src/backend/access/transam/xlog.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * xlog.c
   4  *              PostgreSQL transaction log manager
   5  *
   6  *
   7  * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
   8  * Portions Copyright (c) 1994, Regents of the University of California
   9  *
  10  * $PostgreSQL: pgsql/src/backend/access/transam/xlog.c,v 1.299 2008/04/23 13:44:58 mha Exp $
  11  *
  12  *-------------------------------------------------------------------------
  13  */
  14
  15 #include "postgres.h"
  16
  17 #include <ctype.h>
  18 #include <fcntl.h>
  19 #include <signal.h>
  20 #include <time.h>
  21 #include <sys/stat.h>
  22 #include <sys/time.h>
  23 #include <sys/wait.h>
  24 #include <unistd.h>
  25
  26 #include "access/clog.h"
  27 #include "access/heapam.h"
  28 #include "access/multixact.h"
  29 #include "access/subtrans.h"
  30 #include "access/transam.h"
  31 #include "access/tuptoaster.h"
  32 #include "access/twophase.h"
  33 #include "access/xact.h"
  34 #include "access/xlog_internal.h"
  35 #include "access/xlogdefs.h"
  36 #include "access/xlogutils.h"
  37 #include "catalog/catversion.h"
  38 #include "catalog/pg_control.h"
  39 #include "catalog/pg_type.h"
  40 #include "funcapi.h"
  41 #include "miscadmin.h"
  42 #include "pgstat.h"
  43 #include "postmaster/bgwriter.h"
  44 #include "storage/bufpage.h"
  45 #include "storage/fd.h"
  46 #include "storage/ipc.h"
  47 #include "storage/pmsignal.h"
  48 #include "storage/procarray.h"
  49 #include "storage/smgr.h"
  50 #include "storage/spin.h"
  51 #include "utils/builtins.h"
  52 #include "utils/pg_locale.h"
  53 #include "utils/ps_status.h"
  54
  55
  56 /* File path names (all relative to $PGDATA) */
  57 #define BACKUP_LABEL_FILE               "backup_label"
  58 #define BACKUP_LABEL_OLD                "backup_label.old"
  59 #define RECOVERY_COMMAND_FILE   "recovery.conf"
  60 #define RECOVERY_COMMAND_DONE   "recovery.done"
  61
  62
  63 /* User-settable parameters */
  64 int                     CheckPointSegments = 3;
  65 int                     XLOGbuffers = 8;
  66 int                     XLogArchiveTimeout = 0;
  67 bool            XLogArchiveMode = false;
  68 char       *XLogArchiveCommand = NULL;
  69 char       *XLOG_sync_method = NULL;
  70 const char      XLOG_sync_method_default[] = DEFAULT_SYNC_METHOD_STR;
  71 bool            fullPageWrites = true;
  72 bool            log_checkpoints = false;
  73
  74 #ifdef WAL_DEBUG
  75 bool            XLOG_DEBUG = false;
  76 #endif
  77
  78 /*
  79  * XLOGfileslop is the maximum number of preallocated future XLOG segments.
  80  * When we are done with an old XLOG segment file, we will recycle it as a
  81  * future XLOG segment as long as there aren't already XLOGfileslop future
  82  * segments; else we'll delete it.  This could be made a separate GUC
  83  * variable, but at present I think it's sufficient to hardwire it as
  84  * 2*CheckPointSegments+1.      Under normal conditions, a checkpoint will free
  85  * no more than 2*CheckPointSegments log segments, and we want to recycle all
  86  * of them; the +1 allows boundary cases to happen without wasting a
  87  * delete/create-segment cycle.
  88  */
  89 #define XLOGfileslop    (2*CheckPointSegments + 1)
  90
  91
  92 /* these are derived from XLOG_sync_method by assign_xlog_sync_method */
  93 int                     sync_method = DEFAULT_SYNC_METHOD;
  94 static int      open_sync_bit = DEFAULT_SYNC_FLAGBIT;
  95
  96 #define XLOG_SYNC_BIT  (enableFsync ? open_sync_bit : 0)
  97
  98
  99 /*
 100  * Statistics for current checkpoint are collected in this global struct.
 101  * Because only the background writer or a stand-alone backend can perform
 102  * checkpoints, this will be unused in normal backends.
 103  */
 104 CheckpointStatsData CheckpointStats;
 105
 106 /*
 107  * ThisTimeLineID will be same in all backends --- it identifies current
 108  * WAL timeline for the database system.
 109  */
 110 TimeLineID      ThisTimeLineID = 0;
 111
 112 /* Are we doing recovery from XLOG? */
 113 bool            InRecovery = false;
 114
 115 /* Are we recovering using offline XLOG archives? */
 116 static bool InArchiveRecovery = false;
 117
 118 /* Was the last xlog file restored from archive, or local? */
 119 static bool restoredFromArchive = false;
 120
 121 /* options taken from recovery.conf */
 122 static char *recoveryRestoreCommand = NULL;
 123 static bool recoveryTarget = false;
 124 static bool recoveryTargetExact = false;
 125 static bool recoveryTargetInclusive = true;
 126 static bool recoveryLogRestartpoints = false;
 127 static TransactionId recoveryTargetXid;
 128 static TimestampTz recoveryTargetTime;
 129 static TimestampTz recoveryLastXTime = 0;
 130
 131 /* if recoveryStopsHere returns true, it saves actual stop xid/time here */
 132 static TransactionId recoveryStopXid;
 133 static TimestampTz recoveryStopTime;
 134 static bool recoveryStopAfter;
 135
 136 /*
 137  * During normal operation, the only timeline we care about is ThisTimeLineID.
 138  * During recovery, however, things are more complicated.  To simplify life
 139  * for rmgr code, we keep ThisTimeLineID set to the "current" timeline as we
 140  * scan through the WAL history (that is, it is the line that was active when
 141  * the currently-scanned WAL record was generated).  We also need these
 142  * timeline values:
 143  *
 144  * recoveryTargetTLI: the desired timeline that we want to end in.
 145  *
 146  * expectedTLIs: an integer list of recoveryTargetTLI and the TLIs of
 147  * its known parents, newest first (so recoveryTargetTLI is always the
 148  * first list member).  Only these TLIs are expected to be seen in the WAL
 149  * segments we read, and indeed only these TLIs will be considered as
 150  * candidate WAL files to open at all.
 151  *
 152  * curFileTLI: the TLI appearing in the name of the current input WAL file.
 153  * (This is not necessarily the same as ThisTimeLineID, because we could
 154  * be scanning data that was copied from an ancestor timeline when the current
 155  * file was created.)  During a sequential scan we do not allow this value
 156  * to decrease.
 157  */
 158 static TimeLineID recoveryTargetTLI;
 159 static List *expectedTLIs;
 160 static TimeLineID curFileTLI;
 161
 162 /*
 163  * ProcLastRecPtr points to the start of the last XLOG record inserted by the
 164  * current backend.  It is updated for all inserts.  XactLastRecEnd points to
 165  * end+1 of the last record, and is reset when we end a top-level transaction,
 166  * or start a new one; so it can be used to tell if the current transaction has
 167  * created any XLOG records.
 168  */
 169 static XLogRecPtr ProcLastRecPtr = {0, 0};
 170
 171 XLogRecPtr      XactLastRecEnd = {0, 0};
 172
 173 /*
 174  * RedoRecPtr is this backend's local copy of the REDO record pointer
 175  * (which is almost but not quite the same as a pointer to the most recent
 176  * CHECKPOINT record).  We update this from the shared-memory copy,
 177  * XLogCtl->Insert.RedoRecPtr, whenever we can safely do so (ie, when we
 178  * hold the Insert lock).  See XLogInsert for details.  We are also allowed
 179  * to update from XLogCtl->Insert.RedoRecPtr if we hold the info_lck;
 180  * see GetRedoRecPtr.  A freshly spawned backend obtains the value during
 181  * InitXLOGAccess.
 182  */
 183 static XLogRecPtr RedoRecPtr;
 184
 185 /*----------
 186  * Shared-memory data structures for XLOG control
 187  *
 188  * LogwrtRqst indicates a byte position that we need to write and/or fsync
 189  * the log up to (all records before that point must be written or fsynced).
 190  * LogwrtResult indicates the byte positions we have already written/fsynced.
 191  * These structs are identical but are declared separately to indicate their
 192  * slightly different functions.
 193  *
 194  * We do a lot of pushups to minimize the amount of access to lockable
 195  * shared memory values.  There are actually three shared-memory copies of
 196  * LogwrtResult, plus one unshared copy in each backend.  Here's how it works:
 197  *              XLogCtl->LogwrtResult is protected by info_lck
 198  *              XLogCtl->Write.LogwrtResult is protected by WALWriteLock
 199  *              XLogCtl->Insert.LogwrtResult is protected by WALInsertLock
 200  * One must hold the associated lock to read or write any of these, but
 201  * of course no lock is needed to read/write the unshared LogwrtResult.
 202  *
 203  * XLogCtl->LogwrtResult and XLogCtl->Write.LogwrtResult are both "always
 204  * right", since both are updated by a write or flush operation before
 205  * it releases WALWriteLock.  The point of keeping XLogCtl->Write.LogwrtResult
 206  * is that it can be examined/modified by code that already holds WALWriteLock
 207  * without needing to grab info_lck as well.
 208  *
 209  * XLogCtl->Insert.LogwrtResult may lag behind the reality of the other two,
 210  * but is updated when convenient.      Again, it exists for the convenience of
 211  * code that is already holding WALInsertLock but not the other locks.
 212  *
 213  * The unshared LogwrtResult may lag behind any or all of these, and again
 214  * is updated when convenient.
 215  *
 216  * The request bookkeeping is simpler: there is a shared XLogCtl->LogwrtRqst
 217  * (protected by info_lck), but we don't need to cache any copies of it.
 218  *
 219  * Note that this all works because the request and result positions can only
 220  * advance forward, never back up, and so we can easily determine which of two
 221  * values is "more up to date".
 222  *
 223  * info_lck is only held long enough to read/update the protected variables,
 224  * so it's a plain spinlock.  The other locks are held longer (potentially
 225  * over I/O operations), so we use LWLocks for them.  These locks are:
 226  *
 227  * WALInsertLock: must be held to insert a record into the WAL buffers.
 228  *
 229  * WALWriteLock: must be held to write WAL buffers to disk (XLogWrite or
 230  * XLogFlush).
 231  *
 232  * ControlFileLock: must be held to read/update control file or create
 233  * new log file.
 234  *
 235  * CheckpointLock: must be held to do a checkpoint (ensures only one
 236  * checkpointer at a time; currently, with all checkpoints done by the
 237  * bgwriter, this is just pro forma).
 238  *
 239  *----------
 240  */
 241
 242 typedef struct XLogwrtRqst
 243 {
 244         XLogRecPtr      Write;                  /* last byte + 1 to write out */
 245         XLogRecPtr      Flush;                  /* last byte + 1 to flush */
 246 } XLogwrtRqst;
 247
 248 typedef struct XLogwrtResult
 249 {
 250         XLogRecPtr      Write;                  /* last byte + 1 written out */
 251         XLogRecPtr      Flush;                  /* last byte + 1 flushed */
 252 } XLogwrtResult;
 253
 254 /*
 255  * Shared state data for XLogInsert.
 256  */
 257 typedef struct XLogCtlInsert
 258 {
 259         XLogwrtResult LogwrtResult; /* a recent value of LogwrtResult */
 260         XLogRecPtr      PrevRecord;             /* start of previously-inserted record */
 261         int                     curridx;                /* current block index in cache */
 262         XLogPageHeader currpage;        /* points to header of block in cache */
 263         char       *currpos;            /* current insertion point in cache */
 264         XLogRecPtr      RedoRecPtr;             /* current redo point for insertions */
 265         bool            forcePageWrites;        /* forcing full-page writes for PITR? */
 266 } XLogCtlInsert;
 267
 268 /*
 269  * Shared state data for XLogWrite/XLogFlush.
 270  */
 271 typedef struct XLogCtlWrite
 272 {
 273         XLogwrtResult LogwrtResult; /* current value of LogwrtResult */
 274         int                     curridx;                /* cache index of next block to write */
 275         pg_time_t       lastSegSwitchTime;              /* time of last xlog segment switch */
 276 } XLogCtlWrite;
 277
 278 /*
 279  * Total shared-memory state for XLOG.
 280  */
 281 typedef struct XLogCtlData
 282 {
 283         /* Protected by WALInsertLock: */
 284         XLogCtlInsert Insert;
 285
 286         /* Protected by info_lck: */
 287         XLogwrtRqst LogwrtRqst;
 288         XLogwrtResult LogwrtResult;
 289         uint32          ckptXidEpoch;   /* nextXID & epoch of latest checkpoint */
 290         TransactionId ckptXid;
 291         XLogRecPtr      asyncCommitLSN; /* LSN of newest async commit */
 292
 293         /* Protected by WALWriteLock: */
 294         XLogCtlWrite Write;
 295
 296         /*
 297          * These values do not change after startup, although the pointed-to pages
 298          * and xlblocks values certainly do.  Permission to read/write the pages
 299          * and xlblocks values depends on WALInsertLock and WALWriteLock.
 300          */
 301         char       *pages;                      /* buffers for unwritten XLOG pages */
 302         XLogRecPtr *xlblocks;           /* 1st byte ptr-s + XLOG_BLCKSZ */
 303         int                     XLogCacheBlck;  /* highest allocated xlog buffer index */
 304         TimeLineID      ThisTimeLineID;
 305
 306         slock_t         info_lck;               /* locks shared variables shown above */
 307 } XLogCtlData;
 308
 309 static XLogCtlData *XLogCtl = NULL;
 310
 311 /*
 312  * We maintain an image of pg_control in shared memory.
 313  */
 314 static ControlFileData *ControlFile = NULL;
 315
 316 /*
 317  * Macros for managing XLogInsert state.  In most cases, the calling routine
 318  * has local copies of XLogCtl->Insert and/or XLogCtl->Insert->curridx,
 319  * so these are passed as parameters instead of being fetched via XLogCtl.
 320  */
 321
 322 /* Free space remaining in the current xlog page buffer */
 323 #define INSERT_FREESPACE(Insert)  \
 324         (XLOG_BLCKSZ - ((Insert)->currpos - (char *) (Insert)->currpage))
 325
 326 /* Construct XLogRecPtr value for current insertion point */
 327 #define INSERT_RECPTR(recptr,Insert,curridx)  \
 328         ( \
 329           (recptr).xlogid = XLogCtl->xlblocks[curridx].xlogid, \
 330           (recptr).xrecoff = \
 331                 XLogCtl->xlblocks[curridx].xrecoff - INSERT_FREESPACE(Insert) \
 332         )
 333
 334 #define PrevBufIdx(idx)         \
 335                 (((idx) == 0) ? XLogCtl->XLogCacheBlck : ((idx) - 1))
 336
 337 #define NextBufIdx(idx)         \
 338                 (((idx) == XLogCtl->XLogCacheBlck) ? 0 : ((idx) + 1))
 339
 340 /*
 341  * Private, possibly out-of-date copy of shared LogwrtResult.
 342  * See discussion above.
 343  */
 344 static XLogwrtResult LogwrtResult = {{0, 0}, {0, 0}};
 345
 346 /*
 347  * openLogFile is -1 or a kernel FD for an open log file segment.
 348  * When it's open, openLogOff is the current seek offset in the file.
 349  * openLogId/openLogSeg identify the segment.  These variables are only
 350  * used to write the XLOG, and so will normally refer to the active segment.
 351  */
 352 static int      openLogFile = -1;
 353 static uint32 openLogId = 0;
 354 static uint32 openLogSeg = 0;
 355 static uint32 openLogOff = 0;
 356
 357 /*
 358  * These variables are used similarly to the ones above, but for reading
 359  * the XLOG.  Note, however, that readOff generally represents the offset
 360  * of the page just read, not the seek position of the FD itself, which
 361  * will be just past that page.
