421                 exit(1);
 422         }
 423
 424         /* Use malloc to ensure we have a maxaligned buffer */
 425         buffer = (char *) pg_malloc(PG_CONTROL_SIZE);
 426
 427         len = read(fd, buffer, PG_CONTROL_SIZE);
 428         if (len < 0)
 429         {
 430                 fprintf(stderr, _("%s: could not read file \"%s\": %s\n"),
 431                                 progname, XLOG_CONTROL_FILE, strerror(errno));
 432                 exit(1);
 433         }
 434         close(fd);
 435
 436         if (len >= sizeof(ControlFileData) &&
 437           ((ControlFileData *) buffer)->pg_control_version == PG_CONTROL_VERSION)
 438         {
 439                 /* Check the CRC. */
 440                 INIT_CRC32(crc);
 441                 COMP_CRC32(crc,
 442                                    buffer,
 443                                    offsetof(ControlFileData, crc));
 444                 FIN_CRC32(crc);
 445
 446                 if (EQ_CRC32(crc, ((ControlFileData *) buffer)->crc))
 447                 {
 448                         /* Valid data... */
 449                         memcpy(&ControlFile, buffer, sizeof(ControlFile));
 450                         return true;
 451                 }
 452
 453                 fprintf(stderr, _("%s: pg_control exists but has invalid CRC; proceed with caution\n"),
 454                                 progname);
 455                 /* We will use the data anyway, but treat it as guessed. */
 456                 memcpy(&ControlFile, buffer, sizeof(ControlFile));
 457                 guessed = true;
 458                 return true;
 459         }
 460
 461         /* Looks like it's a mess. */
 462         fprintf(stderr, _("%s: pg_control exists but is broken or unknown version; ignoring it\n"),
 463                         progname);
 464         return false;
 465 }
 466
 467
 468 /*
 469  * Guess at pg_control values when we can't read the old ones.
 470  */
 471 static void
 472 GuessControlValues(void)
 473 {
 474         uint64          sysidentifier;
 475         struct timeval tv;
 476
 477         /*
 478          * Set up a completely default set of pg_control values.
 479          */
 480         guessed = true;
 481         memset(&ControlFile, 0, sizeof(ControlFile));
 482
 483         ControlFile.pg_control_version = PG_CONTROL_VERSION;
 484         ControlFile.catalog_version_no = CATALOG_VERSION_NO;
 485
 486         /*
 487          * Create a new unique installation identifier, since we can no longer use
 488          * any old XLOG records.  See notes in xlog.c about the algorithm.
 489          */
 490         gettimeofday(&tv, NULL);
 491         sysidentifier = ((uint64) tv.tv_sec) << 32;
 492         sysidentifier |= (uint32) (tv.tv_sec | tv.tv_usec);
 493
 494         ControlFile.system_identifier = sysidentifier;
 495
 496         ControlFile.checkPointCopy.redo = SizeOfXLogLongPHD;
 497         ControlFile.checkPointCopy.ThisTimeLineID = 1;
 498         ControlFile.checkPointCopy.PrevTimeLineID = 1;
 499         ControlFile.checkPointCopy.fullPageWrites = false;
 500         ControlFile.checkPointCopy.nextXidEpoch = 0;
 501         ControlFile.checkPointCopy.nextXid = FirstNormalTransactionId;
 502         ControlFile.checkPointCopy.nextOid = FirstBootstrapObjectId;
 503         ControlFile.checkPointCopy.nextMulti = FirstMultiXactId;
 504         ControlFile.checkPointCopy.nextMultiOffset = 0;
 505         ControlFile.checkPointCopy.oldestXid = FirstNormalTransactionId;
 506         ControlFile.checkPointCopy.oldestXidDB = InvalidOid;
 507         ControlFile.checkPointCopy.oldestMulti = FirstMultiXactId;
 508         ControlFile.checkPointCopy.oldestMultiDB = InvalidOid;
 509         ControlFile.checkPointCopy.time = (pg_time_t) time(NULL);
 510         ControlFile.checkPointCopy.oldestActiveXid = InvalidTransactionId;
 511
 512         ControlFile.state = DB_SHUTDOWNED;
 513         ControlFile.time = (pg_time_t) time(NULL);
 514         ControlFile.checkPoint = ControlFile.checkPointCopy.redo;
 515         ControlFile.unloggedLSN = 1;
 516
