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) 1996-2012, 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         char            path[MAXPGPATH];
  95
  96         set_pglocale_pgservice(argv[0], PG_TEXTDOMAIN("pg_resetxlog"));
  97
  98         progname = get_progname(argv[0]);
  99
 100         if (argc > 1)
 101         {
 102                 if (strcmp(argv[1], "--help") == 0 || strcmp(argv[1], "-?") == 0)
 103                 {
 104                         usage();
 105                         exit(0);
 106                 }
 107                 if (strcmp(argv[1], "--version") == 0 || strcmp(argv[1], "-V") == 0)
 108                 {
 109                         puts("pg_resetxlog (PostgreSQL) " PG_VERSION);
 110                         exit(0);
 111                 }
 112         }
 113
 114
 115         while ((c = getopt(argc, argv, "fl:m:no:O:x:e:")) != -1)
 116         {
 117                 switch (c)
 118                 {
 119                         case 'f':
 120                                 force = true;
 121                                 break;
 122
 123                         case 'n':
 124                                 noupdate = true;
 125                                 break;
 126
 127                         case 'e':
 128                                 set_xid_epoch = strtoul(optarg, &endptr, 0);
 129                                 if (endptr == optarg || *endptr != '\0')
 130                                 {
 131                                         fprintf(stderr, _("%s: invalid argument for option -e\n"), progname);
 132                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 133                                         exit(1);
 134                                 }
 135                                 if (set_xid_epoch == -1)
 136                                 {
 137                                         fprintf(stderr, _("%s: transaction ID epoch (-e) must not be -1\n"), progname);
 138                                         exit(1);
 139                                 }
 140                                 break;
 141
 142                         case 'x':
 143                                 set_xid = strtoul(optarg, &endptr, 0);
 144                                 if (endptr == optarg || *endptr != '\0')
 145                                 {
 146                                         fprintf(stderr, _("%s: invalid argument for option -x\n"), progname);
 147                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 148                                         exit(1);
 149                                 }
 150                                 if (set_xid == 0)
 151                                 {
 152                                         fprintf(stderr, _("%s: transaction ID (-x) must not be 0\n"), progname);
 153                                         exit(1);
 154                                 }
 155                                 break;
 156
 157                         case 'o':
 158                                 set_oid = strtoul(optarg, &endptr, 0);
 159                                 if (endptr == optarg || *endptr != '\0')
 160                                 {
 161                                         fprintf(stderr, _("%s: invalid argument for option -o\n"), progname);
 162                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 163                                         exit(1);
 164                                 }
 165                                 if (set_oid == 0)
 166                                 {
 167                                         fprintf(stderr, _("%s: OID (-o) must not be 0\n"), progname);
 168                                         exit(1);
 169                                 }
 170                                 break;
 171
 172                         case 'm':
 173                                 set_mxid = strtoul(optarg, &endptr, 0);
 174                                 if (endptr == optarg || *endptr != '\0')
 175                                 {
 176                                         fprintf(stderr, _("%s: invalid argument for option -m\n"), progname);
 177                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 178                                         exit(1);
 179                                 }
 180                                 if (set_mxid == 0)
 181                                 {
 182                                         fprintf(stderr, _("%s: multitransaction ID (-m) must not be 0\n"), progname);
 183                                         exit(1);
 184                                 }
 185                                 break;
 186
 187                         case 'O':
 188                                 set_mxoff = strtoul(optarg, &endptr, 0);
 189                                 if (endptr == optarg || *endptr != '\0')
 190                                 {
 191                                         fprintf(stderr, _("%s: invalid argument for option -O\n"), progname);
 192                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 193                                         exit(1);
 194                                 }
 195                                 if (set_mxoff == -1)
 196                                 {
 197                                         fprintf(stderr, _("%s: multitransaction offset (-O) must not be -1\n"), progname);
 198                                         exit(1);
 199                                 }
 200                                 break;
 201
 202                         case 'l':
 203                                 if (strspn(optarg, "01234567890ABCDEFabcdef") != 24)
 204                                 {
 205                                         fprintf(stderr, _("%s: invalid argument for option -l\n"), progname);
 206                                         fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 207                                         exit(1);
 208                                 }
 209                                 XLogFromFileName(optarg, &minXlogTli, &minXlogSegNo);
 210                                 break;
 211
 212                         default:
 213                                 fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 214                                 exit(1);
 215                 }
 216         }
 217
 218         if (optind == argc)
 219         {
 220                 fprintf(stderr, _("%s: no data directory specified\n"), progname);
 221                 fprintf(stderr, _("Try \"%s --help\" for more information.\n"), progname);
 222                 exit(1);
 223         }
 224
 225         /*
 226          * Don't allow pg_resetxlog to be run as root, to avoid overwriting the
 227          * ownership of files in the data directory. We need only check for root
 228          * -- any other user won't have sufficient permissions to modify files in
 229          * the data directory.