 362  */
 363 static int      readFile = -1;
 364 static uint32 readId = 0;
 365 static uint32 readSeg = 0;
 366 static uint32 readOff = 0;
 367
 368 /* Buffer for currently read page (XLOG_BLCKSZ bytes) */
 369 static char *readBuf = NULL;
 370
 371 /* Buffer for current ReadRecord result (expandable) */
 372 static char *readRecordBuf = NULL;
 373 static uint32 readRecordBufSize = 0;
 374
 375 /* State information for XLOG reading */
 376 static XLogRecPtr ReadRecPtr;   /* start of last record read */
 377 static XLogRecPtr EndRecPtr;    /* end+1 of last record read */
 378 static XLogRecord *nextRecord = NULL;
 379 static TimeLineID lastPageTLI = 0;
 380
 381 static bool InRedo = false;
 382
 383
 384 static void XLogArchiveNotify(const char *xlog);
 385 static void XLogArchiveNotifySeg(uint32 log, uint32 seg);
 386 static bool XLogArchiveCheckDone(const char *xlog, bool create_if_missing);
 387 static void XLogArchiveCleanup(const char *xlog);
 388 static void readRecoveryCommandFile(void);
 389 static void exitArchiveRecovery(TimeLineID endTLI,
 390                                         uint32 endLogId, uint32 endLogSeg);
 391 static bool recoveryStopsHere(XLogRecord *record, bool *includeThis);
 392 static void CheckPointGuts(XLogRecPtr checkPointRedo, int flags);
 393
 394 static bool XLogCheckBuffer(XLogRecData *rdata, bool doPageWrites,
 395                                 XLogRecPtr *lsn, BkpBlock *bkpb);
 396 static bool AdvanceXLInsertBuffer(bool new_segment);
 397 static void XLogWrite(XLogwrtRqst WriteRqst, bool flexible, bool xlog_switch);
 398 static int XLogFileInit(uint32 log, uint32 seg,
 399                          bool *use_existent, bool use_lock);
 400 static bool InstallXLogFileSegment(uint32 *log, uint32 *seg, char *tmppath,
 401                                            bool find_free, int *max_advance,
 402                                            bool use_lock);
 403 static int      XLogFileOpen(uint32 log, uint32 seg);
 404 static int      XLogFileRead(uint32 log, uint32 seg, int emode);
 405 static void XLogFileClose(void);
 406 static bool RestoreArchivedFile(char *path, const char *xlogfname,
 407                                         const char *recovername, off_t expectedSize);
 408 static void PreallocXlogFiles(XLogRecPtr endptr);
 409 static void RemoveOldXlogFiles(uint32 log, uint32 seg, XLogRecPtr endptr);
 410 static void CleanupBackupHistory(void);
 411 static XLogRecord *ReadRecord(XLogRecPtr *RecPtr, int emode);
 412 static bool ValidXLOGHeader(XLogPageHeader hdr, int emode);
 413 static XLogRecord *ReadCheckpointRecord(XLogRecPtr RecPtr, int whichChkpt);
 414 static List *readTimeLineHistory(TimeLineID targetTLI);
 415 static bool existsTimeLineHistory(TimeLineID probeTLI);
 416 static TimeLineID findNewestTimeLine(TimeLineID startTLI);
 417 static void writeTimeLineHistory(TimeLineID newTLI, TimeLineID parentTLI,
 418                                          TimeLineID endTLI,
 419                                          uint32 endLogId, uint32 endLogSeg);
 420 static void WriteControlFile(void);
 421 static void ReadControlFile(void);
 422 static char *str_time(pg_time_t tnow);
 423 #ifdef WAL_DEBUG
 424 static void xlog_outrec(StringInfo buf, XLogRecord *record);
 425 #endif
 426 static void issue_xlog_fsync(void);
 427 static void pg_start_backup_callback(int code, Datum arg);
 428 static bool read_backup_label(XLogRecPtr *checkPointLoc,
 429                                   XLogRecPtr *minRecoveryLoc);
 430 static void rm_redo_error_callback(void *arg);
 431
 432
 433 /*
 434  * Insert an XLOG record having the specified RMID and info bytes,
 435  * with the body of the record being the data chunk(s) described by
 436  * the rdata chain (see xlog.h for notes about rdata).
 437  *
 438  * Returns XLOG pointer to end of record (beginning of next record).
 439  * This can be used as LSN for data pages affected by the logged action.
 440  * (LSN is the XLOG point up to which the XLOG must be flushed to disk
 441  * before the data page can be written out.  This implements the basic
 442  * WAL rule "write the log before the data".)
 443  *
 444  * NB: this routine feels free to scribble on the XLogRecData structs,
 445  * though not on the data they reference.  This is OK since the XLogRecData
 446  * structs are always just temporaries in the calling code.
 447  */
 448 XLogRecPtr
 449 XLogInsert(RmgrId rmid, uint8 info, XLogRecData *rdata)
 450 {
 451         XLogCtlInsert *Insert = &XLogCtl->Insert;
 452         XLogRecord *record;
 453         XLogContRecord *contrecord;
 454         XLogRecPtr      RecPtr;
 455         XLogRecPtr      WriteRqst;
 456         uint32          freespace;
 457         int                     curridx;
 458         XLogRecData *rdt;
 459         Buffer          dtbuf[XLR_MAX_BKP_BLOCKS];
 460         bool            dtbuf_bkp[XLR_MAX_BKP_BLOCKS];
 461         BkpBlock        dtbuf_xlg[XLR_MAX_BKP_BLOCKS];
 462         XLogRecPtr      dtbuf_lsn[XLR_MAX_BKP_BLOCKS];
 463         XLogRecData dtbuf_rdt1[XLR_MAX_BKP_BLOCKS];
 464         XLogRecData dtbuf_rdt2[XLR_MAX_BKP_BLOCKS];
 465         XLogRecData dtbuf_rdt3[XLR_MAX_BKP_BLOCKS];
 466         pg_crc32        rdata_crc;
 467         uint32          len,
 468                                 write_len;
 469         unsigned        i;
 470         bool            updrqst;
 471         bool            doPageWrites;
 472         bool            isLogSwitch = (rmid == RM_XLOG_ID && info == XLOG_SWITCH);
 473
 474         /* info's high bits are reserved for use by me */
 475         if (info & XLR_INFO_MASK)
 476                 elog(PANIC, "invalid xlog info mask %02X", info);
 477
 478         /*
 479          * In bootstrap mode, we don't actually log anything but XLOG resources;
 480          * return a phony record pointer.
 481          */
 482         if (IsBootstrapProcessingMode() && rmid != RM_XLOG_ID)
 483         {
 484                 RecPtr.xlogid = 0;
 485                 RecPtr.xrecoff = SizeOfXLogLongPHD;             /* start of 1st chkpt record */
 486                 return RecPtr;
 487         }
 488
 489         /*
 490          * Here we scan the rdata chain, determine which buffers must be backed
 491          * up, and compute the CRC values for the data.  Note that the record
 492          * header isn't added into the CRC initially since we don't know the final
 493          * length or info bits quite yet.  Thus, the CRC will represent the CRC of
 494          * the whole record in the order "rdata, then backup blocks, then record
 495          * header".
 496          *
 497          * We may have to loop back to here if a race condition is detected below.
 498          * We could prevent the race by doing all this work while holding the
 499          * insert lock, but it seems better to avoid doing CRC calculations while
 500          * holding the lock.  This means we have to be careful about modifying the
 501          * rdata chain until we know we aren't going to loop back again.  The only
 502          * change we allow ourselves to make earlier is to set rdt->data = NULL in
 503          * chain items we have decided we will have to back up the whole buffer
 504          * for.  This is OK because we will certainly decide the same thing again
 505          * for those items if we do it over; doing it here saves an extra pass
 506          * over the chain later.
 507          */
 508 begin:;
 509         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
 510         {
 511                 dtbuf[i] = InvalidBuffer;
 512                 dtbuf_bkp[i] = false;
 513         }
 514
 515         /*
 516          * Decide if we need to do full-page writes in this XLOG record: true if
 517          * full_page_writes is on or we have a PITR request for it.  Since we
 518          * don't yet have the insert lock, forcePageWrites could change under us,
 519          * but we'll recheck it once we have the lock.
 520          */
 521         doPageWrites = fullPageWrites || Insert->forcePageWrites;
 522
 523         INIT_CRC32(rdata_crc);
 524         len = 0;
 525         for (rdt = rdata;;)
 526         {
 527                 if (rdt->buffer == InvalidBuffer)
 528                 {
 529                         /* Simple data, just include it */
 530                         len += rdt->len;
 531                         COMP_CRC32(rdata_crc, rdt->data, rdt->len);
 532                 }
 533                 else
 534                 {
 535                         /* Find info for buffer */
 536                         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
 537                         {
 538                                 if (rdt->buffer == dtbuf[i])
 539                                 {
 540                                         /* Buffer already referenced by earlier chain item */
 541                                         if (dtbuf_bkp[i])
 542                                                 rdt->data = NULL;
 543                                         else if (rdt->data)
 544                                         {
 545                                                 len += rdt->len;
 546                                                 COMP_CRC32(rdata_crc, rdt->data, rdt->len);
 547                                         }
 548                                         break;
 549                                 }
 550                                 if (dtbuf[i] == InvalidBuffer)
 551                                 {
 552                                         /* OK, put it in this slot */
 553                                         dtbuf[i] = rdt->buffer;
 554                                         if (XLogCheckBuffer(rdt, doPageWrites,
 555                                                                                 &(dtbuf_lsn[i]), &(dtbuf_xlg[i])))
 556                                         {
 557                                                 dtbuf_bkp[i] = true;
 558                                                 rdt->data = NULL;
 559                                         }
 560                                         else if (rdt->data)
 561                                         {
 562                                                 len += rdt->len;
 563                                                 COMP_CRC32(rdata_crc, rdt->data, rdt->len);
 564                                         }
 565                                         break;
 566                                 }
 567                         }
 568                         if (i >= XLR_MAX_BKP_BLOCKS)
 569                                 elog(PANIC, "can backup at most %d blocks per xlog record",
 570                                          XLR_MAX_BKP_BLOCKS);
 571                 }
 572                 /* Break out of loop when rdt points to last chain item */
 573                 if (rdt->next == NULL)
 574                         break;
 575                 rdt = rdt->next;
 576         }
 577
 578         /*
 579          * Now add the backup block headers and data into the CRC
 580          */
 581         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
 582         {
 583                 if (dtbuf_bkp[i])
 584                 {
 585                         BkpBlock   *bkpb = &(dtbuf_xlg[i]);
 586                         char       *page;
 587
 588                         COMP_CRC32(rdata_crc,
 589                                            (char *) bkpb,
 590                                            sizeof(BkpBlock));
 591                         page = (char *) BufferGetBlock(dtbuf[i]);
 592                         if (bkpb->hole_length == 0)
 593                         {
 594                                 COMP_CRC32(rdata_crc,
 595                                                    page,
 596                                                    BLCKSZ);
 597                         }
 598                         else
 599                         {
 600                                 /* must skip the hole */
 601                                 COMP_CRC32(rdata_crc,
 602                                                    page,
 603                                                    bkpb->hole_offset);
 604                                 COMP_CRC32(rdata_crc,
 605                                                    page + (bkpb->hole_offset + bkpb->hole_length),
 606                                                    BLCKSZ - (bkpb->hole_offset + bkpb->hole_length));
 607                         }
 608                 }
 609         }
 610
 611         /*
 612          * NOTE: We disallow len == 0 because it provides a useful bit of extra
 613          * error checking in ReadRecord.  This means that all callers of
 614          * XLogInsert must supply at least some not-in-a-buffer data.  However, we
 615          * make an exception for XLOG SWITCH records because we don't want them to
 616          * ever cross a segment boundary.
 617          */
 618         if (len == 0 && !isLogSwitch)
 619                 elog(PANIC, "invalid xlog record length %u", len);
 620
 621         START_CRIT_SECTION();
 622
 623         /* Now wait to get insert lock */
 624         LWLockAcquire(WALInsertLock, LW_EXCLUSIVE);
 625
 626         /*
 627          * Check to see if my RedoRecPtr is out of date.  If so, may have to go
 628          * back and recompute everything.  This can only happen just after a
 629          * checkpoint, so it's better to be slow in this case and fast otherwise.
 630          *
 631          * If we aren't doing full-page writes then RedoRecPtr doesn't actually
 632          * affect the contents of the XLOG record, so we'll update our local copy
 633          * but not force a recomputation.
 634          */
 635         if (!XLByteEQ(RedoRecPtr, Insert->RedoRecPtr))
 636         {
 637                 Assert(XLByteLT(RedoRecPtr, Insert->RedoRecPtr));
 638                 RedoRecPtr = Insert->RedoRecPtr;
 639
 640                 if (doPageWrites)
 641                 {
 642                         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
 643                         {
 644                                 if (dtbuf[i] == InvalidBuffer)
 645                                         continue;
 646                                 if (dtbuf_bkp[i] == false &&
 647                                         XLByteLE(dtbuf_lsn[i], RedoRecPtr))
 648                                 {
 649                                         /*
 650                                          * Oops, this buffer now needs to be backed up, but we
 651                                          * didn't think so above.  Start over.
 652                                          */
 653                                         LWLockRelease(WALInsertLock);
 654                                         END_CRIT_SECTION();
 655                                         goto begin;
 656                                 }
 657                         }
 658                 }
 659         }
 660
 661         /*
 662          * Also check to see if forcePageWrites was just turned on; if we weren't
 663          * already doing full-page writes then go back and recompute. (If it was
 664          * just turned off, we could recompute the record without full pages, but
 665          * we choose not to bother.)
 666          */
 667         if (Insert->forcePageWrites && !doPageWrites)
 668         {
 669                 /* Oops, must redo it with full-page data */
 670                 LWLockRelease(WALInsertLock);
 671                 END_CRIT_SECTION();
 672                 goto begin;
 673         }
 674
 675         /*
 676          * Make additional rdata chain entries for the backup blocks, so that we
 677          * don't need to special-case them in the write loop.  Note that we have
 678          * now irrevocably changed the input rdata chain.  At the exit of this
 679          * loop, write_len includes the backup block data.
 680          *
 681          * Also set the appropriate info bits to show which buffers were backed
 682          * up. The i'th XLR_SET_BKP_BLOCK bit corresponds to the i'th distinct
 683          * buffer value (ignoring InvalidBuffer) appearing in the rdata chain.
 684          */
 685         write_len = len;
 686         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
 687         {
 688                 BkpBlock   *bkpb;
 689                 char       *page;
 690
 691                 if (!dtbuf_bkp[i])
 692                         continue;
 693
 694                 info |= XLR_SET_BKP_BLOCK(i);
 695
 696                 bkpb = &(dtbuf_xlg[i]);
 697                 page = (char *) BufferGetBlock(dtbuf[i]);
 698
 699                 rdt->next = &(dtbuf_rdt1[i]);
 700                 rdt = rdt->next;
 701
 702                 rdt->data = (char *) bkpb;
 703                 rdt->len = sizeof(BkpBlock);
 704                 write_len += sizeof(BkpBlock);
 705
 706                 rdt->next = &(dtbuf_rdt2[i]);
 707                 rdt = rdt->next;
 708
 709                 if (bkpb->hole_length == 0)
 710                 {
 711                         rdt->data = page;
 712                         rdt->len = BLCKSZ;
 713                         write_len += BLCKSZ;
 714                         rdt->next = NULL;
 715                 }
 716                 else
 717                 {
 718                         /* must skip the hole */
 719                         rdt->data = page;
 720                         rdt->len = bkpb->hole_offset;
 721                         write_len += bkpb->hole_offset;
 722
 723                         rdt->next = &(dtbuf_rdt3[i]);
 724                         rdt = rdt->next;
 725
 726                         rdt->data = page + (bkpb->hole_offset + bkpb->hole_length);
 727                         rdt->len = BLCKSZ - (bkpb->hole_offset + bkpb->hole_length);
 728                         write_len += rdt->len;
 729                         rdt->next = NULL;
 730                 }
 731         }
 732
 733         /*
 734          * If we backed up any full blocks and online backup is not in progress,
 735          * mark the backup blocks as removable.  This allows the WAL archiver to
 736          * know whether it is safe to compress archived WAL data by transforming
 737          * full-block records into the non-full-block format.
 738          *
 739          * Note: we could just set the flag whenever !forcePageWrites, but
 740          * defining it like this leaves the info bit free for some potential other
 741          * use in records without any backup blocks.
 742          */
 743         if ((info & XLR_BKP_BLOCK_MASK) && !Insert->forcePageWrites)
 744                 info |= XLR_BKP_REMOVABLE;
 745
 746         /*
 747          * If there isn't enough space on the current XLOG page for a record
 748          * header, advance to the next page (leaving the unused space as zeroes).
 749          */
 750         updrqst = false;
 751         freespace = INSERT_FREESPACE(Insert);
 752         if (freespace < SizeOfXLogRecord)
 753         {
 754                 updrqst = AdvanceXLInsertBuffer(false);
 755                 freespace = INSERT_FREESPACE(Insert);
 756         }
 757
 758         /* Compute record's XLOG location */
 759         curridx = Insert->curridx;
 760         INSERT_RECPTR(RecPtr, Insert, curridx);
 761
 762         /*
 763          * If the record is an XLOG_SWITCH, and we are exactly at the start of a
 764          * segment, we need not insert it (and don't want to because we'd like
 765          * consecutive switch requests to be no-ops).  Instead, make sure
 766          * everything is written and flushed through the end of the prior segment,
 767          * and return the prior segment's end address.
