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 static uint32 set_xid_epoch = (uint32) -1;
  68 static TransactionId set_xid = 0;
  69 static Oid      set_oid = 0;
  70 static MultiXactId set_mxid = 0;
  71 static MultiXactOffset set_mxoff = (MultiXactOffset) -1;
  72 static uint32 minXlogTli = 0;
  73 static XLogSegNo minXlogSegNo = 0;
  74
  75 static bool ReadControlFile(void);
  76 static void GuessControlValues(void);
  77 static void PrintControlValues(bool guessed);
  78 static void PrintNewControlValues(void);
  79 static void RewriteControlFile(void);
  80 static void FindEndOfXLOG(void);
  81 static void KillExistingXLOG(void);
  82 static void KillExistingArchiveStatus(void);
  83 static void WriteEmptyXLOG(void);
  84 static void usage(void);
  85
  86
  87 int
  88 main(int argc, char *argv[])
  89 {
  90         int                     c;
  91         bool            force = false;
  92         bool            noupdate = false;
  93         MultiXactId set_oldestmxid = 0;
  94         char       *endptr;
  95         char       *endptr2;
  96         char       *DataDir;
  97         int                     fd;
  98
  99         set_pglocale_pgservice(argv[0], PG_TEXTDOMAIN("pg_resetxlog"));
 100
 101         progname = get_progname(argv[0]);
 102
 103         if (argc > 1)
 104         {
 105                 if (strcmp(argv[1], "--help") == 0 || strcmp(argv[1], "-?") == 0)
 106                 {
 107                         usage();
 108                         exit(0);
 109                 }
 110                 if (strcmp(argv[1], "--version") == 0 || strcmp(argv[1], "-V") == 0)
 111                 {
 112                         puts("pg_resetxlog (PostgreSQL) " PG_VERSION);
 113                         exit(0);
 114                 }
 115         }
 116
 117
 118         while ((c = getopt(argc, argv, "fl:m:no:O:x:e:")) != -1)
 119         {
 120                 switch (c)
 121                 {
 122                         case 'f':
 123                                 force = true;
 124                                 break;
 125
 126                         case 'n':
 127                                 noupdate = true;
 128                                 break;
 129
 130                         case 'e':
 131                                 set_xid_epoch = strtoul(optarg, &endptr, 0);
 132                                 if (endptr == optarg || *endptr != '\0')
 133                                 {
 134                                         fprintf(stderr, _("%s: invalid argument for option -e\n"), progname);
 135                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 136                                         exit(1);
 137                                 }
 138                                 if (set_xid_epoch == -1)
 139                                 {
 140                                         fprintf(stderr, _("%s: transaction ID epoch (-e) must not be -1\n"), progname);
 141                                         exit(1);
 142                                 }
 143                                 break;
 144
 145                         case 'x':
 146                                 set_xid = strtoul(optarg, &endptr, 0);
 147                                 if (endptr == optarg || *endptr != '\0')
 148                                 {
 149                                         fprintf(stderr, _("%s: invalid argument for option -x\n"), progname);
 150                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 151                                         exit(1);
 152                                 }
 153                                 if (set_xid == 0)
 154                                 {
 155                                         fprintf(stderr, _("%s: transaction ID (-x) must not be 0\n"), progname);
 156                                         exit(1);
 157                                 }
 158                                 break;
 159
 160                         case 'o':
 161                                 set_oid = strtoul(optarg, &endptr, 0);
 162                                 if (endptr == optarg || *endptr != '\0')
 163                                 {
 164                                         fprintf(stderr, _("%s: invalid argument for option -o\n"), progname);
 165                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 166                                         exit(1);
 167                                 }
 168                                 if (set_oid == 0)
 169                                 {
 170                                         fprintf(stderr, _("%s: OID (-o) must not be 0\n"), progname);
 171                                         exit(1);
 172                                 }
 173                                 break;
 174
 175                         case 'm':
 176                                 set_mxid = strtoul(optarg, &endptr, 0);
 177                                 if (endptr == optarg || *endptr != ',')
 178                                 {
 179                                         fprintf(stderr, _("%s: invalid argument for option -m\n"), progname);
 180                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 181                                         exit(1);
 182                                 }
 183
 184                                 set_oldestmxid = strtoul(endptr + 1, &endptr2, 0);
 185                                 if (endptr2 == endptr + 1 || *endptr2 != '\0')
 186                                 {
 187                                         fprintf(stderr, _("%s: invalid argument for option -m\n"), progname);
 188                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 189                                         exit(1);
 190                                 }
 191                                 if (set_mxid == 0)
 192                                 {
 193                                         fprintf(stderr, _("%s: multitransaction ID (-m) must not be 0\n"), progname);
 194                                         exit(1);
 195                                 }
 196
 197                                 /*
 198                                  * XXX It'd be nice to have more sanity checks here, e.g. so
 199                                  * that oldest is not wrapped around w.r.t. nextMulti.