 517         /* minRecoveryPoint, backupStartPoint and backupEndPoint can be left zero */
 518
 519         ControlFile.wal_level = WAL_LEVEL_MINIMAL;
 520         ControlFile.MaxConnections = 100;
 521         ControlFile.max_prepared_xacts = 0;
 522         ControlFile.max_locks_per_xact = 64;
 523
 524         ControlFile.maxAlign = MAXIMUM_ALIGNOF;
 525         ControlFile.floatFormat = FLOATFORMAT_VALUE;
 526         ControlFile.blcksz = BLCKSZ;
 527         ControlFile.relseg_size = RELSEG_SIZE;
 528         ControlFile.xlog_blcksz = XLOG_BLCKSZ;
 529         ControlFile.xlog_seg_size = XLOG_SEG_SIZE;
 530         ControlFile.nameDataLen = NAMEDATALEN;
 531         ControlFile.indexMaxKeys = INDEX_MAX_KEYS;
 532         ControlFile.toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
 533 #ifdef HAVE_INT64_TIMESTAMP
 534         ControlFile.enableIntTimes = true;
 535 #else
 536         ControlFile.enableIntTimes = false;
 537 #endif
 538         ControlFile.float4ByVal = FLOAT4PASSBYVAL;
 539         ControlFile.float8ByVal = FLOAT8PASSBYVAL;
 540
 541         /*
 542          * XXX eventually, should try to grovel through old XLOG to develop more
 543          * accurate values for TimeLineID, nextXID, etc.
 544          */
 545 }
 546
 547
 548 /*
 549  * Print the guessed pg_control values when we had to guess.
 550  *
 551  * NB: this display should be just those fields that will not be
 552  * reset by RewriteControlFile().
 553  */
 554 static void
 555 PrintControlValues(bool guessed)
 556 {
 557         char            sysident_str[32];
 558         char            fname[MAXFNAMELEN];
 559
 560         if (guessed)
 561                 printf(_("Guessed pg_control values:\n\n"));
 562         else
 563                 printf(_("pg_control values:\n\n"));
 564
 565         /*
 566          * Format system_identifier separately to keep platform-dependent format
 567          * code out of the translatable message string.
 568          */
 569         snprintf(sysident_str, sizeof(sysident_str), UINT64_FORMAT,
 570                          ControlFile.system_identifier);
 571
 572         XLogFileName(fname, ControlFile.checkPointCopy.ThisTimeLineID, newXlogSegNo);
 573
 574         printf(_("First log segment after reset:        %s\n"),
 575                    fname);
 576         printf(_("pg_control version number:            %u\n"),
 577                    ControlFile.pg_control_version);
 578         printf(_("Catalog version number:               %u\n"),
 579                    ControlFile.catalog_version_no);
 580         printf(_("Database system identifier:           %s\n"),
 581                    sysident_str);
 582         printf(_("Latest checkpoint's TimeLineID:       %u\n"),
 583                    ControlFile.checkPointCopy.ThisTimeLineID);
 584         printf(_("Latest checkpoint's full_page_writes: %s\n"),
 585                    ControlFile.checkPointCopy.fullPageWrites ? _("on") : _("off"));
 586         printf(_("Latest checkpoint's NextXID:          %u/%u\n"),
 587                    ControlFile.checkPointCopy.nextXidEpoch,
 588                    ControlFile.checkPointCopy.nextXid);
 589         printf(_("Latest checkpoint's NextOID:          %u\n"),
 590                    ControlFile.checkPointCopy.nextOid);
 591         printf(_("Latest checkpoint's NextMultiXactId:  %u\n"),
 592                    ControlFile.checkPointCopy.nextMulti);
 593         printf(_("Latest checkpoint's NextMultiOffset:  %u\n"),
 594                    ControlFile.checkPointCopy.nextMultiOffset);
 595         printf(_("Latest checkpoint's oldestXID:        %u\n"),
 596                    ControlFile.checkPointCopy.oldestXid);
 597         printf(_("Latest checkpoint's oldestXID's DB:   %u\n"),
 598                    ControlFile.checkPointCopy.oldestXidDB);
 599         printf(_("Latest checkpoint's oldestActiveXID:  %u\n"),
 600                    ControlFile.checkPointCopy.oldestActiveXid);
 601         printf(_("Latest checkpoint's oldestMultiXid:   %u\n"),
 602                    ControlFile.checkPointCopy.oldestMulti);
 603         printf(_("Latest checkpoint's oldestMulti's DB: %u\n"),