 768          */
 769         if (isLogSwitch &&
 770                 (RecPtr.xrecoff % XLogSegSize) == SizeOfXLogLongPHD)
 771         {
 772                 /* We can release insert lock immediately */
 773                 LWLockRelease(WALInsertLock);
 774
 775                 RecPtr.xrecoff -= SizeOfXLogLongPHD;
 776                 if (RecPtr.xrecoff == 0)
 777                 {
 778                         /* crossing a logid boundary */
 779                         RecPtr.xlogid -= 1;
 780                         RecPtr.xrecoff = XLogFileSize;
 781                 }
 782
 783                 LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
 784                 LogwrtResult = XLogCtl->Write.LogwrtResult;
 785                 if (!XLByteLE(RecPtr, LogwrtResult.Flush))
 786                 {
 787                         XLogwrtRqst FlushRqst;
 788
 789                         FlushRqst.Write = RecPtr;
 790                         FlushRqst.Flush = RecPtr;
 791                         XLogWrite(FlushRqst, false, false);
 792                 }
 793                 LWLockRelease(WALWriteLock);
 794
 795                 END_CRIT_SECTION();
 796
 797                 return RecPtr;
 798         }
 799
 800         /* Insert record header */
 801
 802         record = (XLogRecord *) Insert->currpos;
 803         record->xl_prev = Insert->PrevRecord;
 804         record->xl_xid = GetCurrentTransactionIdIfAny();
 805         record->xl_tot_len = SizeOfXLogRecord + write_len;
 806         record->xl_len = len;           /* doesn't include backup blocks */
 807         record->xl_info = info;
 808         record->xl_rmid = rmid;
 809
 810         /* Now we can finish computing the record's CRC */
 811         COMP_CRC32(rdata_crc, (char *) record + sizeof(pg_crc32),
 812                            SizeOfXLogRecord - sizeof(pg_crc32));
 813         FIN_CRC32(rdata_crc);
 814         record->xl_crc = rdata_crc;
 815
 816 #ifdef WAL_DEBUG
 817         if (XLOG_DEBUG)
 818         {
 819                 StringInfoData buf;
 820
 821                 initStringInfo(&buf);
 822                 appendStringInfo(&buf, "INSERT @ %X/%X: ",
 823                                                  RecPtr.xlogid, RecPtr.xrecoff);
 824                 xlog_outrec(&buf, record);
 825                 if (rdata->data != NULL)
 826                 {
 827                         appendStringInfo(&buf, " - ");
 828                         RmgrTable[record->xl_rmid].rm_desc(&buf, record->xl_info, rdata->data);
 829                 }
 830                 elog(LOG, "%s", buf.data);
 831                 pfree(buf.data);
 832         }
 833 #endif
 834
 835         /* Record begin of record in appropriate places */
 836         ProcLastRecPtr = RecPtr;
 837         Insert->PrevRecord = RecPtr;
 838
 839         Insert->currpos += SizeOfXLogRecord;
 840         freespace -= SizeOfXLogRecord;
 841
 842         /*
 843          * Append the data, including backup blocks if any
 844          */
 845         while (write_len)
 846         {
 847                 while (rdata->data == NULL)
 848                         rdata = rdata->next;
 849
 850                 if (freespace > 0)
 851                 {
 852                         if (rdata->len > freespace)
 853                         {
 854                                 memcpy(Insert->currpos, rdata->data, freespace);
 855                                 rdata->data += freespace;
 856                                 rdata->len -= freespace;
 857                                 write_len -= freespace;
 858                         }
 859                         else
 860                         {
 861                                 memcpy(Insert->currpos, rdata->data, rdata->len);
 862                                 freespace -= rdata->len;
 863                                 write_len -= rdata->len;
 864                                 Insert->currpos += rdata->len;
 865                                 rdata = rdata->next;
 866                                 continue;
 867                         }
 868                 }
 869
 870                 /* Use next buffer */
 871                 updrqst = AdvanceXLInsertBuffer(false);
 872                 curridx = Insert->curridx;
 873                 /* Insert cont-record header */
 874                 Insert->currpage->xlp_info |= XLP_FIRST_IS_CONTRECORD;
 875                 contrecord = (XLogContRecord *) Insert->currpos;
 876                 contrecord->xl_rem_len = write_len;
 877                 Insert->currpos += SizeOfXLogContRecord;
 878                 freespace = INSERT_FREESPACE(Insert);
 879         }
 880
 881         /* Ensure next record will be properly aligned */
 882         Insert->currpos = (char *) Insert->currpage +
 883                 MAXALIGN(Insert->currpos - (char *) Insert->currpage);
 884         freespace = INSERT_FREESPACE(Insert);
 885
 886         /*
 887          * The recptr I return is the beginning of the *next* record. This will be
 888          * stored as LSN for changed data pages...
 889          */
 890         INSERT_RECPTR(RecPtr, Insert, curridx);
 891
 892         /*
 893          * If the record is an XLOG_SWITCH, we must now write and flush all the
 894          * existing data, and then forcibly advance to the start of the next
 895          * segment.  It's not good to do this I/O while holding the insert lock,
 896          * but there seems too much risk of confusion if we try to release the
 897          * lock sooner.  Fortunately xlog switch needn't be a high-performance
 898          * operation anyway...
 899          */
 900         if (isLogSwitch)
 901         {
 902                 XLogCtlWrite *Write = &XLogCtl->Write;
 903                 XLogwrtRqst FlushRqst;
 904                 XLogRecPtr      OldSegEnd;
 905
 906                 LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
 907
 908                 /*
 909                  * Flush through the end of the page containing XLOG_SWITCH, and
 910                  * perform end-of-segment actions (eg, notifying archiver).
 911                  */
 912                 WriteRqst = XLogCtl->xlblocks[curridx];
 913                 FlushRqst.Write = WriteRqst;
 914                 FlushRqst.Flush = WriteRqst;
 915                 XLogWrite(FlushRqst, false, true);
 916
 917                 /* Set up the next buffer as first page of next segment */
 918                 /* Note: AdvanceXLInsertBuffer cannot need to do I/O here */
 919                 (void) AdvanceXLInsertBuffer(true);
 920
 921                 /* There should be no unwritten data */
 922                 curridx = Insert->curridx;
 923                 Assert(curridx == Write->curridx);
 924
 925                 /* Compute end address of old segment */
 926                 OldSegEnd = XLogCtl->xlblocks[curridx];
 927                 OldSegEnd.xrecoff -= XLOG_BLCKSZ;
 928                 if (OldSegEnd.xrecoff == 0)
 929                 {
 930                         /* crossing a logid boundary */
 931                         OldSegEnd.xlogid -= 1;
 932                         OldSegEnd.xrecoff = XLogFileSize;
 933                 }
 934
 935                 /* Make it look like we've written and synced all of old segment */
 936                 LogwrtResult.Write = OldSegEnd;
 937                 LogwrtResult.Flush = OldSegEnd;
 938
 939                 /*
 940                  * Update shared-memory status --- this code should match XLogWrite
 941                  */
 942                 {
 943                         /* use volatile pointer to prevent code rearrangement */
 944                         volatile XLogCtlData *xlogctl = XLogCtl;
 945
 946                         SpinLockAcquire(&xlogctl->info_lck);
 947                         xlogctl->LogwrtResult = LogwrtResult;
 948                         if (XLByteLT(xlogctl->LogwrtRqst.Write, LogwrtResult.Write))
 949                                 xlogctl->LogwrtRqst.Write = LogwrtResult.Write;
 950                         if (XLByteLT(xlogctl->LogwrtRqst.Flush, LogwrtResult.Flush))
 951                                 xlogctl->LogwrtRqst.Flush = LogwrtResult.Flush;
 952                         SpinLockRelease(&xlogctl->info_lck);
 953                 }
 954
 955                 Write->LogwrtResult = LogwrtResult;
 956
 957                 LWLockRelease(WALWriteLock);
 958
 959                 updrqst = false;                /* done already */
 960         }
 961         else
 962         {
 963                 /* normal case, ie not xlog switch */
 964
 965                 /* Need to update shared LogwrtRqst if some block was filled up */
 966                 if (freespace < SizeOfXLogRecord)
 967                 {
 968                         /* curridx is filled and available for writing out */
 969                         updrqst = true;
 970                 }
 971                 else
 972                 {
 973                         /* if updrqst already set, write through end of previous buf */
 974                         curridx = PrevBufIdx(curridx);
 975                 }
 976                 WriteRqst = XLogCtl->xlblocks[curridx];
 977         }
 978
 979         LWLockRelease(WALInsertLock);
 980
 981         if (updrqst)
 982         {
 983                 /* use volatile pointer to prevent code rearrangement */
 984                 volatile XLogCtlData *xlogctl = XLogCtl;
 985
 986                 SpinLockAcquire(&xlogctl->info_lck);
 987                 /* advance global request to include new block(s) */
 988                 if (XLByteLT(xlogctl->LogwrtRqst.Write, WriteRqst))
 989                         xlogctl->LogwrtRqst.Write = WriteRqst;
 990                 /* update local result copy while I have the chance */
 991                 LogwrtResult = xlogctl->LogwrtResult;
 992                 SpinLockRelease(&xlogctl->info_lck);
 993         }
 994
 995         XactLastRecEnd = RecPtr;
 996
 997         END_CRIT_SECTION();
 998
 999         return RecPtr;
1000 }
1001
1002 /*
1003  * Determine whether the buffer referenced by an XLogRecData item has to
1004  * be backed up, and if so fill a BkpBlock struct for it.  In any case
1005  * save the buffer's LSN at *lsn.
1006  */
1007 static bool
1008 XLogCheckBuffer(XLogRecData *rdata, bool doPageWrites,
1009                                 XLogRecPtr *lsn, BkpBlock *bkpb)
1010 {
1011         PageHeader      page;
1012
1013         page = (PageHeader) BufferGetBlock(rdata->buffer);
1014
1015         /*
1016          * XXX We assume page LSN is first data on *every* page that can be passed
1017          * to XLogInsert, whether it otherwise has the standard page layout or
1018          * not.
1019          */
1020         *lsn = page->pd_lsn;
1021
1022         if (doPageWrites &&
1023                 XLByteLE(page->pd_lsn, RedoRecPtr))
1024         {
1025                 /*
1026                  * The page needs to be backed up, so set up *bkpb
1027                  */
1028                 bkpb->node = BufferGetFileNode(rdata->buffer);
1029                 bkpb->block = BufferGetBlockNumber(rdata->buffer);
1030
1031                 if (rdata->buffer_std)
1032                 {
1033                         /* Assume we can omit data between pd_lower and pd_upper */
1034                         uint16          lower = page->pd_lower;
1035                         uint16          upper = page->pd_upper;
1036
1037                         if (lower >= SizeOfPageHeaderData &&
1038                                 upper > lower &&
1039                                 upper <= BLCKSZ)
1040                         {
1041                                 bkpb->hole_offset = lower;
1042                                 bkpb->hole_length = upper - lower;
1043                         }
1044                         else
1045                         {
1046                                 /* No "hole" to compress out */
1047                                 bkpb->hole_offset = 0;
1048                                 bkpb->hole_length = 0;
1049                         }
1050                 }
1051                 else
1052                 {
1053                         /* Not a standard page header, don't try to eliminate "hole" */
1054                         bkpb->hole_offset = 0;
1055                         bkpb->hole_length = 0;
1056                 }
1057
1058                 return true;                    /* buffer requires backup */
1059         }
1060
1061         return false;                           /* buffer does not need to be backed up */
1062 }
1063
1064 /*
1065  * XLogArchiveNotify
1066  *
1067  * Create an archive notification file
1068  *
1069  * The name of the notification file is the message that will be picked up
1070  * by the archiver, e.g. we write 0000000100000001000000C6.ready
1071  * and the archiver then knows to archive XLOGDIR/0000000100000001000000C6,
1072  * then when complete, rename it to 0000000100000001000000C6.done
1073  */
1074 static void
1075 XLogArchiveNotify(const char *xlog)
1076 {
1077         char            archiveStatusPath[MAXPGPATH];
1078         FILE       *fd;
1079
1080         /* insert an otherwise empty file called <XLOG>.ready */
1081         StatusFilePath(archiveStatusPath, xlog, ".ready");
1082         fd = AllocateFile(archiveStatusPath, "w");
1083         if (fd == NULL)
1084         {
1085                 ereport(LOG,
1086                                 (errcode_for_file_access(),
1087                                  errmsg("could not create archive status file \"%s\": %m",
1088                                                 archiveStatusPath)));
1089                 return;
1090         }
1091         if (FreeFile(fd))
1092         {
1093                 ereport(LOG,
1094                                 (errcode_for_file_access(),
1095                                  errmsg("could not write archive status file \"%s\": %m",
1096                                                 archiveStatusPath)));
1097                 return;
1098         }
1099
1100         /* Notify archiver that it's got something to do */
1101         if (IsUnderPostmaster)
1102                 SendPostmasterSignal(PMSIGNAL_WAKEN_ARCHIVER);
1103 }
1104
1105 /*
1106  * Convenience routine to notify using log/seg representation of filename
1107  */
1108 static void
1109 XLogArchiveNotifySeg(uint32 log, uint32 seg)
1110 {
1111         char            xlog[MAXFNAMELEN];
1112
1113         XLogFileName(xlog, ThisTimeLineID, log, seg);
1114         XLogArchiveNotify(xlog);
1115 }
1116
1117 /*
1118  * XLogArchiveCheckDone
1119  *
1120  * This is called when we are ready to delete or recycle an old XLOG segment
1121  * file or backup history file.  If it is okay to delete it then return true.
1122  * If it is not time to delete it, make sure a .ready file exists, and return
1123  * false.
1124  *
1125  * If <XLOG>.done exists, then return true; else if <XLOG>.ready exists,
1126  * then return false; else create <XLOG>.ready and return false.
1127  *
1128  * The reason we do things this way is so that if the original attempt to
1129  * create <XLOG>.ready fails, we'll retry during subsequent checkpoints.
1130  */
1131 static bool
1132 XLogArchiveCheckDone(const char *xlog, bool create_if_missing)
1133 {
1134         char            archiveStatusPath[MAXPGPATH];
1135         struct stat stat_buf;
1136
1137         /* Always deletable if archiving is off */
1138         if (!XLogArchivingActive())
1139                 return true;
1140
1141         /* First check for .done --- this means archiver is done with it */
1142         StatusFilePath(archiveStatusPath, xlog, ".done");
1143         if (stat(archiveStatusPath, &stat_buf) == 0)
1144                 return true;
1145
1146         /* check for .ready --- this means archiver is still busy with it */
1147         StatusFilePath(archiveStatusPath, xlog, ".ready");
1148         if (stat(archiveStatusPath, &stat_buf) == 0)
1149                 return false;
1150
1151         /* Race condition --- maybe archiver just finished, so recheck */
1152         StatusFilePath(archiveStatusPath, xlog, ".done");
1153         if (stat(archiveStatusPath, &stat_buf) == 0)
1154                 return true;
1155
1156         /* Retry creation of the .ready file */
1157         if (create_if_missing)
1158                 XLogArchiveNotify(xlog);
1159
1160         return false;
1161 }
1162
1163 /*
1164  * XLogArchiveCleanup
1165  *
1166  * Cleanup archive notification file(s) for a particular xlog segment
1167  */
1168 static void
1169 XLogArchiveCleanup(const char *xlog)
1170 {
1171         char            archiveStatusPath[MAXPGPATH];
1172
1173         /* Remove the .done file */
1174         StatusFilePath(archiveStatusPath, xlog, ".done");
1175         unlink(archiveStatusPath);
1176         /* should we complain about failure? */
1177
1178         /* Remove the .ready file if present --- normally it shouldn't be */
1179         StatusFilePath(archiveStatusPath, xlog, ".ready");
1180         unlink(archiveStatusPath);
1181         /* should we complain about failure? */
1182 }
1183
1184 /*
1185  * Advance the Insert state to the next buffer page, writing out the next
1186  * buffer if it still contains unwritten data.
1187  *
1188  * If new_segment is TRUE then we set up the next buffer page as the first
1189  * page of the next xlog segment file, possibly but not usually the next
1190  * consecutive file page.
1191  *
1192  * The global LogwrtRqst.Write pointer needs to be advanced to include the
1193  * just-filled page.  If we can do this for free (without an extra lock),
1194  * we do so here.  Otherwise the caller must do it.  We return TRUE if the
1195  * request update still needs to be done, FALSE if we did it internally.
1196  *
1197  * Must be called with WALInsertLock held.
1198  */
1199 static bool
1200 AdvanceXLInsertBuffer(bool new_segment)
1201 {
1202         XLogCtlInsert *Insert = &XLogCtl->Insert;
1203         XLogCtlWrite *Write = &XLogCtl->Write;
1204         int                     nextidx = NextBufIdx(Insert->curridx);
1205         bool            update_needed = true;
1206         XLogRecPtr      OldPageRqstPtr;
1207         XLogwrtRqst WriteRqst;
1208         XLogRecPtr      NewPageEndPtr;
1209         XLogPageHeader NewPage;
1210
1211         /* Use Insert->LogwrtResult copy if it's more fresh */
1212         if (XLByteLT(LogwrtResult.Write, Insert->LogwrtResult.Write))
1213                 LogwrtResult = Insert->LogwrtResult;
1214
1215         /*
1216          * Get ending-offset of the buffer page we need to replace (this may be
1217          * zero if the buffer hasn't been used yet).  Fall through if it's already
1218          * written out.