 200                                  */
 201                                 if (set_oldestmxid == 0)
 202                                 {
 203                                         fprintf(stderr, _("%s: oldest multitransaction ID (-m) must not be 0\n"),
 204                                                         progname);
 205                                         exit(1);
 206                                 }
 207                                 break;
 208
 209                         case 'O':
 210                                 set_mxoff = strtoul(optarg, &endptr, 0);
 211                                 if (endptr == optarg || *endptr != '\0')
 212                                 {
 213                                         fprintf(stderr, _("%s: invalid argument for option -O\n"), progname);
 214                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 215                                         exit(1);
 216                                 }
 217                                 if (set_mxoff == -1)
 218                                 {
 219                                         fprintf(stderr, _("%s: multitransaction offset (-O) must not be -1\n"), progname);
 220                                         exit(1);
 221                                 }
 222                                 break;
 223
 224                         case 'l':
 225                                 if (strspn(optarg, "01234567890ABCDEFabcdef") != 24)
 226                                 {
 227                                         fprintf(stderr, _("%s: invalid argument for option -l\n"), progname);
 228                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 229                                         exit(1);
 230                                 }
 231                                 XLogFromFileName(optarg, &minXlogTli, &minXlogSegNo);
 232                                 break;
 233
 234                         default:
 235                                 fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 236                                 exit(1);
 237                 }
 238         }
 239
 240         if (optind == argc)
 241         {
 242                 fprintf(stderr, _("%s: no data directory specified\n"), progname);
 243                 fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 244                 exit(1);
 245         }
 246
 247         /*
 248          * Don't allow pg_resetxlog to be run as root, to avoid overwriting the
 249          * ownership of files in the data directory. We need only check for root
 250          * -- any other user won't have sufficient permissions to modify files in
 251          * the data directory.
 252          */
 253 #ifndef WIN32
 254         if (geteuid() == 0)
 255         {
 256                 fprintf(stderr, _("%s: cannot be executed by \"root\"\n"),
 257                                 progname);
 258                 fprintf(stderr, _("You must run %s as the PostgreSQL superuser.\n"),
 259                                 progname);
 260                 exit(1);
 261         }
 262 #endif
 263
 264         DataDir = argv[optind];
 265
 266         if (chdir(DataDir) < 0)
 267         {
 268                 fprintf(stderr, _("%s: could not change directory to \"%s\": %s\n"),
 269                                 progname, DataDir, strerror(errno));
 270                 exit(1);
 271         }
 272
 273         /*
 274          * Check for a postmaster lock file --- if there is one, refuse to
 275          * proceed, on grounds we might be interfering with a live installation.
 276          */
 277         if ((fd = open("postmaster.pid", O_RDONLY, 0)) < 0)
 278         {
 279                 if (errno != ENOENT)
 280                 {
 281                         fprintf(stderr, _("%s: could not open file \"%s\" for reading: %s\n"),
 282                                         progname, "postmaster.pid", strerror(errno));
 283                         exit(1);
 284                 }
 285         }
 286         else
 287         {
 288                 fprintf(stderr, _("%s: lock file \"%s\" exists\n"
 289                                                   "Is a server running?  If not, delete the lock file and try again.\n"),
 290                                 progname, "postmaster.pid");
 291                 exit(1);
 292         }
 293
 294         /*
 295          * Attempt to read the existing pg_control file
 296          */
 297         if (!ReadControlFile())
 298                 GuessControlValues();
 299
 300         /*
 301          * Also look at existing segment files to set up newXlogSegNo
 302          */
 303         FindEndOfXLOG();
 304
 305         /*
 306          * If we're not going to proceed with the reset, print the current control
 307          * file parameters.
 308          */
 309         if ((guessed && !force) || noupdate)
 310                 PrintControlValues(guessed);
 311
 312         /*
 313          * Adjust fields if required by switches.  (Do this now so that printout,
 314          * if any, includes these values.)
 315          */
 316         if (set_xid_epoch != -1)
 317                 ControlFile.checkPointCopy.nextXidEpoch = set_xid_epoch;
 318
 319         if (set_xid != 0)
 320         {
 321                 ControlFile.checkPointCopy.nextXid = set_xid;
 322
 323                 /*
 324                  * For the moment, just set oldestXid to a value that will force
 325                  * immediate autovacuum-for-wraparound.  It's not clear whether adding
 326                  * user control of this is useful, so let's just do something that's
 327                  * reasonably safe.  The magic constant here corresponds to the
 328                  * maximum allowed value of autovacuum_freeze_max_age.
 329                  */
 330                 ControlFile.checkPointCopy.oldestXid = set_xid - 2000000000;
 331                 if (ControlFile.checkPointCopy.oldestXid < FirstNormalTransactionId)
 332                         ControlFile.checkPointCopy.oldestXid += FirstNormalTransactionId;
 333                 ControlFile.checkPointCopy.oldestXidDB = InvalidOid;
 334         }
 335
 336         if (set_oid != 0)
 337                 ControlFile.checkPointCopy.nextOid = set_oid;
 338
 339         if (set_mxid != 0)
 340         {
 341                 ControlFile.checkPointCopy.nextMulti = set_mxid;
 342
 343                 ControlFile.checkPointCopy.oldestMulti = set_oldestmxid;
 344                 if (ControlFile.checkPointCopy.oldestMulti < FirstMultiXactId)
 345                         ControlFile.checkPointCopy.oldestMulti += FirstMultiXactId;
 346                 ControlFile.checkPointCopy.oldestMultiDB = InvalidOid;
 347         }
 348
 349         if (set_mxoff != -1)
 350                 ControlFile.checkPointCopy.nextMultiOffset = set_mxoff;
 351
 352         if (minXlogTli > ControlFile.checkPointCopy.ThisTimeLineID)
 353         {
 354                 ControlFile.checkPointCopy.ThisTimeLineID = minXlogTli;
 355                 ControlFile.checkPointCopy.PrevTimeLineID = minXlogTli;
 356         }
 357
 358         if (minXlogSegNo > newXlogSegNo)
 359                 newXlogSegNo = minXlogSegNo;
 360
 361         /*
 362          * If we had to guess anything, and -f was not given, just print the
 363          * guessed values and exit.  Also print if -n is given.