 230          */
 231 #ifndef WIN32
 232         if (geteuid() == 0)
 233         {
 234                 fprintf(stderr, _("%s: cannot be executed by \"root\"\n"),
 235                                 progname);
 236                 fprintf(stderr, _("You must run %s as the PostgreSQL superuser.\n"),
 237                                 progname);
 238                 exit(1);
 239         }
 240 #endif
 241
 242         DataDir = argv[optind];
 243
 244         if (chdir(DataDir) < 0)
 245         {
 246                 fprintf(stderr, _("%s: could not change directory to \"%s\": %s\n"),
 247                                 progname, DataDir, strerror(errno));
 248                 exit(1);
 249         }
 250
 251         /*
 252          * Check for a postmaster lock file --- if there is one, refuse to
 253          * proceed, on grounds we might be interfering with a live installation.
 254          */
 255         snprintf(path, MAXPGPATH, "%s/postmaster.pid", DataDir);
 256
 257         if ((fd = open(path, O_RDONLY, 0)) < 0)
 258         {
 259                 if (errno != ENOENT)
 260                 {
 261                         fprintf(stderr, _("%s: could not open file \"%s\" for reading: %s\n"), progname, path, strerror(errno));
 262                         exit(1);
 263                 }
 264         }
 265         else
 266         {
 267                 fprintf(stderr, _("%s: lock file \"%s\" exists\n"
 268                                                   "Is a server running?  If not, delete the lock file and try again.\n"),
 269                                 progname, path);
 270                 exit(1);
 271         }
 272
 273         /*
 274          * Attempt to read the existing pg_control file
 275          */
 276         if (!ReadControlFile())
 277                 GuessControlValues();
 278
 279         /*
 280          * Also look at existing segment files to set up newXlogSegNo
 281          */
 282         FindEndOfXLOG();
 283
 284         /*
 285          * Adjust fields if required by switches.  (Do this now so that printout,
 286          * if any, includes these values.)
 287          */
 288         if (set_xid_epoch != -1)
 289                 ControlFile.checkPointCopy.nextXidEpoch = set_xid_epoch;
 290
 291         if (set_xid != 0)
 292         {
 293                 ControlFile.checkPointCopy.nextXid = set_xid;
 294
 295                 /*
 296                  * For the moment, just set oldestXid to a value that will force
 297                  * immediate autovacuum-for-wraparound.  It's not clear whether adding
 298                  * user control of this is useful, so let's just do something that's
 299                  * reasonably safe.  The magic constant here corresponds to the
 300                  * maximum allowed value of autovacuum_freeze_max_age.
 301                  */
 302                 ControlFile.checkPointCopy.oldestXid = set_xid - 2000000000;
 303                 if (ControlFile.checkPointCopy.oldestXid < FirstNormalTransactionId)
 304                         ControlFile.checkPointCopy.oldestXid += FirstNormalTransactionId;
 305                 ControlFile.checkPointCopy.oldestXidDB = InvalidOid;
 306         }
 307
 308         if (set_oid != 0)
 309                 ControlFile.checkPointCopy.nextOid = set_oid;
 310
 311         if (set_mxid != 0)
 312                 ControlFile.checkPointCopy.nextMulti = set_mxid;
 313
 314         if (set_mxoff != -1)
 315                 ControlFile.checkPointCopy.nextMultiOffset = set_mxoff;
 316
 317         if (minXlogTli > ControlFile.checkPointCopy.ThisTimeLineID)
 318                 ControlFile.checkPointCopy.ThisTimeLineID = minXlogTli;
 319
 320         if (minXlogSegNo > newXlogSegNo)
 321                 newXlogSegNo = minXlogSegNo;
 322
 323         /*
 324          * If we had to guess anything, and -f was not given, just print the
 325          * guessed values and exit.  Also print if -n is given.
 326          */
 327         if ((guessed && !force) || noupdate)
 328         {
 329                 PrintControlValues(guessed);
 330                 if (!noupdate)
 331                 {
 332                         printf(_("\nIf these values seem acceptable, use -f to force reset.\n"));
 333                         exit(1);
 334                 }
 335                 else
 336                         exit(0);
 337         }
 338
 339         /*
 340          * Don't reset from a dirty pg_control without -f, either.
 341          */
 342         if (ControlFile.state != DB_SHUTDOWNED && !force)
 343         {
 344                 printf(_("The database server was not shut down cleanly.\n"
 345                            "Resetting the transaction log might cause data to be lost.\n"
 346                                  "If you want to proceed anyway, use -f to force reset.\n"));
 347                 exit(1);
 348         }
 349
 350         /*
 351          * Else, do the dirty deed.