1219          */
1220         OldPageRqstPtr = XLogCtl->xlblocks[nextidx];
1221         if (!XLByteLE(OldPageRqstPtr, LogwrtResult.Write))
1222         {
1223                 /* nope, got work to do... */
1224                 XLogRecPtr      FinishedPageRqstPtr;
1225
1226                 FinishedPageRqstPtr = XLogCtl->xlblocks[Insert->curridx];
1227
1228                 /* Before waiting, get info_lck and update LogwrtResult */
1229                 {
1230                         /* use volatile pointer to prevent code rearrangement */
1231                         volatile XLogCtlData *xlogctl = XLogCtl;
1232
1233                         SpinLockAcquire(&xlogctl->info_lck);
1234                         if (XLByteLT(xlogctl->LogwrtRqst.Write, FinishedPageRqstPtr))
1235                                 xlogctl->LogwrtRqst.Write = FinishedPageRqstPtr;
1236                         LogwrtResult = xlogctl->LogwrtResult;
1237                         SpinLockRelease(&xlogctl->info_lck);
1238                 }
1239
1240                 update_needed = false;  /* Did the shared-request update */
1241
1242                 if (XLByteLE(OldPageRqstPtr, LogwrtResult.Write))
1243                 {
1244                         /* OK, someone wrote it already */
1245                         Insert->LogwrtResult = LogwrtResult;
1246                 }
1247                 else
1248                 {
1249                         /* Must acquire write lock */
1250                         LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
1251                         LogwrtResult = Write->LogwrtResult;
1252                         if (XLByteLE(OldPageRqstPtr, LogwrtResult.Write))
1253                         {
1254                                 /* OK, someone wrote it already */
1255                                 LWLockRelease(WALWriteLock);
1256                                 Insert->LogwrtResult = LogwrtResult;
1257                         }
1258                         else
1259                         {
1260                                 /*
1261                                  * Have to write buffers while holding insert lock. This is
1262                                  * not good, so only write as much as we absolutely must.
1263                                  */
1264                                 WriteRqst.Write = OldPageRqstPtr;
1265                                 WriteRqst.Flush.xlogid = 0;
1266                                 WriteRqst.Flush.xrecoff = 0;
1267                                 XLogWrite(WriteRqst, false, false);
1268                                 LWLockRelease(WALWriteLock);
1269                                 Insert->LogwrtResult = LogwrtResult;
1270                         }
1271                 }
1272         }
1273
1274         /*
1275          * Now the next buffer slot is free and we can set it up to be the next
1276          * output page.
1277          */
1278         NewPageEndPtr = XLogCtl->xlblocks[Insert->curridx];
1279
1280         if (new_segment)
1281         {
1282                 /* force it to a segment start point */
1283                 NewPageEndPtr.xrecoff += XLogSegSize - 1;
1284                 NewPageEndPtr.xrecoff -= NewPageEndPtr.xrecoff % XLogSegSize;
1285         }
1286
1287         if (NewPageEndPtr.xrecoff >= XLogFileSize)
1288         {
1289                 /* crossing a logid boundary */
1290                 NewPageEndPtr.xlogid += 1;
1291                 NewPageEndPtr.xrecoff = XLOG_BLCKSZ;
1292         }
1293         else
1294                 NewPageEndPtr.xrecoff += XLOG_BLCKSZ;
1295         XLogCtl->xlblocks[nextidx] = NewPageEndPtr;
1296         NewPage = (XLogPageHeader) (XLogCtl->pages + nextidx * (Size) XLOG_BLCKSZ);
1297
1298         Insert->curridx = nextidx;
1299         Insert->currpage = NewPage;
1300
1301         Insert->currpos = ((char *) NewPage) +SizeOfXLogShortPHD;
1302
1303         /*
1304          * Be sure to re-zero the buffer so that bytes beyond what we've written
1305          * will look like zeroes and not valid XLOG records...
1306          */
1307         MemSet((char *) NewPage, 0, XLOG_BLCKSZ);
1308
1309         /*
1310          * Fill the new page's header
1311          */
1312         NewPage   ->xlp_magic = XLOG_PAGE_MAGIC;
1313
1314         /* NewPage->xlp_info = 0; */    /* done by memset */
1315         NewPage   ->xlp_tli = ThisTimeLineID;
1316         NewPage   ->xlp_pageaddr.xlogid = NewPageEndPtr.xlogid;
1317         NewPage   ->xlp_pageaddr.xrecoff = NewPageEndPtr.xrecoff - XLOG_BLCKSZ;
1318
1319         /*
1320          * If first page of an XLOG segment file, make it a long header.
1321          */
1322         if ((NewPage->xlp_pageaddr.xrecoff % XLogSegSize) == 0)
1323         {
1324                 XLogLongPageHeader NewLongPage = (XLogLongPageHeader) NewPage;
1325
1326                 NewLongPage->xlp_sysid = ControlFile->system_identifier;
1327                 NewLongPage->xlp_seg_size = XLogSegSize;
1328                 NewLongPage->xlp_xlog_blcksz = XLOG_BLCKSZ;
1329                 NewPage   ->xlp_info |= XLP_LONG_HEADER;
1330
1331                 Insert->currpos = ((char *) NewPage) +SizeOfXLogLongPHD;
1332         }
1333
1334         return update_needed;
1335 }
1336
1337 /*
1338  * Check whether we've consumed enough xlog space that a checkpoint is needed.
1339  *
1340  * Caller must have just finished filling the open log file (so that
1341  * openLogId/openLogSeg are valid).  We measure the distance from RedoRecPtr
1342  * to the open log file and see if that exceeds CheckPointSegments.
1343  *
1344  * Note: it is caller's responsibility that RedoRecPtr is up-to-date.
1345  */
1346 static bool
1347 XLogCheckpointNeeded(void)
1348 {
1349         /*
1350          * A straight computation of segment number could overflow 32 bits. Rather
1351          * than assuming we have working 64-bit arithmetic, we compare the
1352          * highest-order bits separately, and force a checkpoint immediately when
1353          * they change.
1354          */
1355         uint32          old_segno,
1356                                 new_segno;
1357         uint32          old_highbits,
1358                                 new_highbits;
1359
1360         old_segno = (RedoRecPtr.xlogid % XLogSegSize) * XLogSegsPerFile +
1361                 (RedoRecPtr.xrecoff / XLogSegSize);
1362         old_highbits = RedoRecPtr.xlogid / XLogSegSize;
1363         new_segno = (openLogId % XLogSegSize) * XLogSegsPerFile + openLogSeg;
1364         new_highbits = openLogId / XLogSegSize;
1365         if (new_highbits != old_highbits ||
1366                 new_segno >= old_segno + (uint32) (CheckPointSegments - 1))
1367                 return true;
1368         return false;
1369 }
1370
1371 /*
1372  * Write and/or fsync the log at least as far as WriteRqst indicates.
1373  *
1374  * If flexible == TRUE, we don't have to write as far as WriteRqst, but
1375  * may stop at any convenient boundary (such as a cache or logfile boundary).
1376  * This option allows us to avoid uselessly issuing multiple writes when a
1377  * single one would do.
1378  *
1379  * If xlog_switch == TRUE, we are intending an xlog segment switch, so
1380  * perform end-of-segment actions after writing the last page, even if
1381  * it's not physically the end of its segment.  (NB: this will work properly
1382  * only if caller specifies WriteRqst == page-end and flexible == false,
1383  * and there is some data to write.)
1384  *
1385  * Must be called with WALWriteLock held.
1386  */
1387 static void
1388 XLogWrite(XLogwrtRqst WriteRqst, bool flexible, bool xlog_switch)
1389 {
1390         XLogCtlWrite *Write = &XLogCtl->Write;
1391         bool            ispartialpage;
1392         bool            last_iteration;
1393         bool            finishing_seg;
1394         bool            use_existent;
1395         int                     curridx;
1396         int                     npages;
1397         int                     startidx;
1398         uint32          startoffset;
1399
1400         /* We should always be inside a critical section here */
1401         Assert(CritSectionCount > 0);
1402
1403         /*
1404          * Update local LogwrtResult (caller probably did this already, but...)
1405          */
1406         LogwrtResult = Write->LogwrtResult;
1407
1408         /*
1409          * Since successive pages in the xlog cache are consecutively allocated,
1410          * we can usually gather multiple pages together and issue just one
1411          * write() call.  npages is the number of pages we have determined can be
1412          * written together; startidx is the cache block index of the first one,
1413          * and startoffset is the file offset at which it should go. The latter
1414          * two variables are only valid when npages > 0, but we must initialize
1415          * all of them to keep the compiler quiet.
1416          */
1417         npages = 0;
1418         startidx = 0;
1419         startoffset = 0;
1420
1421         /*
1422          * Within the loop, curridx is the cache block index of the page to
1423          * consider writing.  We advance Write->curridx only after successfully
1424          * writing pages.  (Right now, this refinement is useless since we are
1425          * going to PANIC if any error occurs anyway; but someday it may come in
1426          * useful.)
1427          */
1428         curridx = Write->curridx;
1429
1430         while (XLByteLT(LogwrtResult.Write, WriteRqst.Write))
1431         {
1432                 /*
1433                  * Make sure we're not ahead of the insert process.  This could happen
1434                  * if we're passed a bogus WriteRqst.Write that is past the end of the
1435                  * last page that's been initialized by AdvanceXLInsertBuffer.
1436                  */
1437                 if (!XLByteLT(LogwrtResult.Write, XLogCtl->xlblocks[curridx]))
1438                         elog(PANIC, "xlog write request %X/%X is past end of log %X/%X",
1439                                  LogwrtResult.Write.xlogid, LogwrtResult.Write.xrecoff,
1440                                  XLogCtl->xlblocks[curridx].xlogid,
1441                                  XLogCtl->xlblocks[curridx].xrecoff);
1442
1443                 /* Advance LogwrtResult.Write to end of current buffer page */
1444                 LogwrtResult.Write = XLogCtl->xlblocks[curridx];
1445                 ispartialpage = XLByteLT(WriteRqst.Write, LogwrtResult.Write);
1446
1447                 if (!XLByteInPrevSeg(LogwrtResult.Write, openLogId, openLogSeg))
1448                 {
1449                         /*
1450                          * Switch to new logfile segment.  We cannot have any pending
1451                          * pages here (since we dump what we have at segment end).
1452                          */
1453                         Assert(npages == 0);
1454                         if (openLogFile >= 0)
1455                                 XLogFileClose();
1456                         XLByteToPrevSeg(LogwrtResult.Write, openLogId, openLogSeg);
1457
1458                         /* create/use new log file */
1459                         use_existent = true;
1460                         openLogFile = XLogFileInit(openLogId, openLogSeg,
1461                                                                            &use_existent, true);
1462                         openLogOff = 0;
1463                 }
1464
1465                 /* Make sure we have the current logfile open */
1466                 if (openLogFile < 0)
1467                 {
1468                         XLByteToPrevSeg(LogwrtResult.Write, openLogId, openLogSeg);
1469                         openLogFile = XLogFileOpen(openLogId, openLogSeg);
1470                         openLogOff = 0;
1471                 }
1472
1473                 /* Add current page to the set of pending pages-to-dump */
1474                 if (npages == 0)
1475                 {
1476                         /* first of group */
1477                         startidx = curridx;
1478                         startoffset = (LogwrtResult.Write.xrecoff - XLOG_BLCKSZ) % XLogSegSize;
1479                 }
1480                 npages++;
1481
1482                 /*
1483                  * Dump the set if this will be the last loop iteration, or if we are
1484                  * at the last page of the cache area (since the next page won't be
1485                  * contiguous in memory), or if we are at the end of the logfile
1486                  * segment.
1487                  */
1488                 last_iteration = !XLByteLT(LogwrtResult.Write, WriteRqst.Write);
1489
1490                 finishing_seg = !ispartialpage &&
1491                         (startoffset + npages * XLOG_BLCKSZ) >= XLogSegSize;
1492
1493                 if (last_iteration ||
1494                         curridx == XLogCtl->XLogCacheBlck ||
1495                         finishing_seg)
1496                 {
1497                         char       *from;
1498                         Size            nbytes;
1499
1500                         /* Need to seek in the file? */
1501                         if (openLogOff != startoffset)
1502                         {
1503                                 if (lseek(openLogFile, (off_t) startoffset, SEEK_SET) < 0)
1504                                         ereport(PANIC,
1505                                                         (errcode_for_file_access(),
1506                                                          errmsg("could not seek in log file %u, "
1507                                                                         "segment %u to offset %u: %m",
1508                                                                         openLogId, openLogSeg, startoffset)));
1509                                 openLogOff = startoffset;
1510                         }
1511
1512                         /* OK to write the page(s) */
1513                         from = XLogCtl->pages + startidx * (Size) XLOG_BLCKSZ;
1514                         nbytes = npages * (Size) XLOG_BLCKSZ;
1515                         errno = 0;
1516                         if (write(openLogFile, from, nbytes) != nbytes)
1517                         {
1518                                 /* if write didn't set errno, assume no disk space */
1519                                 if (errno == 0)
1520                                         errno = ENOSPC;
1521                                 ereport(PANIC,
1522                                                 (errcode_for_file_access(),
1523                                                  errmsg("could not write to log file %u, segment %u "
1524                                                                 "at offset %u, length %lu: %m",
1525                                                                 openLogId, openLogSeg,
1526                                                                 openLogOff, (unsigned long) nbytes)));
1527                         }
1528
1529                         /* Update state for write */
1530                         openLogOff += nbytes;
1531                         Write->curridx = ispartialpage ? curridx : NextBufIdx(curridx);
1532                         npages = 0;
1533
1534                         /*
1535                          * If we just wrote the whole last page of a logfile segment,
1536                          * fsync the segment immediately.  This avoids having to go back
1537                          * and re-open prior segments when an fsync request comes along
1538                          * later. Doing it here ensures that one and only one backend will
1539                          * perform this fsync.
1540                          *
1541                          * We also do this if this is the last page written for an xlog
1542                          * switch.
1543                          *
1544                          * This is also the right place to notify the Archiver that the
1545                          * segment is ready to copy to archival storage, and to update the
1546                          * timer for archive_timeout, and to signal for a checkpoint if
1547                          * too many logfile segments have been used since the last
1548                          * checkpoint.
1549                          */
1550                         if (finishing_seg || (xlog_switch && last_iteration))
1551                         {
1552                                 issue_xlog_fsync();
1553                                 LogwrtResult.Flush = LogwrtResult.Write;                /* end of page */
1554
1555                                 if (XLogArchivingActive())
1556                                         XLogArchiveNotifySeg(openLogId, openLogSeg);
1557
1558                                 Write->lastSegSwitchTime = (pg_time_t) time(NULL);
1559
1560                                 /*
1561                                  * Signal bgwriter to start a checkpoint if we've consumed too
1562                                  * much xlog since the last one.  For speed, we first check
1563                                  * using the local copy of RedoRecPtr, which might be out of
1564                                  * date; if it looks like a checkpoint is needed, forcibly
1565                                  * update RedoRecPtr and recheck.
1566                                  */
1567                                 if (IsUnderPostmaster &&
1568                                         XLogCheckpointNeeded())
1569                                 {
1570                                         (void) GetRedoRecPtr();
1571                                         if (XLogCheckpointNeeded())
1572                                                 RequestCheckpoint(CHECKPOINT_CAUSE_XLOG);
1573                                 }
1574                         }
1575                 }
1576
1577                 if (ispartialpage)
1578                 {
1579                         /* Only asked to write a partial page */
1580                         LogwrtResult.Write = WriteRqst.Write;
1581                         break;
1582                 }
1583                 curridx = NextBufIdx(curridx);
1584
1585                 /* If flexible, break out of loop as soon as we wrote something */
1586                 if (flexible && npages == 0)
1587                         break;
1588         }
1589
1590         Assert(npages == 0);
1591         Assert(curridx == Write->curridx);
1592
1593         /*
1594          * If asked to flush, do so
1595          */
1596         if (XLByteLT(LogwrtResult.Flush, WriteRqst.Flush) &&
1597                 XLByteLT(LogwrtResult.Flush, LogwrtResult.Write))
1598         {
1599                 /*
1600                  * Could get here without iterating above loop, in which case we might
1601                  * have no open file or the wrong one.  However, we do not need to
1602                  * fsync more than one file.
1603                  */
1604                 if (sync_method != SYNC_METHOD_OPEN)
1605                 {
1606                         if (openLogFile >= 0 &&
1607                                 !XLByteInPrevSeg(LogwrtResult.Write, openLogId, openLogSeg))
1608                                 XLogFileClose();
1609                         if (openLogFile < 0)
1610                         {
1611                                 XLByteToPrevSeg(LogwrtResult.Write, openLogId, openLogSeg);
1612                                 openLogFile = XLogFileOpen(openLogId, openLogSeg);
1613                                 openLogOff = 0;
1614                         }
1615                         issue_xlog_fsync();
1616                 }
1617                 LogwrtResult.Flush = LogwrtResult.Write;
1618         }
1619
1620         /*
1621          * Update shared-memory status
1622          *
1623          * We make sure that the shared 'request' values do not fall behind the
1624          * 'result' values.  This is not absolutely essential, but it saves some
1625          * code in a couple of places.
1626          */
1627         {
1628                 /* use volatile pointer to prevent code rearrangement */
1629                 volatile XLogCtlData *xlogctl = XLogCtl;
1630
1631                 SpinLockAcquire(&xlogctl->info_lck);
1632                 xlogctl->LogwrtResult = LogwrtResult;
1633                 if (XLByteLT(xlogctl->LogwrtRqst.Write, LogwrtResult.Write))
1634                         xlogctl->LogwrtRqst.Write = LogwrtResult.Write;
1635                 if (XLByteLT(xlogctl->LogwrtRqst.Flush, LogwrtResult.Flush))
1636                         xlogctl->LogwrtRqst.Flush = LogwrtResult.Flush;
1637                 SpinLockRelease(&xlogctl->info_lck);
1638         }
1639
1640         Write->LogwrtResult = LogwrtResult;
1641 }
1642
1643 /*
1644  * Record the LSN for an asynchronous transaction commit.