 364          */
 365         if ((guessed && !force) || noupdate)
 366         {
 367                 PrintNewControlValues();
 368                 if (!noupdate)
 369                 {
 370                         printf(_("\nIf these values seem acceptable, use -f to force reset.\n"));
 371                         exit(1);
 372                 }
 373                 else
 374                         exit(0);
 375         }
 376
 377         /*
 378          * Don't reset from a dirty pg_control without -f, either.
 379          */
 380         if (ControlFile.state != DB_SHUTDOWNED && !force)
 381         {
 382                 printf(_("The database server was not shut down cleanly.\n"
 383                            "Resetting the transaction log might cause data to be lost.\n"
 384                                  "If you want to proceed anyway, use -f to force reset.\n"));
 385                 exit(1);
 386         }
 387
 388         /*
 389          * Else, do the dirty deed.
 390          */
 391         RewriteControlFile();
 392         KillExistingXLOG();
 393         KillExistingArchiveStatus();
 394         WriteEmptyXLOG();
 395
 396         printf(_("Transaction log reset\n"));
 397         return 0;
 398 }
 399
 400
 401 /*
 402  * Try to read the existing pg_control file.
 403  *
 404  * This routine is also responsible for updating old pg_control versions
 405  * to the current format.  (Currently we don't do anything of the sort.)
 406  */
 407 static bool
 408 ReadControlFile(void)
 409 {
 410         int                     fd;
 411         int                     len;
 412         char       *buffer;
 413         pg_crc32        crc;
 414
 415         if ((fd = open(XLOG_CONTROL_FILE, O_RDONLY | PG_BINARY, 0)) < 0)
 416         {
 417                 /*
 418                  * If pg_control is not there at all, or we can't read it, the odds
 419                  * are we've been handed a bad DataDir path, so give up. User can do
 420                  * "touch pg_control" to force us to proceed.
 421                  */
 422                 fprintf(stderr, _("%s: could not open file \"%s\" for reading: %s\n"),
 423                                 progname, XLOG_CONTROL_FILE, strerror(errno));
 424                 if (errno == ENOENT)
 425                         fprintf(stderr, _("If you are sure the data directory path is correct, execute\n"
 426                                                           "  touch %s\n"
 427                                                           "and try again.\n"),
 428                                         XLOG_CONTROL_FILE);
 429                 exit(1);
 430         }
 431
 432         /* Use malloc to ensure we have a maxaligned buffer */
 433         buffer = (char *) pg_malloc(PG_CONTROL_SIZE);
 434
 435         len = read(fd, buffer, PG_CONTROL_SIZE);
 436         if (len < 0)
 437         {
 438                 fprintf(stderr, _("%s: could not read file \"%s\": %s\n"),
 439                                 progname, XLOG_CONTROL_FILE, strerror(errno));
 440                 exit(1);
 441         }
 442         close(fd);
 443
 444         if (len >= sizeof(ControlFileData) &&
 445           ((ControlFileData *) buffer)->pg_control_version == PG_CONTROL_VERSION)
 446         {
 447                 /* Check the CRC. */
 448                 INIT_CRC32(crc);
 449                 COMP_CRC32(crc,
 450                                    buffer,
 451                                    offsetof(ControlFileData, crc));
 452                 FIN_CRC32(crc);
 453
 454                 if (EQ_CRC32(crc, ((ControlFileData *) buffer)->crc))
 455                 {
 456                         /* Valid data... */
 457                         memcpy(&ControlFile, buffer, sizeof(ControlFile));
 458                         return true;
 459                 }
 460
 461                 fprintf(stderr, _("%s: pg_control exists but has invalid CRC; proceed with caution\n"),
 462                                 progname);
 463                 /* We will use the data anyway, but treat it as guessed. */
 464                 memcpy(&ControlFile, buffer, sizeof(ControlFile));
 465                 guessed = true;
 466                 return true;
 467         }
 468
 469         /* Looks like it's a mess. */
 470         fprintf(stderr, _("%s: pg_control exists but is broken or unknown version; ignoring it\n"),
 471                         progname);
 472         return false;
 473 }
 474
 475
 476 /*
 477  * Guess at pg_control values when we can't read the old ones.
 478  */
 479 static void
 480 GuessControlValues(void)
 481 {
 482         uint64          sysidentifier;
 483         struct timeval tv;
 484
 485         /*
 486          * Set up a completely default set of pg_control values.