 352          */
 353         RewriteControlFile();
 354         KillExistingXLOG();
 355         KillExistingArchiveStatus();
 356         WriteEmptyXLOG();
 357
 358         printf(_("Transaction log reset\n"));
 359         return 0;
 360 }
 361
 362
 363 /*
 364  * Try to read the existing pg_control file.
 365  *
 366  * This routine is also responsible for updating old pg_control versions
 367  * to the current format.  (Currently we don't do anything of the sort.)
 368  */
 369 static bool
 370 ReadControlFile(void)
 371 {
 372         int                     fd;
 373         int                     len;
 374         char       *buffer;
 375         pg_crc32        crc;
 376
 377         if ((fd = open(XLOG_CONTROL_FILE, O_RDONLY | PG_BINARY, 0)) < 0)
 378         {
 379                 /*
 380                  * If pg_control is not there at all, or we can't read it, the odds
 381                  * are we've been handed a bad DataDir path, so give up. User can do
 382                  * "touch pg_control" to force us to proceed.
 383                  */
 384                 fprintf(stderr, _("%s: could not open file \"%s\" for reading: %s\n"),
 385                                 progname, XLOG_CONTROL_FILE, strerror(errno));
 386                 if (errno == ENOENT)
 387                         fprintf(stderr, _("If you are sure the data directory path is correct, execute\n"
 388                                                           "  touch %s\n"
 389                                                           "and try again.\n"),
 390                                         XLOG_CONTROL_FILE);
 391                 exit(1);
 392         }
 393
 394         /* Use malloc to ensure we have a maxaligned buffer */
 395         buffer = (char *) malloc(PG_CONTROL_SIZE);
 396
 397         len = read(fd, buffer, PG_CONTROL_SIZE);
 398         if (len < 0)
 399         {
 400                 fprintf(stderr, _("%s: could not read file \"%s\": %s\n"),
 401                                 progname, XLOG_CONTROL_FILE, strerror(errno));
 402                 exit(1);
 403         }
 404         close(fd);
 405
 406         if (len >= sizeof(ControlFileData) &&
 407           ((ControlFileData *) buffer)->pg_control_version == PG_CONTROL_VERSION)
 408         {
 409                 /* Check the CRC. */
 410                 INIT_CRC32(crc);
 411                 COMP_CRC32(crc,
 412                                    buffer,
 413                                    offsetof(ControlFileData, crc));
 414                 FIN_CRC32(crc);
 415
 416                 if (EQ_CRC32(crc, ((ControlFileData *) buffer)->crc))
 417                 {
 418                         /* Valid data... */
 419                         memcpy(&ControlFile, buffer, sizeof(ControlFile));
 420                         return true;
 421                 }
 422
 423                 fprintf(stderr, _("%s: pg_control exists but has invalid CRC; proceed with caution\n"),
 424                                 progname);
 425                 /* We will use the data anyway, but treat it as guessed. */
 426                 memcpy(&ControlFile, buffer, sizeof(ControlFile));
 427                 guessed = true;
 428                 return true;
 429         }
 430
 431         /* Looks like it's a mess. */
 432         fprintf(stderr, _("%s: pg_control exists but is broken or unknown version; ignoring it\n"),
 433                         progname);
 434         return false;
 435 }
 436
 437
 438 /*
 439  * Guess at pg_control values when we can't read the old ones.
 440  */
 441 static void
 442 GuessControlValues(void)
 443 {
 444         uint64          sysidentifier;
 445         struct timeval tv;
 446
 447         /*
 448          * Set up a completely default set of pg_control values.
 449          */
 450         guessed = true;
 451         memset(&ControlFile, 0, sizeof(ControlFile));
 452
 453         ControlFile.pg_control_version = PG_CONTROL_VERSION;
 454         ControlFile.catalog_version_no = CATALOG_VERSION_NO;
 455
 456         /*
 457          * Create a new unique installation identifier, since we can no longer use
 458          * any old XLOG records.  See notes in xlog.c about the algorithm.