1645  * (This should not be called for aborts, nor for synchronous commits.)
1646  */
1647 void
1648 XLogSetAsyncCommitLSN(XLogRecPtr asyncCommitLSN)
1649 {
1650         /* use volatile pointer to prevent code rearrangement */
1651         volatile XLogCtlData *xlogctl = XLogCtl;
1652
1653         SpinLockAcquire(&xlogctl->info_lck);
1654         if (XLByteLT(xlogctl->asyncCommitLSN, asyncCommitLSN))
1655                 xlogctl->asyncCommitLSN = asyncCommitLSN;
1656         SpinLockRelease(&xlogctl->info_lck);
1657 }
1658
1659 /*
1660  * Ensure that all XLOG data through the given position is flushed to disk.
1661  *
1662  * NOTE: this differs from XLogWrite mainly in that the WALWriteLock is not
1663  * already held, and we try to avoid acquiring it if possible.
1664  */
1665 void
1666 XLogFlush(XLogRecPtr record)
1667 {
1668         XLogRecPtr      WriteRqstPtr;
1669         XLogwrtRqst WriteRqst;
1670
1671         /* Disabled during REDO */
1672         if (InRedo)
1673                 return;
1674
1675         /* Quick exit if already known flushed */
1676         if (XLByteLE(record, LogwrtResult.Flush))
1677                 return;
1678
1679 #ifdef WAL_DEBUG
1680         if (XLOG_DEBUG)
1681                 elog(LOG, "xlog flush request %X/%X; write %X/%X; flush %X/%X",
1682                          record.xlogid, record.xrecoff,
1683                          LogwrtResult.Write.xlogid, LogwrtResult.Write.xrecoff,
1684                          LogwrtResult.Flush.xlogid, LogwrtResult.Flush.xrecoff);
1685 #endif
1686
1687         START_CRIT_SECTION();
1688
1689         /*
1690          * Since fsync is usually a horribly expensive operation, we try to
1691          * piggyback as much data as we can on each fsync: if we see any more data
1692          * entered into the xlog buffer, we'll write and fsync that too, so that
1693          * the final value of LogwrtResult.Flush is as large as possible. This
1694          * gives us some chance of avoiding another fsync immediately after.
1695          */
1696
1697         /* initialize to given target; may increase below */
1698         WriteRqstPtr = record;
1699
1700         /* read LogwrtResult and update local state */
1701         {
1702                 /* use volatile pointer to prevent code rearrangement */
1703                 volatile XLogCtlData *xlogctl = XLogCtl;
1704
1705                 SpinLockAcquire(&xlogctl->info_lck);
1706                 if (XLByteLT(WriteRqstPtr, xlogctl->LogwrtRqst.Write))
1707                         WriteRqstPtr = xlogctl->LogwrtRqst.Write;
1708                 LogwrtResult = xlogctl->LogwrtResult;
1709                 SpinLockRelease(&xlogctl->info_lck);
1710         }
1711
1712         /* done already? */
1713         if (!XLByteLE(record, LogwrtResult.Flush))
1714         {
1715                 /* now wait for the write lock */
1716                 LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
1717                 LogwrtResult = XLogCtl->Write.LogwrtResult;
1718                 if (!XLByteLE(record, LogwrtResult.Flush))
1719                 {
1720                         /* try to write/flush later additions to XLOG as well */
1721                         if (LWLockConditionalAcquire(WALInsertLock, LW_EXCLUSIVE))
1722                         {
1723                                 XLogCtlInsert *Insert = &XLogCtl->Insert;
1724                                 uint32          freespace = INSERT_FREESPACE(Insert);
1725
1726                                 if (freespace < SizeOfXLogRecord)               /* buffer is full */
1727                                         WriteRqstPtr = XLogCtl->xlblocks[Insert->curridx];
1728                                 else
1729                                 {
1730                                         WriteRqstPtr = XLogCtl->xlblocks[Insert->curridx];
1731                                         WriteRqstPtr.xrecoff -= freespace;
1732                                 }
1733                                 LWLockRelease(WALInsertLock);
1734                                 WriteRqst.Write = WriteRqstPtr;
1735                                 WriteRqst.Flush = WriteRqstPtr;
1736                         }
1737                         else
1738                         {
1739                                 WriteRqst.Write = WriteRqstPtr;
1740                                 WriteRqst.Flush = record;
1741                         }
1742                         XLogWrite(WriteRqst, false, false);
1743                 }
1744                 LWLockRelease(WALWriteLock);
1745         }
1746
1747         END_CRIT_SECTION();
1748
1749         /*
1750          * If we still haven't flushed to the request point then we have a
1751          * problem; most likely, the requested flush point is past end of XLOG.
1752          * This has been seen to occur when a disk page has a corrupted LSN.
1753          *
1754          * Formerly we treated this as a PANIC condition, but that hurts the
1755          * system's robustness rather than helping it: we do not want to take down
1756          * the whole system due to corruption on one data page.  In particular, if
1757          * the bad page is encountered again during recovery then we would be
1758          * unable to restart the database at all!  (This scenario has actually
1759          * happened in the field several times with 7.1 releases. Note that we
1760          * cannot get here while InRedo is true, but if the bad page is brought in
1761          * and marked dirty during recovery then CreateCheckPoint will try to
1762          * flush it at the end of recovery.)
1763          *
1764          * The current approach is to ERROR under normal conditions, but only
1765          * WARNING during recovery, so that the system can be brought up even if
1766          * there's a corrupt LSN.  Note that for calls from xact.c, the ERROR will
1767          * be promoted to PANIC since xact.c calls this routine inside a critical
1768          * section.  However, calls from bufmgr.c are not within critical sections
1769          * and so we will not force a restart for a bad LSN on a data page.
1770          */
1771         if (XLByteLT(LogwrtResult.Flush, record))
1772                 elog(InRecovery ? WARNING : ERROR,
1773                 "xlog flush request %X/%X is not satisfied --- flushed only to %X/%X",
1774                          record.xlogid, record.xrecoff,
1775                          LogwrtResult.Flush.xlogid, LogwrtResult.Flush.xrecoff);
1776 }
1777
1778 /*
1779  * Flush xlog, but without specifying exactly where to flush to.
1780  *
1781  * We normally flush only completed blocks; but if there is nothing to do on
1782  * that basis, we check for unflushed async commits in the current incomplete
1783  * block, and flush through the latest one of those.  Thus, if async commits
1784  * are not being used, we will flush complete blocks only.      We can guarantee
1785  * that async commits reach disk after at most three cycles; normally only
1786  * one or two.  (We allow XLogWrite to write "flexibly", meaning it can stop
1787  * at the end of the buffer ring; this makes a difference only with very high
1788  * load or long wal_writer_delay, but imposes one extra cycle for the worst
1789  * case for async commits.)
1790  *
1791  * This routine is invoked periodically by the background walwriter process.
1792  */
1793 void
1794 XLogBackgroundFlush(void)
1795 {
1796         XLogRecPtr      WriteRqstPtr;
1797         bool            flexible = true;
1798
1799         /* read LogwrtResult and update local state */
1800         {
1801                 /* use volatile pointer to prevent code rearrangement */
1802                 volatile XLogCtlData *xlogctl = XLogCtl;
1803
1804                 SpinLockAcquire(&xlogctl->info_lck);
1805                 LogwrtResult = xlogctl->LogwrtResult;
1806                 WriteRqstPtr = xlogctl->LogwrtRqst.Write;
1807                 SpinLockRelease(&xlogctl->info_lck);
1808         }
1809
1810         /* back off to last completed page boundary */
1811         WriteRqstPtr.xrecoff -= WriteRqstPtr.xrecoff % XLOG_BLCKSZ;
1812
1813         /* if we have already flushed that far, consider async commit records */
1814         if (XLByteLE(WriteRqstPtr, LogwrtResult.Flush))
1815         {
1816                 /* use volatile pointer to prevent code rearrangement */
1817                 volatile XLogCtlData *xlogctl = XLogCtl;
1818
1819                 SpinLockAcquire(&xlogctl->info_lck);
1820                 WriteRqstPtr = xlogctl->asyncCommitLSN;
1821                 SpinLockRelease(&xlogctl->info_lck);
1822                 flexible = false;               /* ensure it all gets written */
1823         }
1824
1825         /* Done if already known flushed */
1826         if (XLByteLE(WriteRqstPtr, LogwrtResult.Flush))
1827                 return;
1828
1829 #ifdef WAL_DEBUG
1830         if (XLOG_DEBUG)
1831                 elog(LOG, "xlog bg flush request %X/%X; write %X/%X; flush %X/%X",
1832                          WriteRqstPtr.xlogid, WriteRqstPtr.xrecoff,
1833                          LogwrtResult.Write.xlogid, LogwrtResult.Write.xrecoff,
1834                          LogwrtResult.Flush.xlogid, LogwrtResult.Flush.xrecoff);
1835 #endif
1836
1837         START_CRIT_SECTION();
1838
1839         /* now wait for the write lock */
1840         LWLockAcquire(WALWriteLock, LW_EXCLUSIVE);
1841         LogwrtResult = XLogCtl->Write.LogwrtResult;
1842         if (!XLByteLE(WriteRqstPtr, LogwrtResult.Flush))
1843         {
1844                 XLogwrtRqst WriteRqst;
1845
1846                 WriteRqst.Write = WriteRqstPtr;
1847                 WriteRqst.Flush = WriteRqstPtr;
1848                 XLogWrite(WriteRqst, flexible, false);
1849         }
1850         LWLockRelease(WALWriteLock);
1851
1852         END_CRIT_SECTION();
1853 }
1854
1855 /*
1856  * Flush any previous asynchronously-committed transactions' commit records.
1857  *
1858  * NOTE: it is unwise to assume that this provides any strong guarantees.
1859  * In particular, because of the inexact LSN bookkeeping used by clog.c,
1860  * we cannot assume that hint bits will be settable for these transactions.
1861  */
1862 void
1863 XLogAsyncCommitFlush(void)
1864 {
1865         XLogRecPtr      WriteRqstPtr;
1866
1867         /* use volatile pointer to prevent code rearrangement */
1868         volatile XLogCtlData *xlogctl = XLogCtl;
1869
1870         SpinLockAcquire(&xlogctl->info_lck);
1871         WriteRqstPtr = xlogctl->asyncCommitLSN;
1872         SpinLockRelease(&xlogctl->info_lck);
1873
1874         XLogFlush(WriteRqstPtr);
1875 }
1876
1877 /*
1878  * Test whether XLOG data has been flushed up to (at least) the given position.
1879  *
1880  * Returns true if a flush is still needed.  (It may be that someone else
1881  * is already in process of flushing that far, however.)
1882  */
1883 bool
1884 XLogNeedsFlush(XLogRecPtr record)
1885 {
1886         /* Quick exit if already known flushed */
1887         if (XLByteLE(record, LogwrtResult.Flush))
1888                 return false;
1889
1890         /* read LogwrtResult and update local state */
1891         {
1892                 /* use volatile pointer to prevent code rearrangement */
1893                 volatile XLogCtlData *xlogctl = XLogCtl;
1894
1895                 SpinLockAcquire(&xlogctl->info_lck);
1896                 LogwrtResult = xlogctl->LogwrtResult;
1897                 SpinLockRelease(&xlogctl->info_lck);
1898         }
1899
1900         /* check again */
1901         if (XLByteLE(record, LogwrtResult.Flush))
1902                 return false;
1903
1904         return true;
1905 }
1906
1907 /*
1908  * Create a new XLOG file segment, or open a pre-existing one.
1909  *
1910  * log, seg: identify segment to be created/opened.
1911  *
1912  * *use_existent: if TRUE, OK to use a pre-existing file (else, any
1913  * pre-existing file will be deleted).  On return, TRUE if a pre-existing
1914  * file was used.
1915  *
1916  * use_lock: if TRUE, acquire ControlFileLock while moving file into
1917  * place.  This should be TRUE except during bootstrap log creation.  The
1918  * caller must *not* hold the lock at call.
1919  *
1920  * Returns FD of opened file.
1921  *
1922  * Note: errors here are ERROR not PANIC because we might or might not be
1923  * inside a critical section (eg, during checkpoint there is no reason to
1924  * take down the system on failure).  They will promote to PANIC if we are
1925  * in a critical section.
1926  */
1927 static int
1928 XLogFileInit(uint32 log, uint32 seg,
1929                          bool *use_existent, bool use_lock)
1930 {
1931         char            path[MAXPGPATH];
1932         char            tmppath[MAXPGPATH];
1933         char       *zbuffer;
1934         uint32          installed_log;
1935         uint32          installed_seg;
1936         int                     max_advance;
1937         int                     fd;
1938         int                     nbytes;
1939
1940         XLogFilePath(path, ThisTimeLineID, log, seg);
1941
1942         /*
1943          * Try to use existent file (checkpoint maker may have created it already)
1944          */
1945         if (*use_existent)
1946         {
1947                 fd = BasicOpenFile(path, O_RDWR | PG_BINARY | XLOG_SYNC_BIT,
1948                                                    S_IRUSR | S_IWUSR);
1949                 if (fd < 0)
1950                 {
1951                         if (errno != ENOENT)
1952                                 ereport(ERROR,
1953                                                 (errcode_for_file_access(),
1954                                                  errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
1955                                                                 path, log, seg)));
1956                 }
1957                 else
1958                         return fd;
1959         }
1960
1961         /*
1962          * Initialize an empty (all zeroes) segment.  NOTE: it is possible that
1963          * another process is doing the same thing.  If so, we will end up
1964          * pre-creating an extra log segment.  That seems OK, and better than
1965          * holding the lock throughout this lengthy process.
1966          */
1967         elog(DEBUG2, "creating and filling new WAL file");
1968
1969         snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
1970
1971         unlink(tmppath);
1972
1973         /* do not use XLOG_SYNC_BIT here --- want to fsync only at end of fill */
1974         fd = BasicOpenFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
1975                                            S_IRUSR | S_IWUSR);
1976         if (fd < 0)
1977                 ereport(ERROR,
1978                                 (errcode_for_file_access(),
1979                                  errmsg("could not create file \"%s\": %m", tmppath)));
1980
1981         /*
1982          * Zero-fill the file.  We have to do this the hard way to ensure that all
1983          * the file space has really been allocated --- on platforms that allow
1984          * "holes" in files, just seeking to the end doesn't allocate intermediate
1985          * space.  This way, we know that we have all the space and (after the
1986          * fsync below) that all the indirect blocks are down on disk.  Therefore,
1987          * fdatasync(2) or O_DSYNC will be sufficient to sync future writes to the
1988          * log file.
1989          *
1990          * Note: palloc zbuffer, instead of just using a local char array, to
1991          * ensure it is reasonably well-aligned; this may save a few cycles
1992          * transferring data to the kernel.
1993          */
1994         zbuffer = (char *) palloc0(XLOG_BLCKSZ);
1995         for (nbytes = 0; nbytes < XLogSegSize; nbytes += XLOG_BLCKSZ)
1996         {
1997                 errno = 0;
1998                 if ((int) write(fd, zbuffer, XLOG_BLCKSZ) != (int) XLOG_BLCKSZ)
1999                 {
2000                         int                     save_errno = errno;
2001
2002                         /*
2003                          * If we fail to make the file, delete it to release disk space
2004                          */
2005                         unlink(tmppath);
2006                         /* if write didn't set errno, assume problem is no disk space */
2007                         errno = save_errno ? save_errno : ENOSPC;
2008
2009                         ereport(ERROR,
2010                                         (errcode_for_file_access(),
2011                                          errmsg("could not write to file \"%s\": %m", tmppath)));
2012                 }
2013         }
2014         pfree(zbuffer);
2015
2016         if (pg_fsync(fd) != 0)
2017                 ereport(ERROR,
2018                                 (errcode_for_file_access(),
2019                                  errmsg("could not fsync file \"%s\": %m", tmppath)));
2020
2021         if (close(fd))
2022                 ereport(ERROR,
2023                                 (errcode_for_file_access(),
2024                                  errmsg("could not close file \"%s\": %m", tmppath)));
2025
2026         /*
2027          * Now move the segment into place with its final name.
2028          *
2029          * If caller didn't want to use a pre-existing file, get rid of any
2030          * pre-existing file.  Otherwise, cope with possibility that someone else
2031          * has created the file while we were filling ours: if so, use ours to
2032          * pre-create a future log segment.
2033          */
2034         installed_log = log;
2035         installed_seg = seg;
2036         max_advance = XLOGfileslop;
2037         if (!InstallXLogFileSegment(&installed_log, &installed_seg, tmppath,
2038                                                                 *use_existent, &max_advance,
2039                                                                 use_lock))
2040         {
2041                 /* No need for any more future segments... */
2042                 unlink(tmppath);
2043         }
2044
2045         elog(DEBUG2, "done creating and filling new WAL file");
2046
2047         /* Set flag to tell caller there was no existent file */
2048         *use_existent = false;
2049
2050         /* Now open original target segment (might not be file I just made) */
2051         fd = BasicOpenFile(path, O_RDWR | PG_BINARY | XLOG_SYNC_BIT,
2052                                            S_IRUSR | S_IWUSR);
2053         if (fd < 0)
2054                 ereport(ERROR,
2055                                 (errcode_for_file_access(),
2056                    errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
2057                                   path, log, seg)));
2058
2059         return fd;
2060 }
2061
2062 /*
2063  * Create a new XLOG file segment by copying a pre-existing one.