 487          */
 488         guessed = true;
 489         memset(&ControlFile, 0, sizeof(ControlFile));
 490
 491         ControlFile.pg_control_version = PG_CONTROL_VERSION;
 492         ControlFile.catalog_version_no = CATALOG_VERSION_NO;
 493
 494         /*
 495          * Create a new unique installation identifier, since we can no longer use
 496          * any old XLOG records.  See notes in xlog.c about the algorithm.
 497          */
 498         gettimeofday(&tv, NULL);
 499         sysidentifier = ((uint64) tv.tv_sec) << 32;
 500         sysidentifier |= (uint32) (tv.tv_sec | tv.tv_usec);
 501
 502         ControlFile.system_identifier = sysidentifier;
 503
 504         ControlFile.checkPointCopy.redo = SizeOfXLogLongPHD;
 505         ControlFile.checkPointCopy.ThisTimeLineID = 1;
 506         ControlFile.checkPointCopy.PrevTimeLineID = 1;
 507         ControlFile.checkPointCopy.fullPageWrites = false;
 508         ControlFile.checkPointCopy.nextXidEpoch = 0;
 509         ControlFile.checkPointCopy.nextXid = FirstNormalTransactionId;
 510         ControlFile.checkPointCopy.nextOid = FirstBootstrapObjectId;
 511         ControlFile.checkPointCopy.nextMulti = FirstMultiXactId;
 512         ControlFile.checkPointCopy.nextMultiOffset = 0;
 513         ControlFile.checkPointCopy.oldestXid = FirstNormalTransactionId;
 514         ControlFile.checkPointCopy.oldestXidDB = InvalidOid;
 515         ControlFile.checkPointCopy.oldestMulti = FirstMultiXactId;
 516         ControlFile.checkPointCopy.oldestMultiDB = InvalidOid;
 517         ControlFile.checkPointCopy.time = (pg_time_t) time(NULL);
 518         ControlFile.checkPointCopy.oldestActiveXid = InvalidTransactionId;
 519
 520         ControlFile.state = DB_SHUTDOWNED;
 521         ControlFile.time = (pg_time_t) time(NULL);
 522         ControlFile.checkPoint = ControlFile.checkPointCopy.redo;
 523         ControlFile.unloggedLSN = 1;
 524
 525         /* minRecoveryPoint, backupStartPoint and backupEndPoint can be left zero */
 526
 527         ControlFile.wal_level = WAL_LEVEL_MINIMAL;
 528         ControlFile.wal_log_hintbits = false;
 529         ControlFile.MaxConnections = 100;
 530         ControlFile.max_worker_processes = 8;
 531         ControlFile.max_prepared_xacts = 0;
 532         ControlFile.max_locks_per_xact = 64;
 533
 534         ControlFile.maxAlign = MAXIMUM_ALIGNOF;
 535         ControlFile.floatFormat = FLOATFORMAT_VALUE;
 536         ControlFile.blcksz = BLCKSZ;
 537         ControlFile.relseg_size = RELSEG_SIZE;
 538         ControlFile.xlog_blcksz = XLOG_BLCKSZ;
 539         ControlFile.xlog_seg_size = XLOG_SEG_SIZE;
 540         ControlFile.nameDataLen = NAMEDATALEN;
 541         ControlFile.indexMaxKeys = INDEX_MAX_KEYS;
 542         ControlFile.toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
 543 #ifdef HAVE_INT64_TIMESTAMP
 544         ControlFile.enableIntTimes = true;
 545 #else
 546         ControlFile.enableIntTimes = false;
 547 #endif
 548         ControlFile.float4ByVal = FLOAT4PASSBYVAL;
 549         ControlFile.float8ByVal = FLOAT8PASSBYVAL;
 550
 551         /*
 552          * XXX eventually, should try to grovel through old XLOG to develop more
 553          * accurate values for TimeLineID, nextXID, etc.
 554          */
 555 }
 556
 557
 558 /*
 559  * Print the guessed pg_control values when we had to guess.
 560  *
 561  * NB: this display should be just those fields that will not be
 562  * reset by RewriteControlFile().
 563  */
 564 static void
 565 PrintControlValues(bool guessed)
 566 {
 567         char            sysident_str[32];
 568
 569         if (guessed)
 570                 printf(_("Guessed pg_control values:\n\n"));
 571         else
 572                 printf(_("Current pg_control values:\n\n"));
 573
 574         /*
 575          * Format system_identifier separately to keep platform-dependent format
 576          * code out of the translatable message string.