 459          */
 460         gettimeofday(&tv, NULL);
 461         sysidentifier = ((uint64) tv.tv_sec) << 32;
 462         sysidentifier |= (uint32) (tv.tv_sec | tv.tv_usec);
 463
 464         ControlFile.system_identifier = sysidentifier;
 465
 466         ControlFile.checkPointCopy.redo.xlogid = 0;
 467         ControlFile.checkPointCopy.redo.xrecoff = SizeOfXLogLongPHD;
 468         ControlFile.checkPointCopy.ThisTimeLineID = 1;
 469         ControlFile.checkPointCopy.fullPageWrites = false;
 470         ControlFile.checkPointCopy.nextXidEpoch = 0;
 471         ControlFile.checkPointCopy.nextXid = FirstNormalTransactionId;
 472         ControlFile.checkPointCopy.nextOid = FirstBootstrapObjectId;
 473         ControlFile.checkPointCopy.nextMulti = FirstMultiXactId;
 474         ControlFile.checkPointCopy.nextMultiOffset = 0;
 475         ControlFile.checkPointCopy.oldestXid = FirstNormalTransactionId;
 476         ControlFile.checkPointCopy.oldestXidDB = InvalidOid;
 477         ControlFile.checkPointCopy.time = (pg_time_t) time(NULL);
 478         ControlFile.checkPointCopy.oldestActiveXid = InvalidTransactionId;
 479
 480         ControlFile.state = DB_SHUTDOWNED;
 481         ControlFile.time = (pg_time_t) time(NULL);
 482         ControlFile.checkPoint = ControlFile.checkPointCopy.redo;
 483
 484         /* minRecoveryPoint, backupStartPoint and backupEndPoint can be left zero */
 485
 486         ControlFile.wal_level = WAL_LEVEL_MINIMAL;
 487         ControlFile.MaxConnections = 100;
 488         ControlFile.max_prepared_xacts = 0;
 489         ControlFile.max_locks_per_xact = 64;
 490
 491         ControlFile.maxAlign = MAXIMUM_ALIGNOF;
 492         ControlFile.floatFormat = FLOATFORMAT_VALUE;
 493         ControlFile.blcksz = BLCKSZ;
 494         ControlFile.relseg_size = RELSEG_SIZE;
 495         ControlFile.xlog_blcksz = XLOG_BLCKSZ;
 496         ControlFile.xlog_seg_size = XLOG_SEG_SIZE;
 497         ControlFile.nameDataLen = NAMEDATALEN;
 498         ControlFile.indexMaxKeys = INDEX_MAX_KEYS;
 499         ControlFile.toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
 500 #ifdef HAVE_INT64_TIMESTAMP
 501         ControlFile.enableIntTimes = true;
 502 #else
 503         ControlFile.enableIntTimes = false;
 504 #endif
 505         ControlFile.float4ByVal = FLOAT4PASSBYVAL;
 506         ControlFile.float8ByVal = FLOAT8PASSBYVAL;
 507
 508         /*
 509          * XXX eventually, should try to grovel through old XLOG to develop more
 510          * accurate values for TimeLineID, nextXID, etc.
 511          */
 512 }
 513
 514
 515 /*
 516  * Print the guessed pg_control values when we had to guess.
 517  *
 518  * NB: this display should be just those fields that will not be
 519  * reset by RewriteControlFile().
 520  */
 521 static void
 522 PrintControlValues(bool guessed)
 523 {
 524         char            sysident_str[32];
 525         char            fname[MAXFNAMELEN];
 526
 527         if (guessed)
 528                 printf(_("Guessed pg_control values:\n\n"));
 529         else
 530                 printf(_("pg_control values:\n\n"));
 531
 532         /*
 533          * Format system_identifier separately to keep platform-dependent format
 534          * code out of the translatable message string.
 535          */
 536         snprintf(sysident_str, sizeof(sysident_str), UINT64_FORMAT,
 537                          ControlFile.system_identifier);
 538
 539         XLogFileName(fname, ControlFile.checkPointCopy.ThisTimeLineID, newXlogSegNo);
 540
 541         printf(_("First log segment after reset:        %s\n"),
 542                    fname);
 543         printf(_("pg_control version number:            %u\n"),
 544                    ControlFile.pg_control_version);
 545         printf(_("Catalog version number:               %u\n"),
 546                    ControlFile.catalog_version_no);
 547         printf(_("Database system identifier:           %s\n"),
 548                    sysident_str);
 549         printf(_("Latest checkpoint's TimeLineID:       %u\n"),
 550                    ControlFile.checkPointCopy.ThisTimeLineID);
 551         printf(_("Latest checkpoint's full_page_writes: %s\n"),
 552                    ControlFile.checkPointCopy.fullPageWrites ? _("on") : _("off"));
 553         printf(_("Latest checkpoint's NextXID:          %u/%u\n"),
 554                    ControlFile.checkPointCopy.nextXidEpoch,
 555                    ControlFile.checkPointCopy.nextXid);
 556         printf(_("Latest checkpoint's NextOID:          %u\n"),
 557                    ControlFile.checkPointCopy.nextOid);
 558         printf(_("Latest checkpoint's NextMultiXactId:  %u\n"),