2064  *
2065  * log, seg: identify segment to be created.
2066  *
2067  * srcTLI, srclog, srcseg: identify segment to be copied (could be from
2068  *              a different timeline)
2069  *
2070  * Currently this is only used during recovery, and so there are no locking
2071  * considerations.      But we should be just as tense as XLogFileInit to avoid
2072  * emplacing a bogus file.
2073  */
2074 static void
2075 XLogFileCopy(uint32 log, uint32 seg,
2076                          TimeLineID srcTLI, uint32 srclog, uint32 srcseg)
2077 {
2078         char            path[MAXPGPATH];
2079         char            tmppath[MAXPGPATH];
2080         char            buffer[XLOG_BLCKSZ];
2081         int                     srcfd;
2082         int                     fd;
2083         int                     nbytes;
2084
2085         /*
2086          * Open the source file
2087          */
2088         XLogFilePath(path, srcTLI, srclog, srcseg);
2089         srcfd = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
2090         if (srcfd < 0)
2091                 ereport(ERROR,
2092                                 (errcode_for_file_access(),
2093                                  errmsg("could not open file \"%s\": %m", path)));
2094
2095         /*
2096          * Copy into a temp file name.
2097          */
2098         snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
2099
2100         unlink(tmppath);
2101
2102         /* do not use XLOG_SYNC_BIT here --- want to fsync only at end of fill */
2103         fd = BasicOpenFile(tmppath, O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
2104                                            S_IRUSR | S_IWUSR);
2105         if (fd < 0)
2106                 ereport(ERROR,
2107                                 (errcode_for_file_access(),
2108                                  errmsg("could not create file \"%s\": %m", tmppath)));
2109
2110         /*
2111          * Do the data copying.
2112          */
2113         for (nbytes = 0; nbytes < XLogSegSize; nbytes += sizeof(buffer))
2114         {
2115                 errno = 0;
2116                 if ((int) read(srcfd, buffer, sizeof(buffer)) != (int) sizeof(buffer))
2117                 {
2118                         if (errno != 0)
2119                                 ereport(ERROR,
2120                                                 (errcode_for_file_access(),
2121                                                  errmsg("could not read file \"%s\": %m", path)));
2122                         else
2123                                 ereport(ERROR,
2124                                                 (errmsg("not enough data in file \"%s\"", path)));
2125                 }
2126                 errno = 0;
2127                 if ((int) write(fd, buffer, sizeof(buffer)) != (int) sizeof(buffer))
2128                 {
2129                         int                     save_errno = errno;
2130
2131                         /*
2132                          * If we fail to make the file, delete it to release disk space
2133                          */
2134                         unlink(tmppath);
2135                         /* if write didn't set errno, assume problem is no disk space */
2136                         errno = save_errno ? save_errno : ENOSPC;
2137
2138                         ereport(ERROR,
2139                                         (errcode_for_file_access(),
2140                                          errmsg("could not write to file \"%s\": %m", tmppath)));
2141                 }
2142         }
2143
2144         if (pg_fsync(fd) != 0)
2145                 ereport(ERROR,
2146                                 (errcode_for_file_access(),
2147                                  errmsg("could not fsync file \"%s\": %m", tmppath)));
2148
2149         if (close(fd))
2150                 ereport(ERROR,
2151                                 (errcode_for_file_access(),
2152                                  errmsg("could not close file \"%s\": %m", tmppath)));
2153
2154         close(srcfd);
2155
2156         /*
2157          * Now move the segment into place with its final name.
2158          */
2159         if (!InstallXLogFileSegment(&log, &seg, tmppath, false, NULL, false))
2160                 elog(ERROR, "InstallXLogFileSegment should not have failed");
2161 }
2162
2163 /*
2164  * Install a new XLOG segment file as a current or future log segment.
2165  *
2166  * This is used both to install a newly-created segment (which has a temp
2167  * filename while it's being created) and to recycle an old segment.
2168  *
2169  * *log, *seg: identify segment to install as (or first possible target).
2170  * When find_free is TRUE, these are modified on return to indicate the
2171  * actual installation location or last segment searched.
2172  *
2173  * tmppath: initial name of file to install.  It will be renamed into place.
2174  *
2175  * find_free: if TRUE, install the new segment at the first empty log/seg
2176  * number at or after the passed numbers.  If FALSE, install the new segment
2177  * exactly where specified, deleting any existing segment file there.
2178  *
2179  * *max_advance: maximum number of log/seg slots to advance past the starting
2180  * point.  Fail if no free slot is found in this range.  On return, reduced
2181  * by the number of slots skipped over.  (Irrelevant, and may be NULL,
2182  * when find_free is FALSE.)
2183  *
2184  * use_lock: if TRUE, acquire ControlFileLock while moving file into
2185  * place.  This should be TRUE except during bootstrap log creation.  The
2186  * caller must *not* hold the lock at call.
2187  *
2188  * Returns TRUE if file installed, FALSE if not installed because of
2189  * exceeding max_advance limit.  On Windows, we also return FALSE if we
2190  * can't rename the file into place because someone's got it open.
2191  * (Any other kind of failure causes ereport().)
2192  */
2193 static bool
2194 InstallXLogFileSegment(uint32 *log, uint32 *seg, char *tmppath,
2195                                            bool find_free, int *max_advance,
2196                                            bool use_lock)
2197 {
2198         char            path[MAXPGPATH];
2199         struct stat stat_buf;
2200
2201         XLogFilePath(path, ThisTimeLineID, *log, *seg);
2202
2203         /*
2204          * We want to be sure that only one process does this at a time.
2205          */
2206         if (use_lock)
2207                 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
2208
2209         if (!find_free)
2210         {
2211                 /* Force installation: get rid of any pre-existing segment file */
2212                 unlink(path);
2213         }
2214         else
2215         {
2216                 /* Find a free slot to put it in */
2217                 while (stat(path, &stat_buf) == 0)
2218                 {
2219                         if (*max_advance <= 0)
2220                         {
2221                                 /* Failed to find a free slot within specified range */
2222                                 if (use_lock)
2223                                         LWLockRelease(ControlFileLock);
2224                                 return false;
2225                         }
2226                         NextLogSeg(*log, *seg);
2227                         (*max_advance)--;
2228                         XLogFilePath(path, ThisTimeLineID, *log, *seg);
2229                 }
2230         }
2231
2232         /*
2233          * Prefer link() to rename() here just to be really sure that we don't
2234          * overwrite an existing logfile.  However, there shouldn't be one, so
2235          * rename() is an acceptable substitute except for the truly paranoid.
2236          */
2237 #if HAVE_WORKING_LINK
2238         if (link(tmppath, path) < 0)
2239                 ereport(ERROR,
2240                                 (errcode_for_file_access(),
2241                                  errmsg("could not link file \"%s\" to \"%s\" (initialization of log file %u, segment %u): %m",
2242                                                 tmppath, path, *log, *seg)));
2243         unlink(tmppath);
2244 #else
2245         if (rename(tmppath, path) < 0)
2246         {
2247 #ifdef WIN32
2248 #if !defined(__CYGWIN__)
2249                 if (GetLastError() == ERROR_ACCESS_DENIED)
2250 #else
2251                 if (errno == EACCES)
2252 #endif
2253                 {
2254                         if (use_lock)
2255                                 LWLockRelease(ControlFileLock);
2256                         return false;
2257                 }
2258 #endif   /* WIN32 */
2259
2260                 ereport(ERROR,
2261                                 (errcode_for_file_access(),
2262                                  errmsg("could not rename file \"%s\" to \"%s\" (initialization of log file %u, segment %u): %m",
2263                                                 tmppath, path, *log, *seg)));
2264         }
2265 #endif
2266
2267         if (use_lock)
2268                 LWLockRelease(ControlFileLock);
2269
2270         return true;
2271 }
2272
2273 /*
2274  * Open a pre-existing logfile segment for writing.
2275  */
2276 static int
2277 XLogFileOpen(uint32 log, uint32 seg)
2278 {
2279         char            path[MAXPGPATH];
2280         int                     fd;
2281
2282         XLogFilePath(path, ThisTimeLineID, log, seg);
2283
2284         fd = BasicOpenFile(path, O_RDWR | PG_BINARY | XLOG_SYNC_BIT,
2285                                            S_IRUSR | S_IWUSR);
2286         if (fd < 0)
2287                 ereport(PANIC,
2288                                 (errcode_for_file_access(),
2289                    errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
2290                                   path, log, seg)));
2291
2292         return fd;
2293 }
2294
2295 /*
2296  * Open a logfile segment for reading (during recovery).
2297  */
2298 static int
2299 XLogFileRead(uint32 log, uint32 seg, int emode)
2300 {
2301         char            path[MAXPGPATH];
2302         char            xlogfname[MAXFNAMELEN];
2303         char            activitymsg[MAXFNAMELEN + 16];
2304         ListCell   *cell;
2305         int                     fd;
2306
2307         /*
2308          * Loop looking for a suitable timeline ID: we might need to read any of
2309          * the timelines listed in expectedTLIs.
2310          *
2311          * We expect curFileTLI on entry to be the TLI of the preceding file in
2312          * sequence, or 0 if there was no predecessor.  We do not allow curFileTLI
2313          * to go backwards; this prevents us from picking up the wrong file when a
2314          * parent timeline extends to higher segment numbers than the child we
2315          * want to read.
2316          */
2317         foreach(cell, expectedTLIs)
2318         {
2319                 TimeLineID      tli = (TimeLineID) lfirst_int(cell);
2320
2321                 if (tli < curFileTLI)
2322                         break;                          /* don't bother looking at too-old TLIs */
2323
2324                 XLogFileName(xlogfname, tli, log, seg);
2325
2326                 if (InArchiveRecovery)
2327                 {
2328                         /* Report recovery progress in PS display */
2329                         snprintf(activitymsg, sizeof(activitymsg), "waiting for %s",
2330                                          xlogfname);
2331                         set_ps_display(activitymsg, false);
2332
2333                         restoredFromArchive = RestoreArchivedFile(path, xlogfname,
2334                                                                                                           "RECOVERYXLOG",
2335                                                                                                           XLogSegSize);
2336                 }
2337                 else
2338                         XLogFilePath(path, tli, log, seg);
2339
2340                 fd = BasicOpenFile(path, O_RDONLY | PG_BINARY, 0);
2341                 if (fd >= 0)
2342                 {
2343                         /* Success! */
2344                         curFileTLI = tli;
2345
2346                         /* Report recovery progress in PS display */
2347                         snprintf(activitymsg, sizeof(activitymsg), "recovering %s",
2348                                          xlogfname);
2349                         set_ps_display(activitymsg, false);
2350
2351                         return fd;
2352                 }
2353                 if (errno != ENOENT)    /* unexpected failure? */
2354                         ereport(PANIC,
2355                                         (errcode_for_file_access(),
2356                         errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
2357                                    path, log, seg)));
2358         }
2359
2360         /* Couldn't find it.  For simplicity, complain about front timeline */
2361         XLogFilePath(path, recoveryTargetTLI, log, seg);
2362         errno = ENOENT;
2363         ereport(emode,
2364                         (errcode_for_file_access(),
2365                    errmsg("could not open file \"%s\" (log file %u, segment %u): %m",
2366                                   path, log, seg)));
2367         return -1;
2368 }
2369
2370 /*
2371  * Close the current logfile segment for writing.
2372  */
2373 static void
2374 XLogFileClose(void)
2375 {
2376         Assert(openLogFile >= 0);
2377
2378         /*
2379          * posix_fadvise is problematic on many platforms: on older x86 Linux it
2380          * just dumps core, and there are reports of problems on PPC platforms as
2381          * well.  The following is therefore disabled for the time being. We could
2382          * consider some kind of configure test to see if it's safe to use, but
2383          * since we lack hard evidence that there's any useful performance gain to
2384          * be had, spending time on that seems unprofitable for now.
2385          */
2386 #ifdef NOT_USED
2387
2388         /*
2389          * WAL segment files will not be re-read in normal operation, so we advise
2390          * OS to release any cached pages.      But do not do so if WAL archiving is
2391          * active, because archiver process could use the cache to read the WAL
2392          * segment.
2393          *
2394          * While O_DIRECT works for O_SYNC, posix_fadvise() works for fsync() and
2395          * O_SYNC, and some platforms only have posix_fadvise().
2396          */
2397 #if defined(HAVE_DECL_POSIX_FADVISE) && defined(POSIX_FADV_DONTNEED)
2398         if (!XLogArchivingActive())
2399                 posix_fadvise(openLogFile, 0, 0, POSIX_FADV_DONTNEED);
2400 #endif
2401 #endif   /* NOT_USED */
2402
2403         if (close(openLogFile))
2404                 ereport(PANIC,
2405                                 (errcode_for_file_access(),
2406                                  errmsg("could not close log file %u, segment %u: %m",
2407                                                 openLogId, openLogSeg)));
2408         openLogFile = -1;
2409 }
2410
2411 /*
2412  * Attempt to retrieve the specified file from off-line archival storage.
2413  * If successful, fill "path" with its complete path (note that this will be
2414  * a temp file name that doesn't follow the normal naming convention), and
2415  * return TRUE.
2416  *
2417  * If not successful, fill "path" with the name of the normal on-line file
2418  * (which may or may not actually exist, but we'll try to use it), and return
2419  * FALSE.
2420  *
2421  * For fixed-size files, the caller may pass the expected size as an
2422  * additional crosscheck on successful recovery.  If the file size is not
2423  * known, set expectedSize = 0.
2424  */
2425 static bool
2426 RestoreArchivedFile(char *path, const char *xlogfname,
2427                                         const char *recovername, off_t expectedSize)
2428 {
2429         char            xlogpath[MAXPGPATH];
2430         char            xlogRestoreCmd[MAXPGPATH];
2431         char            lastRestartPointFname[MAXPGPATH];
2432         char       *dp;
2433         char       *endp;
2434         const char *sp;
2435         int                     rc;
2436         bool            signaled;
2437         struct stat stat_buf;
2438         uint32          restartLog;
2439         uint32          restartSeg;
2440
2441         /*
2442          * When doing archive recovery, we always prefer an archived log file even
2443          * if a file of the same name exists in XLOGDIR.  The reason is that the
2444          * file in XLOGDIR could be an old, un-filled or partly-filled version
2445          * that was copied and restored as part of backing up $PGDATA.
2446          *
2447          * We could try to optimize this slightly by checking the local copy
2448          * lastchange timestamp against the archived copy, but we have no API to
2449          * do this, nor can we guarantee that the lastchange timestamp was
2450          * preserved correctly when we copied to archive. Our aim is robustness,
2451          * so we elect not to do this.
2452          *
2453          * If we cannot obtain the log file from the archive, however, we will try
2454          * to use the XLOGDIR file if it exists.  This is so that we can make use
2455          * of log segments that weren't yet transferred to the archive.
2456          *
2457          * Notice that we don't actually overwrite any files when we copy back
2458          * from archive because the recoveryRestoreCommand may inadvertently
2459          * restore inappropriate xlogs, or they may be corrupt, so we may wish to
2460          * fallback to the segments remaining in current XLOGDIR later. The
2461          * copy-from-archive filename is always the same, ensuring that we don't
2462          * run out of disk space on long recoveries.
2463          */
2464         snprintf(xlogpath, MAXPGPATH, XLOGDIR "/%s", recovername);
2465
2466         /*
2467          * Make sure there is no existing file named recovername.
2468          */
2469         if (stat(xlogpath, &stat_buf) != 0)
2470         {
2471                 if (errno != ENOENT)
2472                         ereport(FATAL,
2473                                         (errcode_for_file_access(),
2474                                          errmsg("could not stat file \"%s\": %m",
2475                                                         xlogpath)));
2476         }
2477         else
2478         {
2479                 if (unlink(xlogpath) != 0)
2480                         ereport(FATAL,
2481                                         (errcode_for_file_access(),
2482                                          errmsg("could not remove file \"%s\": %m",
2483                                                         xlogpath)));
2484         }
2485
2486         /*
2487          * construct the command to be executed
2488          */
2489         dp = xlogRestoreCmd;
2490         endp = xlogRestoreCmd + MAXPGPATH - 1;
2491         *endp = '\0';
2492
2493         for (sp = recoveryRestoreCommand; *sp; sp++)
2494         {
2495                 if (*sp == '%')
2496                 {
2497                         switch (sp[1])
2498                         {
2499                                 case 'p':
2500                                         /* %p: relative path of target file */
2501                                         sp++;
2502                                         StrNCpy(dp, xlogpath, endp - dp);
2503                                         make_native_path(dp);
2504                                         dp += strlen(dp);
2505                                         break;
2506                                 case 'f':
2507                                         /* %f: filename of desired file */
2508                                         sp++;
2509                                         StrNCpy(dp, xlogfname, endp - dp);
2510                                         dp += strlen(dp);
2511                                         break;
2512                                 case 'r':
2513                                         /* %r: filename of last restartpoint */
2514                                         sp++;
2515                                         XLByteToSeg(ControlFile->checkPointCopy.redo,
2516                                                                 restartLog, restartSeg);
2517                                         XLogFileName(lastRestartPointFname,
2518                                                                  ControlFile->checkPointCopy.ThisTimeLineID,
2519                                                                  restartLog, restartSeg);
2520                                         StrNCpy(dp, lastRestartPointFname, endp - dp);
2521                                         dp += strlen(dp);
2522                                         break;
2523                                 case '%':
2524                                         /* convert %% to a single % */
2525                                         sp++;
2526                                         if (dp < endp)
2527                                                 *dp++ = *sp;
2528                                         break;
2529                                 default:
2530                                         /* otherwise treat the % as not special */
2531                                         if (dp < endp)
2532                                                 *dp++ = *sp;
2533                                         break;
2534                         }
2535                 }
2536                 else
2537                 {
2538                         if (dp < endp)
2539                                 *dp++ = *sp;
2540                 }
2541         }
2542         *dp = '\0';
2543
2544         ereport(DEBUG3,
2545                         (errmsg_internal("executing restore command \"%s\"",
2546                                                          xlogRestoreCmd)));
2547
2548         /*
2549          * Copy xlog from archival storage to XLOGDIR
2550          */
2551         rc = system(xlogRestoreCmd);
2552         if (rc == 0)
2553         {
2554                 /*
2555                  * command apparently succeeded, but let's make sure the file is
2556                  * really there now and has the correct size.