 604                    ControlFile.checkPointCopy.oldestMultiDB);
 605         printf(_("Maximum data alignment:               %u\n"),
 606                    ControlFile.maxAlign);
 607         /* we don't print floatFormat since can't say much useful about it */
 608         printf(_("Database block size:                  %u\n"),
 609                    ControlFile.blcksz);
 610         printf(_("Blocks per segment of large relation: %u\n"),
 611                    ControlFile.relseg_size);
 612         printf(_("WAL block size:                       %u\n"),
 613                    ControlFile.xlog_blcksz);
 614         printf(_("Bytes per WAL segment:                %u\n"),
 615                    ControlFile.xlog_seg_size);
 616         printf(_("Maximum length of identifiers:        %u\n"),
 617                    ControlFile.nameDataLen);
 618         printf(_("Maximum columns in an index:          %u\n"),
 619                    ControlFile.indexMaxKeys);
 620         printf(_("Maximum size of a TOAST chunk:        %u\n"),
 621                    ControlFile.toast_max_chunk_size);
 622         printf(_("Date/time type storage:               %s\n"),
 623                    (ControlFile.enableIntTimes ? _("64-bit integers") : _("floating-point numbers")));
 624         printf(_("Float4 argument passing:              %s\n"),
 625                    (ControlFile.float4ByVal ? _("by value") : _("by reference")));
 626         printf(_("Float8 argument passing:              %s\n"),
 627                    (ControlFile.float8ByVal ? _("by value") : _("by reference")));
 628         printf(_("Data page checksum version:           %u\n"),
 629                    ControlFile.data_checksum_version);
 630 }
 631
 632
 633 /*
 634  * Write out the new pg_control file.
 635  */
 636 static void
 637 RewriteControlFile(void)
 638 {
 639         int                     fd;
 640         char            buffer[PG_CONTROL_SIZE];                /* need not be aligned */
 641
 642         /*
 643          * Adjust fields as needed to force an empty XLOG starting at
 644          * newXlogSegNo.
 645          */
 646         XLogSegNoOffsetToRecPtr(newXlogSegNo, SizeOfXLogLongPHD,
 647                                                         ControlFile.checkPointCopy.redo);
 648         ControlFile.checkPointCopy.time = (pg_time_t) time(NULL);
 649
 650         ControlFile.state = DB_SHUTDOWNED;
 651         ControlFile.time = (pg_time_t) time(NULL);
 652         ControlFile.checkPoint = ControlFile.checkPointCopy.redo;
 653         ControlFile.prevCheckPoint = 0;
 654         ControlFile.minRecoveryPoint = 0;
 655         ControlFile.minRecoveryPointTLI = 0;
 656         ControlFile.backupStartPoint = 0;
 657         ControlFile.backupEndPoint = 0;
 658         ControlFile.backupEndRequired = false;
 659
 660         /*
 661          * Force the defaults for max_* settings. The values don't really matter
 662          * as long as wal_level='minimal'; the postmaster will reset these fields
 663          * anyway at startup.
 664          */
 665         ControlFile.wal_level = WAL_LEVEL_MINIMAL;
 666         ControlFile.MaxConnections = 100;
 667         ControlFile.max_prepared_xacts = 0;
 668         ControlFile.max_locks_per_xact = 64;
 669
 670         /* Now we can force the recorded xlog seg size to the right thing. */
 671         ControlFile.xlog_seg_size = XLogSegSize;
 672
 673         /* Contents are protected with a CRC */
 674         INIT_CRC32(ControlFile.crc);
 675         COMP_CRC32(ControlFile.crc,
 676                            (char *) &ControlFile,
 677                            offsetof(ControlFileData, crc));
 678         FIN_CRC32(ControlFile.crc);
 679
 680         /*
 681          * We write out PG_CONTROL_SIZE bytes into pg_control, zero-padding the
 682          * excess over sizeof(ControlFileData).  This reduces the odds of
 683          * premature-EOF errors when reading pg_control.  We'll still fail when we
 684          * check the contents of the file, but hopefully with a more specific
 685          * error than "couldn't read pg_control".