 559                    ControlFile.checkPointCopy.nextMulti);
 560         printf(_("Latest checkpoint's NextMultiOffset:  %u\n"),
 561                    ControlFile.checkPointCopy.nextMultiOffset);
 562         printf(_("Latest checkpoint's oldestXID:        %u\n"),
 563                    ControlFile.checkPointCopy.oldestXid);
 564         printf(_("Latest checkpoint's oldestXID's DB:   %u\n"),
 565                    ControlFile.checkPointCopy.oldestXidDB);
 566         printf(_("Latest checkpoint's oldestActiveXID:  %u\n"),
 567                    ControlFile.checkPointCopy.oldestActiveXid);
 568         printf(_("Maximum data alignment:               %u\n"),
 569                    ControlFile.maxAlign);
 570         /* we don't print floatFormat since can't say much useful about it */
 571         printf(_("Database block size:                  %u\n"),
 572                    ControlFile.blcksz);
 573         printf(_("Blocks per segment of large relation: %u\n"),
 574                    ControlFile.relseg_size);
 575         printf(_("WAL block size:                       %u\n"),
 576                    ControlFile.xlog_blcksz);
 577         printf(_("Bytes per WAL segment:                %u\n"),
 578                    ControlFile.xlog_seg_size);
 579         printf(_("Maximum length of identifiers:        %u\n"),
 580                    ControlFile.nameDataLen);
 581         printf(_("Maximum columns in an index:          %u\n"),
 582                    ControlFile.indexMaxKeys);
 583         printf(_("Maximum size of a TOAST chunk:        %u\n"),
 584                    ControlFile.toast_max_chunk_size);
 585         printf(_("Date/time type storage:               %s\n"),
 586                    (ControlFile.enableIntTimes ? _("64-bit integers") : _("floating-point numbers")));
 587         printf(_("Float4 argument passing:              %s\n"),
 588                    (ControlFile.float4ByVal ? _("by value") : _("by reference")));
 589         printf(_("Float8 argument passing:              %s\n"),
 590                    (ControlFile.float8ByVal ? _("by value") : _("by reference")));
 591 }
 592
 593
 594 /*
 595  * Write out the new pg_control file.
 596  */
 597 static void
 598 RewriteControlFile(void)
 599 {
 600         int                     fd;
 601         char            buffer[PG_CONTROL_SIZE];                /* need not be aligned */
 602
 603         /*
 604          * Adjust fields as needed to force an empty XLOG starting at
 605          * newXlogSegNo.
 606          */
 607         XLogSegNoOffsetToRecPtr(newXlogSegNo, SizeOfXLogLongPHD,
 608                                                         ControlFile.checkPointCopy.redo);
 609         ControlFile.checkPointCopy.time = (pg_time_t) time(NULL);
 610
 611         ControlFile.state = DB_SHUTDOWNED;
 612         ControlFile.time = (pg_time_t) time(NULL);
 613         ControlFile.checkPoint = ControlFile.checkPointCopy.redo;
 614         ControlFile.prevCheckPoint.xlogid = 0;
 615         ControlFile.prevCheckPoint.xrecoff = 0;
 616         ControlFile.minRecoveryPoint.xlogid = 0;
 617         ControlFile.minRecoveryPoint.xrecoff = 0;
 618         ControlFile.backupStartPoint.xlogid = 0;
 619         ControlFile.backupStartPoint.xrecoff = 0;
 620         ControlFile.backupEndPoint.xlogid = 0;
 621         ControlFile.backupEndPoint.xrecoff = 0;
 622         ControlFile.backupEndRequired = false;
 623
 624         /*
 625          * Force the defaults for max_* settings. The values don't really matter
 626          * as long as wal_level='minimal'; the postmaster will reset these fields
 627          * anyway at startup.
 628          */
 629         ControlFile.wal_level = WAL_LEVEL_MINIMAL;
 630         ControlFile.MaxConnections = 100;
 631         ControlFile.max_prepared_xacts = 0;
 632         ControlFile.max_locks_per_xact = 64;
 633
 634         /* Now we can force the recorded xlog seg size to the right thing. */
 635         ControlFile.xlog_seg_size = XLogSegSize;
 636
 637         /* Contents are protected with a CRC */
 638         INIT_CRC32(ControlFile.crc);
 639         COMP_CRC32(ControlFile.crc,
 640                            (char *) &ControlFile,
 641                            offsetof(ControlFileData, crc));
 642         FIN_CRC32(ControlFile.crc);
 643
 644         /*
 645          * We write out PG_CONTROL_SIZE bytes into pg_control, zero-padding the
 646          * excess over sizeof(ControlFileData).  This reduces the odds of
 647          * premature-EOF errors when reading pg_control.  We'll still fail when we
 648          * check the contents of the file, but hopefully with a more specific
 649          * error than "couldn't read pg_control".