2557                  *
2558                  * XXX I made wrong-size a fatal error to ensure the DBA would notice
2559                  * it, but is that too strong?  We could try to plow ahead with a
2560                  * local copy of the file ... but the problem is that there probably
2561                  * isn't one, and we'd incorrectly conclude we've reached the end of
2562                  * WAL and we're done recovering ...
2563                  */
2564                 if (stat(xlogpath, &stat_buf) == 0)
2565                 {
2566                         if (expectedSize > 0 && stat_buf.st_size != expectedSize)
2567                                 ereport(FATAL,
2568                                                 (errmsg("archive file \"%s\" has wrong size: %lu instead of %lu",
2569                                                                 xlogfname,
2570                                                                 (unsigned long) stat_buf.st_size,
2571                                                                 (unsigned long) expectedSize)));
2572                         else
2573                         {
2574                                 ereport(LOG,
2575                                                 (errmsg("restored log file \"%s\" from archive",
2576                                                                 xlogfname)));
2577                                 strcpy(path, xlogpath);
2578                                 return true;
2579                         }
2580                 }
2581                 else
2582                 {
2583                         /* stat failed */
2584                         if (errno != ENOENT)
2585                                 ereport(FATAL,
2586                                                 (errcode_for_file_access(),
2587                                                  errmsg("could not stat file \"%s\": %m",
2588                                                                 xlogpath)));
2589                 }
2590         }
2591
2592         /*
2593          * Remember, we rollforward UNTIL the restore fails so failure here is
2594          * just part of the process... that makes it difficult to determine
2595          * whether the restore failed because there isn't an archive to restore,
2596          * or because the administrator has specified the restore program
2597          * incorrectly.  We have to assume the former.
2598          *
2599          * However, if the failure was due to any sort of signal, it's best to
2600          * punt and abort recovery.  (If we "return false" here, upper levels will
2601          * assume that recovery is complete and start up the database!) It's
2602          * essential to abort on child SIGINT and SIGQUIT, because per spec
2603          * system() ignores SIGINT and SIGQUIT while waiting; if we see one of
2604          * those it's a good bet we should have gotten it too.  Aborting on other
2605          * signals such as SIGTERM seems a good idea as well.
2606          *
2607          * Per the Single Unix Spec, shells report exit status > 128 when a called
2608          * command died on a signal.  Also, 126 and 127 are used to report
2609          * problems such as an unfindable command; treat those as fatal errors
2610          * too.
2611          */
2612         signaled = WIFSIGNALED(rc) || WEXITSTATUS(rc) > 125;
2613
2614         ereport(signaled ? FATAL : DEBUG2,
2615                 (errmsg("could not restore file \"%s\" from archive: return code %d",
2616                                 xlogfname, rc)));
2617
2618         /*
2619          * if an archived file is not available, there might still be a version of
2620          * this file in XLOGDIR, so return that as the filename to open.
2621          *
2622          * In many recovery scenarios we expect this to fail also, but if so that
2623          * just means we've reached the end of WAL.
2624          */
2625         snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlogfname);
2626         return false;
2627 }
2628
2629 /*
2630  * Preallocate log files beyond the specified log endpoint.
2631  *
2632  * XXX this is currently extremely conservative, since it forces only one
2633  * future log segment to exist, and even that only if we are 75% done with
2634  * the current one.  This is only appropriate for very low-WAL-volume systems.
2635  * High-volume systems will be OK once they've built up a sufficient set of
2636  * recycled log segments, but the startup transient is likely to include
2637  * a lot of segment creations by foreground processes, which is not so good.
2638  */
2639 static void
2640 PreallocXlogFiles(XLogRecPtr endptr)
2641 {
2642         uint32          _logId;
2643         uint32          _logSeg;
2644         int                     lf;
2645         bool            use_existent;
2646
2647         XLByteToPrevSeg(endptr, _logId, _logSeg);
2648         if ((endptr.xrecoff - 1) % XLogSegSize >=
2649                 (uint32) (0.75 * XLogSegSize))
2650         {
2651                 NextLogSeg(_logId, _logSeg);
2652                 use_existent = true;
2653                 lf = XLogFileInit(_logId, _logSeg, &use_existent, true);
2654                 close(lf);
2655                 if (!use_existent)
2656                         CheckpointStats.ckpt_segs_added++;
2657         }
2658 }
2659
2660 /*
2661  * Recycle or remove all log files older or equal to passed log/seg#
2662  *
2663  * endptr is current (or recent) end of xlog; this is used to determine
2664  * whether we want to recycle rather than delete no-longer-wanted log files.
2665  */
2666 static void
2667 RemoveOldXlogFiles(uint32 log, uint32 seg, XLogRecPtr endptr)
2668 {
2669         uint32          endlogId;
2670         uint32          endlogSeg;
2671         int                     max_advance;
2672         DIR                *xldir;
2673         struct dirent *xlde;
2674         char            lastoff[MAXFNAMELEN];
2675         char            path[MAXPGPATH];
2676
2677         /*
2678          * Initialize info about where to try to recycle to.  We allow recycling
2679          * segments up to XLOGfileslop segments beyond the current XLOG location.
2680          */
2681         XLByteToPrevSeg(endptr, endlogId, endlogSeg);
2682         max_advance = XLOGfileslop;
2683
2684         xldir = AllocateDir(XLOGDIR);
2685         if (xldir == NULL)
2686                 ereport(ERROR,
2687                                 (errcode_for_file_access(),
2688                                  errmsg("could not open transaction log directory \"%s\": %m",
2689                                                 XLOGDIR)));
2690
2691         XLogFileName(lastoff, ThisTimeLineID, log, seg);
2692
2693         while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
2694         {
2695                 /*
2696                  * We ignore the timeline part of the XLOG segment identifiers in
2697                  * deciding whether a segment is still needed.  This ensures that we
2698                  * won't prematurely remove a segment from a parent timeline. We could
2699                  * probably be a little more proactive about removing segments of
2700                  * non-parent timelines, but that would be a whole lot more
2701                  * complicated.
2702                  *
2703                  * We use the alphanumeric sorting property of the filenames to decide
2704                  * which ones are earlier than the lastoff segment.
2705                  */
2706                 if (strlen(xlde->d_name) == 24 &&
2707                         strspn(xlde->d_name, "0123456789ABCDEF") == 24 &&
2708                         strcmp(xlde->d_name + 8, lastoff + 8) <= 0)
2709                 {
2710                         if (XLogArchiveCheckDone(xlde->d_name, true))
2711                         {
2712                                 snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
2713
2714                                 /*
2715                                  * Before deleting the file, see if it can be recycled as a
2716                                  * future log segment.
2717                                  */
2718                                 if (InstallXLogFileSegment(&endlogId, &endlogSeg, path,
2719                                                                                    true, &max_advance,
2720                                                                                    true))
2721                                 {
2722                                         ereport(DEBUG2,
2723                                                         (errmsg("recycled transaction log file \"%s\"",
2724                                                                         xlde->d_name)));
2725                                         CheckpointStats.ckpt_segs_recycled++;
2726                                         /* Needn't recheck that slot on future iterations */
2727                                         if (max_advance > 0)
2728                                         {
2729                                                 NextLogSeg(endlogId, endlogSeg);
2730                                                 max_advance--;
2731                                         }
2732                                 }
2733                                 else
2734                                 {
2735                                         /* No need for any more future segments... */
2736                                         ereport(DEBUG2,
2737                                                         (errmsg("removing transaction log file \"%s\"",
2738                                                                         xlde->d_name)));
2739                                         unlink(path);
2740                                         CheckpointStats.ckpt_segs_removed++;
2741                                 }
2742
2743                                 XLogArchiveCleanup(xlde->d_name);
2744                         }
2745                 }
2746         }
2747
2748         FreeDir(xldir);
2749 }
2750
2751 /*
2752  * Remove previous backup history files.  This also retries creation of
2753  * .ready files for any backup history files for which XLogArchiveNotify
2754  * failed earlier.
2755  */
2756 static void
2757 CleanupBackupHistory(void)
2758 {
2759         DIR                *xldir;
2760         struct dirent *xlde;
2761         char            path[MAXPGPATH];
2762
2763         xldir = AllocateDir(XLOGDIR);
2764         if (xldir == NULL)
2765                 ereport(ERROR,
2766                                 (errcode_for_file_access(),
2767                                  errmsg("could not open transaction log directory \"%s\": %m",
2768                                                 XLOGDIR)));
2769
2770         while ((xlde = ReadDir(xldir, XLOGDIR)) != NULL)
2771         {
2772                 if (strlen(xlde->d_name) > 24 &&
2773                         strspn(xlde->d_name, "0123456789ABCDEF") == 24 &&
2774                         strcmp(xlde->d_name + strlen(xlde->d_name) - strlen(".backup"),
2775                                    ".backup") == 0)
2776                 {
2777                         if (XLogArchiveCheckDone(xlde->d_name, true))
2778                         {
2779                                 ereport(DEBUG2,
2780                                 (errmsg("removing transaction log backup history file \"%s\"",
2781                                                 xlde->d_name)));
2782                                 snprintf(path, MAXPGPATH, XLOGDIR "/%s", xlde->d_name);
2783                                 unlink(path);
2784                                 XLogArchiveCleanup(xlde->d_name);
2785                         }
2786                 }
2787         }
2788
2789         FreeDir(xldir);
2790 }
2791
2792 /*
2793  * Restore the backup blocks present in an XLOG record, if any.
2794  *
2795  * We assume all of the record has been read into memory at *record.
2796  *
2797  * Note: when a backup block is available in XLOG, we restore it
2798  * unconditionally, even if the page in the database appears newer.
2799  * This is to protect ourselves against database pages that were partially
2800  * or incorrectly written during a crash.  We assume that the XLOG data
2801  * must be good because it has passed a CRC check, while the database
2802  * page might not be.  This will force us to replay all subsequent
2803  * modifications of the page that appear in XLOG, rather than possibly
2804  * ignoring them as already applied, but that's not a huge drawback.
2805  */
2806 static void
2807 RestoreBkpBlocks(XLogRecord *record, XLogRecPtr lsn)
2808 {
2809         Relation        reln;
2810         Buffer          buffer;
2811         Page            page;
2812         BkpBlock        bkpb;
2813         char       *blk;
2814         int                     i;
2815
2816         blk = (char *) XLogRecGetData(record) + record->xl_len;
2817         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
2818         {
2819                 if (!(record->xl_info & XLR_SET_BKP_BLOCK(i)))
2820                         continue;
2821
2822                 memcpy(&bkpb, blk, sizeof(BkpBlock));
2823                 blk += sizeof(BkpBlock);
2824
2825                 reln = XLogOpenRelation(bkpb.node);
2826                 buffer = XLogReadBuffer(reln, bkpb.block, true);
2827                 Assert(BufferIsValid(buffer));
2828                 page = (Page) BufferGetPage(buffer);
2829
2830                 if (bkpb.hole_length == 0)
2831                 {
2832                         memcpy((char *) page, blk, BLCKSZ);
2833                 }
2834                 else
2835                 {
2836                         /* must zero-fill the hole */
2837                         MemSet((char *) page, 0, BLCKSZ);
2838                         memcpy((char *) page, blk, bkpb.hole_offset);
2839                         memcpy((char *) page + (bkpb.hole_offset + bkpb.hole_length),
2840                                    blk + bkpb.hole_offset,
2841                                    BLCKSZ - (bkpb.hole_offset + bkpb.hole_length));
2842                 }
2843
2844                 PageSetLSN(page, lsn);
2845                 PageSetTLI(page, ThisTimeLineID);
2846                 MarkBufferDirty(buffer);
2847                 UnlockReleaseBuffer(buffer);
2848
2849                 blk += BLCKSZ - bkpb.hole_length;
2850         }
2851 }
2852
2853 /*
2854  * CRC-check an XLOG record.  We do not believe the contents of an XLOG
2855  * record (other than to the minimal extent of computing the amount of
2856  * data to read in) until we've checked the CRCs.
2857  *
2858  * We assume all of the record has been read into memory at *record.
2859  */
2860 static bool
2861 RecordIsValid(XLogRecord *record, XLogRecPtr recptr, int emode)
2862 {
2863         pg_crc32        crc;
2864         int                     i;
2865         uint32          len = record->xl_len;
2866         BkpBlock        bkpb;
2867         char       *blk;
2868
2869         /* First the rmgr data */
2870         INIT_CRC32(crc);
2871         COMP_CRC32(crc, XLogRecGetData(record), len);
2872
2873         /* Add in the backup blocks, if any */
2874         blk = (char *) XLogRecGetData(record) + len;
2875         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
2876         {
2877                 uint32          blen;
2878
2879                 if (!(record->xl_info & XLR_SET_BKP_BLOCK(i)))
2880                         continue;
2881
2882                 memcpy(&bkpb, blk, sizeof(BkpBlock));
2883                 if (bkpb.hole_offset + bkpb.hole_length > BLCKSZ)
2884                 {
2885                         ereport(emode,
2886                                         (errmsg("incorrect hole size in record at %X/%X",
2887                                                         recptr.xlogid, recptr.xrecoff)));
2888                         return false;
2889                 }
2890                 blen = sizeof(BkpBlock) + BLCKSZ - bkpb.hole_length;
2891                 COMP_CRC32(crc, blk, blen);
2892                 blk += blen;
2893         }
2894
2895         /* Check that xl_tot_len agrees with our calculation */
2896         if (blk != (char *) record + record->xl_tot_len)
2897         {
2898                 ereport(emode,
2899                                 (errmsg("incorrect total length in record at %X/%X",
2900                                                 recptr.xlogid, recptr.xrecoff)));
2901                 return false;
2902         }
2903
2904         /* Finally include the record header */
2905         COMP_CRC32(crc, (char *) record + sizeof(pg_crc32),
2906                            SizeOfXLogRecord - sizeof(pg_crc32));
2907         FIN_CRC32(crc);
2908
2909         if (!EQ_CRC32(record->xl_crc, crc))
2910         {
2911                 ereport(emode,
2912                 (errmsg("incorrect resource manager data checksum in record at %X/%X",
2913                                 recptr.xlogid, recptr.xrecoff)));
2914                 return false;
2915         }
2916
2917         return true;
2918 }
2919
2920 /*
2921  * Attempt to read an XLOG record.
2922  *
2923  * If RecPtr is not NULL, try to read a record at that position.  Otherwise
2924  * try to read a record just after the last one previously read.
2925  *
2926  * If no valid record is available, returns NULL, or fails if emode is PANIC.
2927  * (emode must be either PANIC or LOG.)
2928  *
2929  * The record is copied into readRecordBuf, so that on successful return,
2930  * the returned record pointer always points there.
2931  */
2932 static XLogRecord *
2933 ReadRecord(XLogRecPtr *RecPtr, int emode)
2934 {
2935         XLogRecord *record;
2936         char       *buffer;
2937         XLogRecPtr      tmpRecPtr = EndRecPtr;
2938         bool            randAccess = false;
2939         uint32          len,
2940                                 total_len;
2941         uint32          targetPageOff;
2942         uint32          targetRecOff;
2943         uint32          pageHeaderSize;
2944
2945         if (readBuf == NULL)
2946         {
2947                 /*
2948                  * First time through, permanently allocate readBuf.  We do it this
2949                  * way, rather than just making a static array, for two reasons: (1)
2950                  * no need to waste the storage in most instantiations of the backend;
2951                  * (2) a static char array isn't guaranteed to have any particular
2952                  * alignment, whereas malloc() will provide MAXALIGN'd storage.
