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