 650          */
 651         if (sizeof(ControlFileData) > PG_CONTROL_SIZE)
 652         {
 653                 fprintf(stderr,
 654                                 _("%s: internal error -- sizeof(ControlFileData) is too large ... fix PG_CONTROL_SIZE\n"),
 655                                 progname);
 656                 exit(1);
 657         }
 658
 659         memset(buffer, 0, PG_CONTROL_SIZE);
 660         memcpy(buffer, &ControlFile, sizeof(ControlFileData));
 661
 662         unlink(XLOG_CONTROL_FILE);
 663
 664         fd = open(XLOG_CONTROL_FILE,
 665                           O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
 666                           S_IRUSR | S_IWUSR);
 667         if (fd < 0)
 668         {
 669                 fprintf(stderr, _("%s: could not create pg_control file: %s\n"),
 670                                 progname, strerror(errno));
 671                 exit(1);
 672         }
 673
 674         errno = 0;
 675         if (write(fd, buffer, PG_CONTROL_SIZE) != PG_CONTROL_SIZE)
 676         {
 677                 /* if write didn't set errno, assume problem is no disk space */
 678                 if (errno == 0)
 679                         errno = ENOSPC;
 680                 fprintf(stderr, _("%s: could not write pg_control file: %s\n"),
 681                                 progname, strerror(errno));
 682                 exit(1);
 683         }
 684
 685         if (fsync(fd) != 0)
 686         {
 687                 fprintf(stderr, _("%s: fsync error: %s\n"), progname, strerror(errno));
 688                 exit(1);
 689         }
 690
 691         close(fd);
 692 }
 693
 694
 695 /*
 696  * Scan existing XLOG files and determine the highest existing WAL address
 697  *
 698  * On entry, ControlFile.checkPointCopy.redo and ControlFile.xlog_seg_size
 699  * are assumed valid (note that we allow the old xlog seg size to differ
 700  * from what we're using).  On exit, newXlogId and newXlogSeg are set to
 701  * suitable values for the beginning of replacement WAL (in our seg size).
 702  */
 703 static void
 704 FindEndOfXLOG(void)
 705 {
 706         DIR                *xldir;
 707         struct dirent *xlde;
 708         uint64          segs_per_xlogid;
 709         uint64          xlogbytepos;
 710
 711         /*
 712          * Initialize the max() computation using the last checkpoint address from
 713          * old pg_control.      Note that for the moment we are working with segment
 714          * numbering according to the old xlog seg size.
 715          */
 716         segs_per_xlogid = (0x100000000L / ControlFile.xlog_seg_size);
 717         newXlogSegNo = ((uint64) ControlFile.checkPointCopy.redo.xlogid) * segs_per_xlogid
 718                 + (ControlFile.checkPointCopy.redo.xrecoff / ControlFile.xlog_seg_size);
 719
 720         /*
 721          * Scan the pg_xlog directory to find existing WAL segment files. We
 722          * assume any present have been used; in most scenarios this should be
 723          * conservative, because of xlog.c's attempts to pre-create files.
 724          */
 725         xldir = opendir(XLOGDIR);
 726         if (xldir == NULL)
 727         {
 728                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 729                                 progname, XLOGDIR, strerror(errno));
 730                 exit(1);
 731         }
 732
 733         errno = 0;
 734         while ((xlde = readdir(xldir)) != NULL)
 735         {
 736                 if (strlen(xlde->d_name) == 24 &&
 737                         strspn(xlde->d_name, "0123456789ABCDEF") == 24)
 738                 {
 739                         unsigned int tli,
 740                                                 log,
 741                                                 seg;
 742                         XLogSegNo       segno;
 743
 744                         sscanf(xlde->d_name, "%08X%08X%08X", &tli, &log, &seg);
 745                         segno = ((uint64) log) * segs_per_xlogid + seg;
 746
 747                         /*
 748                          * Note: we take the max of all files found, regardless of their
 749                          * timelines.  Another possibility would be to ignore files of
 750                          * timelines other than the target TLI, but this seems safer.
 751                          * Better too large a result than too small...
 752                          */
 753                         if (segno > newXlogSegNo)
 754                                 newXlogSegNo = segno;
 755                 }
 756                 errno = 0;
 757         }
 758 #ifdef WIN32
 759
 760         /*
 761          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 762          * released version
 763          */
 764         if (GetLastError() == ERROR_NO_MORE_FILES)
 765                 errno = 0;
 766 #endif
 767
 768         if (errno)
 769         {
 770                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 771                                 progname, XLOGDIR, strerror(errno));
 772                 exit(1);
 773         }
 774         closedir(xldir);
 775
 776         /*
 777          * Finally, convert to new xlog seg size, and advance by one to ensure we
 778          * are in virgin territory.