 577          */
 578         snprintf(sysident_str, sizeof(sysident_str), UINT64_FORMAT,
 579                          ControlFile.system_identifier);
 580
 581         printf(_("pg_control version number:            %u\n"),
 582                    ControlFile.pg_control_version);
 583         printf(_("Catalog version number:               %u\n"),
 584                    ControlFile.catalog_version_no);
 585         printf(_("Database system identifier:           %s\n"),
 586                    sysident_str);
 587         printf(_("Latest checkpoint's TimeLineID:       %u\n"),
 588                    ControlFile.checkPointCopy.ThisTimeLineID);
 589         printf(_("Latest checkpoint's full_page_writes: %s\n"),
 590                    ControlFile.checkPointCopy.fullPageWrites ? _("on") : _("off"));
 591         printf(_("Latest checkpoint's NextXID:          %u/%u\n"),
 592                    ControlFile.checkPointCopy.nextXidEpoch,
 593                    ControlFile.checkPointCopy.nextXid);
 594         printf(_("Latest checkpoint's NextOID:          %u\n"),
 595                    ControlFile.checkPointCopy.nextOid);
 596         printf(_("Latest checkpoint's NextMultiXactId:  %u\n"),
 597                    ControlFile.checkPointCopy.nextMulti);
 598         printf(_("Latest checkpoint's NextMultiOffset:  %u\n"),
 599                    ControlFile.checkPointCopy.nextMultiOffset);
 600         printf(_("Latest checkpoint's oldestXID:        %u\n"),
 601                    ControlFile.checkPointCopy.oldestXid);
 602         printf(_("Latest checkpoint's oldestXID's DB:   %u\n"),
 603                    ControlFile.checkPointCopy.oldestXidDB);
 604         printf(_("Latest checkpoint's oldestActiveXID:  %u\n"),
 605                    ControlFile.checkPointCopy.oldestActiveXid);
 606         printf(_("Latest checkpoint's oldestMultiXid:   %u\n"),
 607                    ControlFile.checkPointCopy.oldestMulti);
 608         printf(_("Latest checkpoint's oldestMulti's DB: %u\n"),
 609                    ControlFile.checkPointCopy.oldestMultiDB);
 610         printf(_("Maximum data alignment:               %u\n"),
 611                    ControlFile.maxAlign);
 612         /* we don't print floatFormat since can't say much useful about it */
 613         printf(_("Database block size:                  %u\n"),
 614                    ControlFile.blcksz);
 615         printf(_("Blocks per segment of large relation: %u\n"),
 616                    ControlFile.relseg_size);
 617         printf(_("WAL block size:                       %u\n"),
 618                    ControlFile.xlog_blcksz);
 619         printf(_("Bytes per WAL segment:                %u\n"),
 620                    ControlFile.xlog_seg_size);
 621         printf(_("Maximum length of identifiers:        %u\n"),
 622                    ControlFile.nameDataLen);
 623         printf(_("Maximum columns in an index:          %u\n"),
 624                    ControlFile.indexMaxKeys);
 625         printf(_("Maximum size of a TOAST chunk:        %u\n"),
 626                    ControlFile.toast_max_chunk_size);
 627         printf(_("Date/time type storage:               %s\n"),
 628                    (ControlFile.enableIntTimes ? _("64-bit integers") : _("floating-point numbers")));
 629         printf(_("Float4 argument passing:              %s\n"),
 630                    (ControlFile.float4ByVal ? _("by value") : _("by reference")));
 631         printf(_("Float8 argument passing:              %s\n"),
 632                    (ControlFile.float8ByVal ? _("by value") : _("by reference")));
 633         printf(_("Data page checksum version:           %u\n"),
 634                    ControlFile.data_checksum_version);
 635 }
 636
 637
 638 /*
 639  * Print the values to be changed.
 640  */
 641 static void
 642 PrintNewControlValues()
 643 {
 644         char            fname[MAXFNAMELEN];
 645
 646         /* This will be always printed in order to keep format same. */
 647         printf(_("\n\nValues to be changed:\n\n"));
 648
 649         XLogFileName(fname, ControlFile.checkPointCopy.ThisTimeLineID, newXlogSegNo);
 650         printf(_("First log segment after reset:        %s\n"), fname);
 651
 652         if (set_mxid != 0)
 653         {
 654                 printf(_("NextMultiXactId:                      %u\n"),
 655                            ControlFile.checkPointCopy.nextMulti);
 656                 printf(_("OldestMultiXid:                       %u\n"),
 657                            ControlFile.checkPointCopy.oldestMulti);
 658                 printf(_("OldestMulti's DB:                     %u\n"),
 659                            ControlFile.checkPointCopy.oldestMultiDB);
 660         }
 661
 662         if (set_mxoff != -1)
 663         {
 664                 printf(_("NextMultiOffset:                      %u\n"),
 665                            ControlFile.checkPointCopy.nextMultiOffset);
 666         }
 667
 668         if (set_oid != 0)
 669         {
 670                 printf(_("NextOID:                              %u\n"),
 671                            ControlFile.checkPointCopy.nextOid);
 672         }
 673
 674         if (set_xid != 0)
 675         {
 676                 printf(_("NextXID:                              %u\n"),
 677                            ControlFile.checkPointCopy.nextXid);
 678                 printf(_("OldestXID:                            %u\n"),
 679                            ControlFile.checkPointCopy.oldestXid);
 680                 printf(_("OldestXID's DB:                       %u\n"),
 681                            ControlFile.checkPointCopy.oldestXidDB);
 682         }
 683
 684         if (set_xid_epoch != -1)
 685         {
 686                 printf(_("NextXID Epoch:                        %u\n"),
 687                            ControlFile.checkPointCopy.nextXidEpoch);
 688         }
 689 }
 690
 691
 692 /*
 693  * Write out the new pg_control file.
 694  */
 695 static void
 696 RewriteControlFile(void)
 697 {
 698         int                     fd;
 699         char            buffer[PG_CONTROL_SIZE];                /* need not be aligned */
 700
 701         /*
 702          * Adjust fields as needed to force an empty XLOG starting at
 703          * newXlogSegNo.