 686          */
 687         if (sizeof(ControlFileData) > PG_CONTROL_SIZE)
 688         {
 689                 fprintf(stderr,
 690                                 _("%s: internal error -- sizeof(ControlFileData) is too large ... fix PG_CONTROL_SIZE\n"),
 691                                 progname);
 692                 exit(1);
 693         }
 694
 695         memset(buffer, 0, PG_CONTROL_SIZE);
 696         memcpy(buffer, &ControlFile, sizeof(ControlFileData));
 697
 698         unlink(XLOG_CONTROL_FILE);
 699
 700         fd = open(XLOG_CONTROL_FILE,
 701                           O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
 702                           S_IRUSR | S_IWUSR);
 703         if (fd < 0)
 704         {
 705                 fprintf(stderr, _("%s: could not create pg_control file: %s\n"),
 706                                 progname, strerror(errno));
 707                 exit(1);
 708         }
 709
 710         errno = 0;
 711         if (write(fd, buffer, PG_CONTROL_SIZE) != PG_CONTROL_SIZE)
 712         {
 713                 /* if write didn't set errno, assume problem is no disk space */
 714                 if (errno == 0)
 715                         errno = ENOSPC;
 716                 fprintf(stderr, _("%s: could not write pg_control file: %s\n"),
 717                                 progname, strerror(errno));
 718                 exit(1);
 719         }
 720
 721         if (fsync(fd) != 0)
 722         {
 723                 fprintf(stderr, _("%s: fsync error: %s\n"), progname, strerror(errno));
 724                 exit(1);
 725         }
 726
 727         close(fd);
 728 }
 729
 730
 731 /*
 732  * Scan existing XLOG files and determine the highest existing WAL address
 733  *
 734  * On entry, ControlFile.checkPointCopy.redo and ControlFile.xlog_seg_size
 735  * are assumed valid (note that we allow the old xlog seg size to differ
 736  * from what we're using).  On exit, newXlogId and newXlogSeg are set to
 737  * suitable values for the beginning of replacement WAL (in our seg size).
 738  */
 739 static void
 740 FindEndOfXLOG(void)
 741 {
 742         DIR                *xldir;
 743         struct dirent *xlde;
 744         uint64          segs_per_xlogid;
 745         uint64          xlogbytepos;
 746
 747         /*
 748          * Initialize the max() computation using the last checkpoint address from
 749          * old pg_control.      Note that for the moment we are working with segment
 750          * numbering according to the old xlog seg size.
 751          */
 752         segs_per_xlogid = (UINT64CONST(0x0000000100000000) / ControlFile.xlog_seg_size);
 753         newXlogSegNo = ControlFile.checkPointCopy.redo / ControlFile.xlog_seg_size;
 754
 755         /*
 756          * Scan the pg_xlog directory to find existing WAL segment files. We
 757          * assume any present have been used; in most scenarios this should be
 758          * conservative, because of xlog.c's attempts to pre-create files.
 759          */
 760         xldir = opendir(XLOGDIR);
 761         if (xldir == NULL)
 762         {
 763                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 764                                 progname, XLOGDIR, strerror(errno));
 765                 exit(1);
 766         }
 767
 768         errno = 0;
 769         while ((xlde = readdir(xldir)) != NULL)
 770         {
 771                 if (strlen(xlde->d_name) == 24 &&
 772                         strspn(xlde->d_name, "0123456789ABCDEF") == 24)
 773                 {
 774                         unsigned int tli,
 775                                                 log,
 776                                                 seg;
 777                         XLogSegNo       segno;
 778
 779                         sscanf(xlde->d_name, "%08X%08X%08X", &tli, &log, &seg);
 780                         segno = ((uint64) log) * segs_per_xlogid + seg;
 781
 782                         /*
 783                          * Note: we take the max of all files found, regardless of their
 784                          * timelines.  Another possibility would be to ignore files of
 785                          * timelines other than the target TLI, but this seems safer.
 786                          * Better too large a result than too small...
 787                          */
 788                         if (segno > newXlogSegNo)
 789                                 newXlogSegNo = segno;
 790                 }
 791                 errno = 0;
 792         }
 793 #ifdef WIN32
 794
 795         /*
 796          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 797          * released version
 798          */
 799         if (GetLastError() == ERROR_NO_MORE_FILES)
 800                 errno = 0;
 801 #endif
 802
 803         if (errno)
 804         {
 805                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 806                                 progname, XLOGDIR, strerror(errno));
 807                 exit(1);
 808         }
 809         closedir(xldir);
 810
 811         /*
 812          * Finally, convert to new xlog seg size, and advance by one to ensure we
 813          * are in virgin territory.