 779          */
 780         xlogbytepos = newXlogSegNo * ControlFile.xlog_seg_size;
 781         newXlogSegNo = (xlogbytepos + XLogSegSize - 1) / XLogSegSize;
 782         newXlogSegNo++;
 783 }
 784
 785
 786 /*
 787  * Remove existing XLOG files
 788  */
 789 static void
 790 KillExistingXLOG(void)
 791 {
 792         DIR                *xldir;
 793         struct dirent *xlde;
 794         char            path[MAXPGPATH];
 795
 796         xldir = opendir(XLOGDIR);
 797         if (xldir == NULL)
 798         {
 799                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 800                                 progname, XLOGDIR, strerror(errno));
 801                 exit(1);
 802         }
 803
 804         errno = 0;
 805         while ((xlde = readdir(xldir)) != NULL)
 806         {
 807                 if (strlen(xlde->d_name) == 24 &&
 808                         strspn(xlde->d_name, "0123456789ABCDEF") == 24)
 809                 {
 810                         snprintf(path, MAXPGPATH, "%s/%s", XLOGDIR, xlde->d_name);
 811                         if (unlink(path) < 0)
 812                         {
 813                                 fprintf(stderr, _("%s: could not delete file \"%s\": %s\n"),
 814                                                 progname, path, strerror(errno));
 815                                 exit(1);
 816                         }
 817                 }
 818                 errno = 0;
 819         }
 820 #ifdef WIN32
 821
 822         /*
 823          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 824          * released version
 825          */
 826         if (GetLastError() == ERROR_NO_MORE_FILES)
 827                 errno = 0;
 828 #endif
 829
 830         if (errno)
 831         {
 832                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 833                                 progname, XLOGDIR, strerror(errno));
 834                 exit(1);
 835         }
 836         closedir(xldir);
 837 }
 838
 839
 840 /*
 841  * Remove existing archive status files
 842  */
 843 static void
 844 KillExistingArchiveStatus(void)
 845 {
 846         DIR                *xldir;
 847         struct dirent *xlde;
 848         char            path[MAXPGPATH];
 849
 850 #define ARCHSTATDIR XLOGDIR "/archive_status"
 851
 852         xldir = opendir(ARCHSTATDIR);
 853         if (xldir == NULL)
 854         {
 855                 fprintf(stderr, _("%s: could not open directory \"%s\": %s\n"),
 856                                 progname, ARCHSTATDIR, strerror(errno));
 857                 exit(1);
 858         }
 859
 860         errno = 0;
 861         while ((xlde = readdir(xldir)) != NULL)
 862         {
 863                 if (strspn(xlde->d_name, "0123456789ABCDEF") == 24 &&
 864                         (strcmp(xlde->d_name + 24, ".ready") == 0 ||
 865                          strcmp(xlde->d_name + 24, ".done") == 0))
 866                 {
 867                         snprintf(path, MAXPGPATH, "%s/%s", ARCHSTATDIR, xlde->d_name);
 868                         if (unlink(path) < 0)
 869                         {
 870                                 fprintf(stderr, _("%s: could not delete file \"%s\": %s\n"),
 871                                                 progname, path, strerror(errno));
 872                                 exit(1);
 873                         }
 874                 }
 875                 errno = 0;
 876         }
 877 #ifdef WIN32
 878
 879         /*
 880          * This fix is in mingw cvs (runtime/mingwex/dirent.c rev 1.4), but not in
 881          * released version
 882          */
 883         if (GetLastError() == ERROR_NO_MORE_FILES)
 884                 errno = 0;
 885 #endif
 886
 887         if (errno)
 888         {
 889                 fprintf(stderr, _("%s: could not read from directory \"%s\": %s\n"),
 890                                 progname, ARCHSTATDIR, strerror(errno));
 891                 exit(1);
 892         }
 893         closedir(xldir);
 894 }
 895
 896
 897 /*
 898  * Write an empty XLOG file, containing only the checkpoint record
 899  * already set up in ControlFile.