 704          */
 705         XLogSegNoOffsetToRecPtr(newXlogSegNo, SizeOfXLogLongPHD,
 706                                                         ControlFile.checkPointCopy.redo);
 707         ControlFile.checkPointCopy.time = (pg_time_t) time(NULL);
 708
 709         ControlFile.state = DB_SHUTDOWNED;
 710         ControlFile.time = (pg_time_t) time(NULL);
 711         ControlFile.checkPoint = ControlFile.checkPointCopy.redo;
 712         ControlFile.prevCheckPoint = 0;
 713         ControlFile.minRecoveryPoint = 0;
 714         ControlFile.minRecoveryPointTLI = 0;
 715         ControlFile.backupStartPoint = 0;
 716         ControlFile.backupEndPoint = 0;
 717         ControlFile.backupEndRequired = false;
 718
 719         /*
 720          * Force the defaults for max_* settings. The values don't really matter
 721          * as long as wal_level='minimal'; the postmaster will reset these fields
 722          * anyway at startup.
 723          */
 724         ControlFile.wal_level = WAL_LEVEL_MINIMAL;
 725         ControlFile.wal_log_hintbits = false;
 726         ControlFile.MaxConnections = 100;
 727         ControlFile.max_worker_processes = 8;
 728         ControlFile.max_prepared_xacts = 0;
 729         ControlFile.max_locks_per_xact = 64;
 730
 731         /* Now we can force the recorded xlog seg size to the right thing. */
 732         ControlFile.xlog_seg_size = XLogSegSize;
 733
 734         /* Contents are protected with a CRC */
 735         INIT_CRC32(ControlFile.crc);
 736         COMP_CRC32(ControlFile.crc,
 737                            (char *) &ControlFile,
 738                            offsetof(ControlFileData, crc));
 739         FIN_CRC32(ControlFile.crc);
 740
 741         /*
 742          * We write out PG_CONTROL_SIZE bytes into pg_control, zero-padding the
 743          * excess over sizeof(ControlFileData).  This reduces the odds of
 744          * premature-EOF errors when reading pg_control.  We'll still fail when we
 745          * check the contents of the file, but hopefully with a more specific
 746          * error than "couldn't read pg_control".
 747          */
 748         if (sizeof(ControlFileData) > PG_CONTROL_SIZE)
 749         {
 750                 fprintf(stderr,
 751                                 _("%s: internal error -- sizeof(ControlFileData) is too large ... fix PG_CONTROL_SIZE\n"),
 752                                 progname);
 753                 exit(1);
 754         }
 755
 756         memset(buffer, 0, PG_CONTROL_SIZE);
 757         memcpy(buffer, &ControlFile, sizeof(ControlFileData));
 758
 759         unlink(XLOG_CONTROL_FILE);
 760
 761         fd = open(XLOG_CONTROL_FILE,
 762                           O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
 763                           S_IRUSR | S_IWUSR);
 764         if (fd < 0)
 765         {
 766                 fprintf(stderr, _("%s: could not create pg_control file: %s\n"),
 767                                 progname, strerror(errno));
 768                 exit(1);
 769         }
 770
 771         errno = 0;
 772         if (write(fd, buffer, PG_CONTROL_SIZE) != PG_CONTROL_SIZE)
 773         {
 774                 /* if write didn't set errno, assume problem is no disk space */
 775                 if (errno == 0)
 776                         errno = ENOSPC;
 777                 fprintf(stderr, _("%s: could not write pg_control file: %s\n"),
 778                                 progname, strerror(errno));
 779                 exit(1);
 780         }
 781
 782         if (fsync(fd) != 0)
 783         {
 784                 fprintf(stderr, _("%s: fsync error: %s\n"), progname, strerror(errno));
 785                 exit(1);
 786         }
 787
 788         close(fd);
 789 }
 790
 791
 792 /*
 793  * Scan existing XLOG files and determine the highest existing WAL address
 794  *
 795  * On entry, ControlFile.checkPointCopy.redo and ControlFile.xlog_seg_size
 796  * are assumed valid (note that we allow the old xlog seg size to differ
 797  * from what we're using).  On exit, newXlogId and newXlogSeg are set to
 798  * suitable values for the beginning of replacement WAL (in our seg size).
 799  */
 800 static void
 801 FindEndOfXLOG(void)
 802 {
 803         DIR                *xldir;
 804         struct dirent *xlde;
 805         uint64          segs_per_xlogid;
 806         uint64          xlogbytepos;
 807
 808         /*
 809          * Initialize the max() computation using the last checkpoint address from
 810          * old pg_control.      Note that for the moment we are working with segment
 811          * numbering according to the old xlog seg size.
 812          */
 813         segs_per_xlogid = (UINT64CONST(0x0000000100000000) / ControlFile.xlog_seg_size);
 814         newXlogSegNo = ControlFile.checkPointCopy.redo / ControlFile.xlog_seg_size;
 815
 816         /*
 817          * Scan the pg_xlog directory to find existing WAL segment files. We
 818          * assume any present have been used; in most scenarios this should be
 819          * conservative, because of xlog.c's attempts to pre-create files.