 814          */
 815         xlogbytepos = newXlogSegNo * ControlFile.xlog_seg_size;
 816         newXlogSegNo = (xlogbytepos + XLogSegSize - 1) / XLogSegSize;
 817         newXlogSegNo++;
 818 }
 819
 820
 821 /*
 822  * Remove existing XLOG files
 823  */
 824 static void
 825 KillExistingXLOG(void)
 826 {
 827         DIR                *xldir;
 828         struct dirent *xlde;
 829         char            path[MAXPGPATH];
 830
 831         xldir = opendir(XLOGDIR);
 832         if (xldir == NULL)
 833         {
 834                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 835                                 progname, XLOGDIR, strerror(errno));
 836                 exit(1);
 837         }
 838
 839         errno = 0;
 840         while ((xlde = readdir(xldir)) != NULL)
 841         {
 842                 if (strlen(xlde->d_name) == 24 &&
 843                         strspn(xlde->d_name, "0123456789ABCDEF") == 24)
 844                 {
 845                         snprintf(path, MAXPGPATH, "%s/%s", XLOGDIR, xlde->d_name);
 846                         if (unlink(path) < 0)
 847                         {
 848                                 fprintf(stderr, _("%s: could not delete file \"%s\": %s\n"),
 849                                                 progname, path, strerror(errno));
 850                                 exit(1);
 851                         }
 852                 }
 853                 errno = 0;
 854         }
 855 #ifdef WIN32
 856
 857         /*
 858          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 859          * released version
 860          */
 861         if (GetLastError() == ERROR_NO_MORE_FILES)
 862                 errno = 0;
 863 #endif
 864
 865         if (errno)
 866         {
 867                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 868                                 progname, XLOGDIR, strerror(errno));
 869                 exit(1);
 870         }
 871         closedir(xldir);
 872 }
 873
 874
 875 /*
 876  * Remove existing archive status files
 877  */
 878 static void
 879 KillExistingArchiveStatus(void)
 880 {
 881         DIR                *xldir;
 882         struct dirent *xlde;
 883         char            path[MAXPGPATH];
 884
 885 #define ARCHSTATDIR XLOGDIR "/archive_status"
 886
 887         xldir = opendir(ARCHSTATDIR);
 888         if (xldir == NULL)
 889         {
 890                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 891                                 progname, ARCHSTATDIR, strerror(errno));
 892                 exit(1);
 893         }
 894
 895         errno = 0;
 896         while ((xlde = readdir(xldir)) != NULL)
 897         {
 898                 if (strspn(xlde->d_name, "0123456789ABCDEF") == 24 &&
 899                         (strcmp(xlde->d_name + 24, ".ready") == 0 ||
 900                          strcmp(xlde->d_name + 24, ".done") == 0))
 901                 {
 902                         snprintf(path, MAXPGPATH, "%s/%s", ARCHSTATDIR, xlde->d_name);
 903                         if (unlink(path) < 0)
 904                         {
 905                                 fprintf(stderr, _("%s: could not delete file \"%s\": %s\n"),
 906                                                 progname, path, strerror(errno));
 907                                 exit(1);
 908                         }
 909                 }
 910                 errno = 0;
 911         }
 912 #ifdef WIN32
 913
 914         /*
 915          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 916          * released version
 917          */
 918         if (GetLastError() == ERROR_NO_MORE_FILES)
 919                 errno = 0;
 920 #endif
 921
 922         if (errno)
 923         {
 924                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 925                                 progname, ARCHSTATDIR, strerror(errno));
 926                 exit(1);
 927         }
 928         closedir(xldir);
 929 }
 930
 931
 932 /*
 933  * Write an empty XLOG file, containing only the checkpoint record
 934  * already set up in ControlFile.