 900  */
 901 static void
 902 WriteEmptyXLOG(void)
 903 {
 904         char       *buffer;
 905         XLogPageHeader page;
 906         XLogLongPageHeader longpage;
 907         XLogRecord *record;
 908         pg_crc32        crc;
 909         char            path[MAXPGPATH];
 910         int                     fd;
 911         int                     nbytes;
 912
 913         /* Use malloc() to ensure buffer is MAXALIGNED */
 914         buffer = (char *) malloc(XLOG_BLCKSZ);
 915         page = (XLogPageHeader) buffer;
 916         memset(buffer, 0, XLOG_BLCKSZ);
 917
 918         /* Set up the XLOG page header */
 919         page->xlp_magic = XLOG_PAGE_MAGIC;
 920         page->xlp_info = XLP_LONG_HEADER;
 921         page->xlp_tli = ControlFile.checkPointCopy.ThisTimeLineID;
 922         page->xlp_pageaddr.xlogid =
 923                 ControlFile.checkPointCopy.redo.xlogid;
 924         page->xlp_pageaddr.xrecoff =
 925                 ControlFile.checkPointCopy.redo.xrecoff - SizeOfXLogLongPHD;
 926         longpage = (XLogLongPageHeader) page;
 927         longpage->xlp_sysid = ControlFile.system_identifier;
 928         longpage->xlp_seg_size = XLogSegSize;
 929         longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
 930
 931         /* Insert the initial checkpoint record */
 932         record = (XLogRecord *) ((char *) page + SizeOfXLogLongPHD);
 933         record->xl_prev.xlogid = 0;
 934         record->xl_prev.xrecoff = 0;
 935         record->xl_xid = InvalidTransactionId;
 936         record->xl_tot_len = SizeOfXLogRecord + sizeof(CheckPoint);
 937         record->xl_len = sizeof(CheckPoint);
 938         record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
 939         record->xl_rmid = RM_XLOG_ID;
 940         memcpy(XLogRecGetData(record), &ControlFile.checkPointCopy,
 941                    sizeof(CheckPoint));
 942
 943         INIT_CRC32(crc);
 944         COMP_CRC32(crc, &ControlFile.checkPointCopy, sizeof(CheckPoint));
 945         COMP_CRC32(crc, (char *) record + sizeof(pg_crc32),
 946                            SizeOfXLogRecord - sizeof(pg_crc32));
 947         FIN_CRC32(crc);
 948         record->xl_crc = crc;
 949
 950         /* Write the first page */
 951         XLogFilePath(path, ControlFile.checkPointCopy.ThisTimeLineID, newXlogSegNo);
 952
 953         unlink(path);
 954
 955         fd = open(path, O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
 956                           S_IRUSR | S_IWUSR);
 957         if (fd < 0)
 958         {
 959                 fprintf(stderr, _("%s: could not open file \"%s\": %s\n"),
 960                                 progname, path, strerror(errno));
 961                 exit(1);
 962         }
 963
 964         errno = 0;
 965         if (write(fd, buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
 966         {
 967                 /* if write didn't set errno, assume problem is no disk space */
 968                 if (errno == 0)
 969                         errno = ENOSPC;
 970                 fprintf(stderr, _("%s: could not write file \"%s\": %s\n"),
 971                                 progname, path, strerror(errno));
 972                 exit(1);
 973         }
 974
 975         /* Fill the rest of the file with zeroes */
 976         memset(buffer, 0, XLOG_BLCKSZ);
 977         for (nbytes = XLOG_BLCKSZ; nbytes < XLogSegSize; nbytes += XLOG_BLCKSZ)
 978         {
 979                 errno = 0;
 980                 if (write(fd, buffer, XLOG_BLCKSZ) != XLOG_BLCKSZ)
 981                 {
 982                         if (errno == 0)
 983                                 errno = ENOSPC;
 984                         fprintf(stderr, _("%s: could not write file \"%s\": %s\n"),
 985                                         progname, path, strerror(errno));
 986                         exit(1);
 987                 }
 988         }
 989
 990         if (fsync(fd) != 0)
 991         {
 992                 fprintf(stderr, _("%s: fsync error: %s\n"), progname, strerror(errno));
 993                 exit(1);
 994         }
 995
 996         close(fd);
 997 }
 998
 999
1000 static void
1001 usage(void)
1002 {
1003         printf(_("%s resets the PostgreSQL transaction log.\n\n"), progname);
1004         printf(_("Usage:\n  %s [OPTION]... DATADIR\n\n"), progname);
1005         printf(_("Options:\n"));
1006         printf(_("  -e XIDEPOCH      set next transaction ID epoch\n"));
1007         printf(_("  -f               force update to be done\n"));
1008         printf(_("  -l TLI,FILE,SEG  force minimum WAL starting location for new transaction log\n"));
1009         printf(_("  -m XID           set next multitransaction ID\n"));
1010         printf(_("  -n               no update, just show extracted control values (for testing)\n"));
1011         printf(_("  -o OID           set next OID\n"));
1012         printf(_("  -O OFFSET        set next multitransaction offset\n"));
1013         printf(_("  -V, --version    output version information, then exit\n"));
1014         printf(_("  -x XID           set next transaction ID\n"));
1015         printf(_("  -?, --help       show this help, then exit\n"));
1016         printf(_("\nReport bugs to <pgsql-bugs@postgresql.org>.\n"));
1017 }