 820          */
 821         xldir = opendir(XLOGDIR);
 822         if (xldir == NULL)
 823         {
 824                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 825                                 progname, XLOGDIR, strerror(errno));
 826                 exit(1);
 827         }
 828
 829         errno = 0;
 830         while ((xlde = readdir(xldir)) != NULL)
 831         {
 832                 if (strlen(xlde->d_name) == 24 &&
 833                         strspn(xlde->d_name, "0123456789ABCDEF") == 24)
 834                 {
 835                         unsigned int tli,
 836                                                 log,
 837                                                 seg;
 838                         XLogSegNo       segno;
 839
 840                         sscanf(xlde->d_name, "%08X%08X%08X", &tli, &log, &seg);
 841                         segno = ((uint64) log) * segs_per_xlogid + seg;
 842
 843                         /*
 844                          * Note: we take the max of all files found, regardless of their
 845                          * timelines.  Another possibility would be to ignore files of
 846                          * timelines other than the target TLI, but this seems safer.
 847                          * Better too large a result than too small...
 848                          */
 849                         if (segno > newXlogSegNo)
 850                                 newXlogSegNo = segno;
 851                 }
 852                 errno = 0;
 853         }
 854 #ifdef WIN32
 855
 856         /*
 857          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 858          * released version
 859          */
 860         if (GetLastError() == ERROR_NO_MORE_FILES)
 861                 errno = 0;
 862 #endif
 863
 864         if (errno)
 865         {
 866                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 867                                 progname, XLOGDIR, strerror(errno));
 868                 exit(1);
 869         }
 870         closedir(xldir);
 871
 872         /*
 873          * Finally, convert to new xlog seg size, and advance by one to ensure we
 874          * are in virgin territory.
 875          */
 876         xlogbytepos = newXlogSegNo * ControlFile.xlog_seg_size;
 877         newXlogSegNo = (xlogbytepos + XLogSegSize - 1) / XLogSegSize;
 878         newXlogSegNo++;
 879 }
 880
 881
 882 /*
 883  * Remove existing XLOG files
 884  */
 885 static void
 886 KillExistingXLOG(void)
 887 {
 888         DIR                *xldir;
 889         struct dirent *xlde;
 890         char            path[MAXPGPATH];
 891
 892         xldir = opendir(XLOGDIR);
 893         if (xldir == NULL)
 894         {
 895                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 896                                 progname, XLOGDIR, strerror(errno));
 897                 exit(1);
 898         }
 899
 900         errno = 0;
 901         while ((xlde = readdir(xldir)) != NULL)
 902         {
 903                 if (strlen(xlde->d_name) == 24 &&
 904                         strspn(xlde->d_name, "0123456789ABCDEF") == 24)
 905                 {
 906                         snprintf(path, MAXPGPATH, "%s/%s", XLOGDIR, xlde->d_name);
 907                         if (unlink(path) < 0)
 908                         {
 909                                 fprintf(stderr, _("%s: could not delete file \"%s\": %s\n"),
 910                                                 progname, path, strerror(errno));
 911                                 exit(1);
 912                         }
 913                 }
 914                 errno = 0;
 915         }
 916 #ifdef WIN32
 917
 918         /*
 919          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 920          * released version
 921          */
 922         if (GetLastError() == ERROR_NO_MORE_FILES)
 923                 errno = 0;
 924 #endif
 925
 926         if (errno)
 927         {
 928                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 929                                 progname, XLOGDIR, strerror(errno));
 930                 exit(1);
 931         }
 932         closedir(xldir);
 933 }
 934
 935
 936 /*
 937  * Remove existing archive status files
 938  */
 939 static void
 940 KillExistingArchiveStatus(void)
 941 {
 942         DIR                *xldir;
 943         struct dirent *xlde;
 944         char            path[MAXPGPATH];
 945
 946 #define ARCHSTATDIR XLOGDIR "/archive_status"
 947
 948         xldir = opendir(ARCHSTATDIR);
 949         if (xldir == NULL)
 950         {
 951                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 952                                 progname, ARCHSTATDIR, strerror(errno));
 953                 exit(1);
 954         }
 955
 956         errno = 0;
 957         while ((xlde = readdir(xldir)) != NULL)
 958         {
 959                 if (strspn(xlde->d_name, "0123456789ABCDEF") == 24 &&
 960                         (strcmp(xlde->d_name + 24, ".ready") == 0 ||
 961                          strcmp(xlde->d_name + 24, ".done") == 0))
 962                 {
 963                         snprintf(path, MAXPGPATH, "%s/%s", ARCHSTATDIR, xlde->d_name);
 964                         if (unlink(path) < 0)
 965                         {
 966                                 fprintf(stderr, _("%s: could not delete file \"%s\": %s\n"),
 967                                                 progname, path, strerror(errno));
 968                                 exit(1);
 969                         }
 970                 }
 971                 errno = 0;
 972         }
 973 #ifdef WIN32
 974
 975         /*
 976          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 977          * released version
 978          */
 979         if (GetLastError() == ERROR_NO_MORE_FILES)
 980                 errno = 0;
 981 #endif
 982
 983         if (errno)
 984         {
 985                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 986                                 progname, ARCHSTATDIR, strerror(errno));
 987                 exit(1);
 988         }
 989         closedir(xldir);
 990 }
 991
 992
 993 /*
 994  * Write an empty XLOG file, containing only the checkpoint record
 995  * already set up in ControlFile.