 935  */
 936 static void
 937 WriteEmptyXLOG(void)
 938 {
 939         char       *buffer;
 940         XLogPageHeader page;
 941         XLogLongPageHeader longpage;
 942         XLogRecord *record;
 943         pg_crc32        crc;
 944         char            path[MAXPGPATH];
 945         int                     fd;
 946         int                     nbytes;
 947
 948         /* Use malloc() to ensure buffer is MAXALIGNED */
 949         buffer = (char *) pg_malloc(XLOG_BLCKSZ);
 950         page = (XLogPageHeader) buffer;
 951         memset(buffer, 0, XLOG_BLCKSZ);
 952
 953         /* Set up the XLOG page header */
 954         page->xlp_magic = XLOG_PAGE_MAGIC;
 955         page->xlp_info = XLP_LONG_HEADER;
 956         page->xlp_tli = ControlFile.checkPointCopy.ThisTimeLineID;
 957         page->xlp_pageaddr = ControlFile.checkPointCopy.redo - SizeOfXLogLongPHD;
 958         longpage = (XLogLongPageHeader) page;
 959         longpage->xlp_sysid = ControlFile.system_identifier;
 960         longpage->xlp_seg_size = XLogSegSize;
 961         longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
 962
 963         /* Insert the initial checkpoint record */
 964         record = (XLogRecord *) ((char *) page + SizeOfXLogLongPHD);
 965         record->xl_prev = 0;
 966         record->xl_xid = InvalidTransactionId;
 967         record->xl_tot_len = SizeOfXLogRecord + sizeof(CheckPoint);
 968         record->xl_len = sizeof(CheckPoint);
 969         record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
 970         record->xl_rmid = RM_XLOG_ID;
 971         memcpy(XLogRecGetData(record), &ControlFile.checkPointCopy,
 972                    sizeof(CheckPoint));
 973
 974         INIT_CRC32(crc);
 975         COMP_CRC32(crc, &ControlFile.checkPointCopy, sizeof(CheckPoint));
 976         COMP_CRC32(crc, (char *) record, offsetof(XLogRecord, xl_crc));
 977         FIN_CRC32(crc);
 978         record->xl_crc = crc;
 979
 980         /* Write the first page */
 981         XLogFilePath(path, ControlFile.checkPointCopy.ThisTimeLineID, newXlogSegNo);
 982
 983         unlink(path);
 984
 985         fd = open(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
 986                           S_IRUSR | S_IWUSR);
 987         if (fd < 0)
 988         {
 989                 fprintf(stderr, _("%s: could not open file \"%s\": %s\n"),
 990                                 progname, path, strerror(errno));
 991                 exit(1);
 992         }
 993
 994         errno = 0;
 995         if (write(fd, buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
 996         {
 997                 /* if write didn't set errno, assume problem is no disk space */
 998                 if (errno == 0)
 999                         errno = ENOSPC;
1000                 fprintf(stderr, _("%s: could not write file \"%s\": %s\n"),
1001                                 progname, path, strerror(errno));
1002                 exit(1);
1003         }
1004
1005         /* Fill the rest of the file with zeroes */
1006         memset(buffer, 0, XLOG_BLCKSZ);
1007         for (nbytes = XLOG_BLCKSZ; nbytes < XLogSegSize; nbytes += XLOG_BLCKSZ)
1008         {
1009                 errno = 0;
1010                 if (write(fd, buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
1011                 {
1012                         if (errno == 0)
1013                                 errno = ENOSPC;
1014                         fprintf(stderr, _("%s: could not write file \"%s\": %s\n"),
1015                                         progname, path, strerror(errno));
1016                         exit(1);
1017                 }
1018         }
1019
1020         if (fsync(fd) != 0)
1021         {
1022                 fprintf(stderr, _("%s: fsync error: %s\n"), progname, strerror(errno));
1023                 exit(1);
1024         }
1025
1026         close(fd);
1027 }
1028
1029
1030 static void
1031 usage(void)
1032 {
1033         printf(_("%s resets the PostgreSQL transaction log.\n\n"), progname);
1034         printf(_("Usage:\n  %s [OPTION]... DATADIR\n\n"), progname);
1035         printf(_("Options:\n"));
1036         printf(_("  -e XIDEPOCH      set next transaction ID epoch\n"));
1037         printf(_("  -f               force update to be done\n"));
1038         printf(_("  -l XLOGFILE      force minimum WAL starting location for new transaction log\n"));
1039         printf(_("  -m XID,XID       set next and oldest multitransaction ID\n"));
1040         printf(_("  -n               no update, just show extracted control values (for testing)\n"));
1041         printf(_("  -o OID           set next OID\n"));
1042         printf(_("  -O OFFSET        set next multitransaction offset\n"));
1043         printf(_("  -V, --version    output version information, then exit\n"));
1044         printf(_("  -x XID           set next transaction ID\n"));
1045         printf(_("  -?, --help       show this help, then exit\n"));
1046         printf(_("\nReport bugs to <pgsql-bugs@postgresql.org>.\n"));
1047 }