 996  */
 997 static void
 998 WriteEmptyXLOG(void)
 999 {
1000         char       *buffer;
1001         XLogPageHeader page;
1002         XLogLongPageHeader longpage;
1003         XLogRecord *record;
1004         pg_crc32        crc;
1005         char            path[MAXPGPATH];
1006         int                     fd;
1007         int                     nbytes;
1008
1009         /* Use malloc() to ensure buffer is MAXALIGNED */
1010         buffer = (char *) pg_malloc(XLOG_BLCKSZ);
1011         page = (XLogPageHeader) buffer;
1012         memset(buffer, 0, XLOG_BLCKSZ);
1013
1014         /* Set up the XLOG page header */
1015         page->xlp_magic = XLOG_PAGE_MAGIC;
1016         page->xlp_info = XLP_LONG_HEADER;
1017         page->xlp_tli = ControlFile.checkPointCopy.ThisTimeLineID;
1018         page->xlp_pageaddr = ControlFile.checkPointCopy.redo - SizeOfXLogLongPHD;
1019         longpage = (XLogLongPageHeader) page;
1020         longpage->xlp_sysid = ControlFile.system_identifier;
1021         longpage->xlp_seg_size = XLogSegSize;
1022         longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
1023
1024         /* Insert the initial checkpoint record */
1025         record = (XLogRecord *) ((char *) page + SizeOfXLogLongPHD);
1026         record->xl_prev = 0;
1027         record->xl_xid = InvalidTransactionId;
1028         record->xl_tot_len = SizeOfXLogRecord + sizeof(CheckPoint);
1029         record->xl_len = sizeof(CheckPoint);
1030         record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
1031         record->xl_rmid = RM_XLOG_ID;
1032         memcpy(XLogRecGetData(record), &ControlFile.checkPointCopy,
1033                    sizeof(CheckPoint));
1034
1035         INIT_CRC32(crc);
1036         COMP_CRC32(crc, &ControlFile.checkPointCopy, sizeof(CheckPoint));
1037         COMP_CRC32(crc, (char *) record, offsetof(XLogRecord, xl_crc));
1038         FIN_CRC32(crc);
1039         record->xl_crc = crc;
1040
1041         /* Write the first page */
1042         XLogFilePath(path, ControlFile.checkPointCopy.ThisTimeLineID, newXlogSegNo);
1043
1044         unlink(path);
1045
1046         fd = open(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
1047                           S_IRUSR | S_IWUSR);
1048         if (fd < 0)
1049         {
1050                 fprintf(stderr, _("%s: could not open file \"%s\": %s\n"),
1051                                 progname, path, strerror(errno));
1052                 exit(1);
1053         }
1054
1055         errno = 0;
1056         if (write(fd, buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
1057         {
1058                 /* if write didn't set errno, assume problem is no disk space */
1059                 if (errno == 0)
1060                         errno = ENOSPC;
1061                 fprintf(stderr, _("%s: could not write file \"%s\": %s\n"),
1062                                 progname, path, strerror(errno));
1063                 exit(1);
1064         }
1065
1066         /* Fill the rest of the file with zeroes */
1067         memset(buffer, 0, XLOG_BLCKSZ);
1068         for (nbytes = XLOG_BLCKSZ; nbytes < XLogSegSize; nbytes += XLOG_BLCKSZ)
1069         {
1070                 errno = 0;
1071                 if (write(fd, buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
1072                 {
1073                         if (errno == 0)
1074                                 errno = ENOSPC;
1075                         fprintf(stderr, _("%s: could not write file \"%s\": %s\n"),
1076                                         progname, path, strerror(errno));
1077                         exit(1);
1078                 }
1079         }
1080
1081         if (fsync(fd) != 0)
1082         {
1083                 fprintf(stderr, _("%s: fsync error: %s\n"), progname, strerror(errno));
1084                 exit(1);
1085         }
1086
1087         close(fd);
1088 }
1089
1090
1091 static void
1092 usage(void)
1093 {
1094         printf(_("%s resets the PostgreSQL transaction log.\n\n"), progname);
1095         printf(_("Usage:\n  %s [OPTION]... DATADIR\n\n"), progname);
1096         printf(_("Options:\n"));
1097         printf(_("  -e XIDEPOCH      set next transaction ID epoch\n"));
1098         printf(_("  -f               force update to be done\n"));
1099         printf(_("  -l XLOGFILE      force minimum WAL starting location for new transaction log\n"));
1100         printf(_("  -m MXID,MXID     set next and oldest multitransaction ID\n"));
1101         printf(_("  -n               no update, just show what would be done (for testing)\n"));
1102         printf(_("  -o OID           set next OID\n"));
1103         printf(_("  -O OFFSET        set next multitransaction offset\n"));
1104         printf(_("  -V, --version    output version information, then exit\n"));
1105         printf(_("  -x XID           set next transaction ID\n"));
1106         printf(_("  -?, --help       show this help, then exit\n"));
1107         printf(_("\nReport bugs to <pgsql-bugs@postgresql.org>.\n"));
1108 }