2953                  */
2954                 readBuf = (char *) malloc(XLOG_BLCKSZ);
2955                 Assert(readBuf != NULL);
2956         }
2957
2958         if (RecPtr == NULL)
2959         {
2960                 RecPtr = &tmpRecPtr;
2961                 /* fast case if next record is on same page */
2962                 if (nextRecord != NULL)
2963                 {
2964                         record = nextRecord;
2965                         goto got_record;
2966                 }
2967                 /* align old recptr to next page */
2968                 if (tmpRecPtr.xrecoff % XLOG_BLCKSZ != 0)
2969                         tmpRecPtr.xrecoff += (XLOG_BLCKSZ - tmpRecPtr.xrecoff % XLOG_BLCKSZ);
2970                 if (tmpRecPtr.xrecoff >= XLogFileSize)
2971                 {
2972                         (tmpRecPtr.xlogid)++;
2973                         tmpRecPtr.xrecoff = 0;
2974                 }
2975                 /* We will account for page header size below */
2976         }
2977         else
2978         {
2979                 if (!XRecOffIsValid(RecPtr->xrecoff))
2980                         ereport(PANIC,
2981                                         (errmsg("invalid record offset at %X/%X",
2982                                                         RecPtr->xlogid, RecPtr->xrecoff)));
2983
2984                 /*
2985                  * Since we are going to a random position in WAL, forget any prior
2986                  * state about what timeline we were in, and allow it to be any
2987                  * timeline in expectedTLIs.  We also set a flag to allow curFileTLI
2988                  * to go backwards (but we can't reset that variable right here, since
2989                  * we might not change files at all).
2990                  */
2991                 lastPageTLI = 0;                /* see comment in ValidXLOGHeader */
2992                 randAccess = true;              /* allow curFileTLI to go backwards too */
2993         }
2994
2995         if (readFile >= 0 && !XLByteInSeg(*RecPtr, readId, readSeg))
2996         {
2997                 close(readFile);
2998                 readFile = -1;
2999         }
3000         XLByteToSeg(*RecPtr, readId, readSeg);
3001         if (readFile < 0)
3002         {
3003                 /* Now it's okay to reset curFileTLI if random fetch */
3004                 if (randAccess)
3005                         curFileTLI = 0;
3006
3007                 readFile = XLogFileRead(readId, readSeg, emode);
3008                 if (readFile < 0)
3009                         goto next_record_is_invalid;
3010
3011                 /*
3012                  * Whenever switching to a new WAL segment, we read the first page of
3013                  * the file and validate its header, even if that's not where the
3014                  * target record is.  This is so that we can check the additional
3015                  * identification info that is present in the first page's "long"
3016                  * header.
3017                  */
3018                 readOff = 0;
3019                 if (read(readFile, readBuf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
3020                 {
3021                         ereport(emode,
3022                                         (errcode_for_file_access(),
3023                                          errmsg("could not read from log file %u, segment %u, offset %u: %m",
3024                                                         readId, readSeg, readOff)));
3025                         goto next_record_is_invalid;
3026                 }
3027                 if (!ValidXLOGHeader((XLogPageHeader) readBuf, emode))
3028                         goto next_record_is_invalid;
3029         }
3030
3031         targetPageOff = ((RecPtr->xrecoff % XLogSegSize) / XLOG_BLCKSZ) * XLOG_BLCKSZ;
3032         if (readOff != targetPageOff)
3033         {
3034                 readOff = targetPageOff;
3035                 if (lseek(readFile, (off_t) readOff, SEEK_SET) < 0)
3036                 {
3037                         ereport(emode,
3038                                         (errcode_for_file_access(),
3039                                          errmsg("could not seek in log file %u, segment %u to offset %u: %m",
3040                                                         readId, readSeg, readOff)));
3041                         goto next_record_is_invalid;
3042                 }
3043                 if (read(readFile, readBuf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
3044                 {
3045                         ereport(emode,
3046                                         (errcode_for_file_access(),
3047                                          errmsg("could not read from log file %u, segment %u, offset %u: %m",
3048                                                         readId, readSeg, readOff)));
3049                         goto next_record_is_invalid;
3050                 }
3051                 if (!ValidXLOGHeader((XLogPageHeader) readBuf, emode))
3052                         goto next_record_is_invalid;
3053         }
3054         pageHeaderSize = XLogPageHeaderSize((XLogPageHeader) readBuf);
3055         targetRecOff = RecPtr->xrecoff % XLOG_BLCKSZ;
3056         if (targetRecOff == 0)
3057         {
3058                 /*
3059                  * Can only get here in the continuing-from-prev-page case, because
3060                  * XRecOffIsValid eliminated the zero-page-offset case otherwise. Need
3061                  * to skip over the new page's header.
3062                  */
3063                 tmpRecPtr.xrecoff += pageHeaderSize;
3064                 targetRecOff = pageHeaderSize;
3065         }
3066         else if (targetRecOff < pageHeaderSize)
3067         {
3068                 ereport(emode,
3069                                 (errmsg("invalid record offset at %X/%X",
3070                                                 RecPtr->xlogid, RecPtr->xrecoff)));
3071                 goto next_record_is_invalid;
3072         }
3073         if ((((XLogPageHeader) readBuf)->xlp_info & XLP_FIRST_IS_CONTRECORD) &&
3074                 targetRecOff == pageHeaderSize)
3075         {
3076                 ereport(emode,
3077                                 (errmsg("contrecord is requested by %X/%X",
3078                                                 RecPtr->xlogid, RecPtr->xrecoff)));
3079                 goto next_record_is_invalid;
3080         }
3081         record = (XLogRecord *) ((char *) readBuf + RecPtr->xrecoff % XLOG_BLCKSZ);
3082
3083 got_record:;
3084
3085         /*
3086          * xl_len == 0 is bad data for everything except XLOG SWITCH, where it is
3087          * required.
3088          */
3089         if (record->xl_rmid == RM_XLOG_ID && record->xl_info == XLOG_SWITCH)
3090         {
3091                 if (record->xl_len != 0)
3092                 {
3093                         ereport(emode,
3094                                         (errmsg("invalid xlog switch record at %X/%X",
3095                                                         RecPtr->xlogid, RecPtr->xrecoff)));
3096                         goto next_record_is_invalid;
3097                 }
3098         }
3099         else if (record->xl_len == 0)
3100         {
3101                 ereport(emode,
3102                                 (errmsg("record with zero length at %X/%X",
3103                                                 RecPtr->xlogid, RecPtr->xrecoff)));
3104                 goto next_record_is_invalid;
3105         }
3106         if (record->xl_tot_len < SizeOfXLogRecord + record->xl_len ||
3107                 record->xl_tot_len > SizeOfXLogRecord + record->xl_len +
3108                 XLR_MAX_BKP_BLOCKS * (sizeof(BkpBlock) + BLCKSZ))
3109         {
3110                 ereport(emode,
3111                                 (errmsg("invalid record length at %X/%X",
3112                                                 RecPtr->xlogid, RecPtr->xrecoff)));
3113                 goto next_record_is_invalid;
3114         }
3115         if (record->xl_rmid > RM_MAX_ID)
3116         {
3117                 ereport(emode,
3118                                 (errmsg("invalid resource manager ID %u at %X/%X",
3119                                                 record->xl_rmid, RecPtr->xlogid, RecPtr->xrecoff)));
3120                 goto next_record_is_invalid;
3121         }
3122         if (randAccess)
3123         {
3124                 /*
3125                  * We can't exactly verify the prev-link, but surely it should be less
3126                  * than the record's own address.
3127                  */
3128                 if (!XLByteLT(record->xl_prev, *RecPtr))
3129                 {
3130                         ereport(emode,
3131                                         (errmsg("record with incorrect prev-link %X/%X at %X/%X",
3132                                                         record->xl_prev.xlogid, record->xl_prev.xrecoff,
3133                                                         RecPtr->xlogid, RecPtr->xrecoff)));
3134                         goto next_record_is_invalid;
3135                 }
3136         }
3137         else
3138         {
3139                 /*
3140                  * Record's prev-link should exactly match our previous location. This
3141                  * check guards against torn WAL pages where a stale but valid-looking
3142                  * WAL record starts on a sector boundary.
3143                  */
3144                 if (!XLByteEQ(record->xl_prev, ReadRecPtr))
3145                 {
3146                         ereport(emode,
3147                                         (errmsg("record with incorrect prev-link %X/%X at %X/%X",
3148                                                         record->xl_prev.xlogid, record->xl_prev.xrecoff,
3149                                                         RecPtr->xlogid, RecPtr->xrecoff)));
3150                         goto next_record_is_invalid;
3151                 }
3152         }
3153
3154         /*
3155          * Allocate or enlarge readRecordBuf as needed.  To avoid useless small
3156          * increases, round its size to a multiple of XLOG_BLCKSZ, and make sure
3157          * it's at least 4*Max(BLCKSZ, XLOG_BLCKSZ) to start with.  (That is
3158          * enough for all "normal" records, but very large commit or abort records
3159          * might need more space.)
3160          */
3161         total_len = record->xl_tot_len;
3162         if (total_len > readRecordBufSize)
3163         {
3164                 uint32          newSize = total_len;
3165
3166                 newSize += XLOG_BLCKSZ - (newSize % XLOG_BLCKSZ);
3167                 newSize = Max(newSize, 4 * Max(BLCKSZ, XLOG_BLCKSZ));
3168                 if (readRecordBuf)
3169                         free(readRecordBuf);
3170                 readRecordBuf = (char *) malloc(newSize);
3171                 if (!readRecordBuf)
3172                 {
3173                         readRecordBufSize = 0;
3174                         /* We treat this as a "bogus data" condition */
3175                         ereport(emode,
3176                                         (errmsg("record length %u at %X/%X too long",
3177                                                         total_len, RecPtr->xlogid, RecPtr->xrecoff)));
3178                         goto next_record_is_invalid;
3179                 }
3180                 readRecordBufSize = newSize;
3181         }
3182
3183         buffer = readRecordBuf;
3184         nextRecord = NULL;
3185         len = XLOG_BLCKSZ - RecPtr->xrecoff % XLOG_BLCKSZ;
3186         if (total_len > len)
3187         {
3188                 /* Need to reassemble record */
3189                 XLogContRecord *contrecord;
3190                 uint32          gotlen = len;
3191
3192                 memcpy(buffer, record, len);
3193                 record = (XLogRecord *) buffer;
3194                 buffer += len;
3195                 for (;;)
3196                 {
3197                         readOff += XLOG_BLCKSZ;
3198                         if (readOff >= XLogSegSize)
3199                         {
3200                                 close(readFile);
3201                                 readFile = -1;
3202                                 NextLogSeg(readId, readSeg);
3203                                 readFile = XLogFileRead(readId, readSeg, emode);
3204                                 if (readFile < 0)
3205                                         goto next_record_is_invalid;
3206                                 readOff = 0;
3207                         }
3208                         if (read(readFile, readBuf, XLOG_BLCKSZ) != XLOG_BLCKSZ)
3209                         {
3210                                 ereport(emode,
3211                                                 (errcode_for_file_access(),
3212                                                  errmsg("could not read from log file %u, segment %u, offset %u: %m",
3213                                                                 readId, readSeg, readOff)));
3214                                 goto next_record_is_invalid;
3215                         }
3216                         if (!ValidXLOGHeader((XLogPageHeader) readBuf, emode))
3217                                 goto next_record_is_invalid;
3218                         if (!(((XLogPageHeader) readBuf)->xlp_info & XLP_FIRST_IS_CONTRECORD))
3219                         {
3220                                 ereport(emode,
3221                                                 (errmsg("there is no contrecord flag in log file %u, segment %u, offset %u",
3222                                                                 readId, readSeg, readOff)));
3223                                 goto next_record_is_invalid;
3224                         }
3225                         pageHeaderSize = XLogPageHeaderSize((XLogPageHeader) readBuf);
3226                         contrecord = (XLogContRecord *) ((char *) readBuf + pageHeaderSize);
3227                         if (contrecord->xl_rem_len == 0 ||
3228                                 total_len != (contrecord->xl_rem_len + gotlen))
3229                         {
3230                                 ereport(emode,
3231                                                 (errmsg("invalid contrecord length %u in log file %u, segment %u, offset %u",
3232                                                                 contrecord->xl_rem_len,
3233                                                                 readId, readSeg, readOff)));
3234                                 goto next_record_is_invalid;
3235                         }
3236                         len = XLOG_BLCKSZ - pageHeaderSize - SizeOfXLogContRecord;
3237                         if (contrecord->xl_rem_len > len)
3238                         {
3239                                 memcpy(buffer, (char *) contrecord + SizeOfXLogContRecord, len);
3240                                 gotlen += len;
3241                                 buffer += len;
3242                                 continue;
3243                         }
3244                         memcpy(buffer, (char *) contrecord + SizeOfXLogContRecord,
3245                                    contrecord->xl_rem_len);
3246                         break;
3247                 }
3248                 if (!RecordIsValid(record, *RecPtr, emode))
3249                         goto next_record_is_invalid;
3250                 pageHeaderSize = XLogPageHeaderSize((XLogPageHeader) readBuf);
3251                 if (XLOG_BLCKSZ - SizeOfXLogRecord >= pageHeaderSize +
3252                         MAXALIGN(SizeOfXLogContRecord + contrecord->xl_rem_len))
3253                 {
3254                         nextRecord = (XLogRecord *) ((char *) contrecord +
3255                                         MAXALIGN(SizeOfXLogContRecord + contrecord->xl_rem_len));
3256                 }
3257                 EndRecPtr.xlogid = readId;
3258                 EndRecPtr.xrecoff = readSeg * XLogSegSize + readOff +
3259                         pageHeaderSize +
3260                         MAXALIGN(SizeOfXLogContRecord + contrecord->xl_rem_len);
3261                 ReadRecPtr = *RecPtr;
3262                 /* needn't worry about XLOG SWITCH, it can't cross page boundaries */
3263                 return record;
3264         }
3265
3266         /* Record does not cross a page boundary */
3267         if (!RecordIsValid(record, *RecPtr, emode))
3268                 goto next_record_is_invalid;
3269         if (XLOG_BLCKSZ - SizeOfXLogRecord >= RecPtr->xrecoff % XLOG_BLCKSZ +
3270                 MAXALIGN(total_len))
3271                 nextRecord = (XLogRecord *) ((char *) record + MAXALIGN(total_len));
3272         EndRecPtr.xlogid = RecPtr->xlogid;
3273         EndRecPtr.xrecoff = RecPtr->xrecoff + MAXALIGN(total_len);
3274         ReadRecPtr = *RecPtr;
3275         memcpy(buffer, record, total_len);
3276
3277         /*
3278          * Special processing if it's an XLOG SWITCH record
3279          */
3280         if (record->xl_rmid == RM_XLOG_ID && record->xl_info == XLOG_SWITCH)
3281         {
3282                 /* Pretend it extends to end of segment */
3283                 EndRecPtr.xrecoff += XLogSegSize - 1;
3284                 EndRecPtr.xrecoff -= EndRecPtr.xrecoff % XLogSegSize;
3285                 nextRecord = NULL;              /* definitely not on same page */
3286
3287                 /*
3288                  * Pretend that readBuf contains the last page of the segment. This is
3289                  * just to avoid Assert failure in StartupXLOG if XLOG ends with this
3290                  * segment.
3291                  */
3292                 readOff = XLogSegSize - XLOG_BLCKSZ;
3293         }
3294         return (XLogRecord *) buffer;
3295
3296 next_record_is_invalid:;
3297         close(readFile);
3298         readFile = -1;
3299         nextRecord = NULL;
3300         return NULL;
3301 }
3302
3303 /*
3304  * Check whether the xlog header of a page just read in looks valid.
3305  *
3306  * This is just a convenience subroutine to avoid duplicated code in
3307  * ReadRecord.  It's not intended for use from anywhere else.
3308  */
3309 static bool
3310 ValidXLOGHeader(XLogPageHeader hdr, int emode)
3311 {
3312         XLogRecPtr      recaddr;
3313
3314         if (hdr->xlp_magic != XLOG_PAGE_MAGIC)
3315         {
3316                 ereport(emode,
3317                                 (errmsg("invalid magic number %04X in log file %u, segment %u, offset %u",
3318                                                 hdr->xlp_magic, readId, readSeg, readOff)));
3319                 return false;
3320         }
3321         if ((hdr->xlp_info & ~XLP_ALL_FLAGS) != 0)
3322         {
3323                 ereport(emode,
3324                                 (errmsg("invalid info bits %04X in log file %u, segment %u, offset %u",
3325                                                 hdr->xlp_info, readId, readSeg, readOff)));
3326                 return false;
3327         }
3328         if (hdr->xlp_info & XLP_LONG_HEADER)
3329         {
3330                 XLogLongPageHeader longhdr = (XLogLongPageHeader) hdr;
3331
3332                 if (longhdr->xlp_sysid != ControlFile->system_identifier)
3333                 {
3334                         char            fhdrident_str[32];
3335                         char            sysident_str[32];
3336
3337                         /*
3338                          * Format sysids separately to keep platform-dependent format code
3339                          * out of the translatable message string.
3340                          */
3341                         snprintf(fhdrident_str, sizeof(fhdrident_str), UINT64_FORMAT,
3342                                          longhdr->xlp_sysid);
3343                         snprintf(sysident_str, sizeof(sysident_str), UINT64_FORMAT,
3344                                          ControlFile->system_identifier);
3345                         ereport(emode,
3346                                         (errmsg("WAL file is from different system"),
3347                                          errdetail("WAL file SYSID is %s, pg_control SYSID is %s",
3348                                                            fhdrident_str, sysident_str)));
3349                         return false;
3350                 }
3351                 if (longhdr->xlp_seg_size != XLogSegSize)
3352                 {
3353                         ereport(emode,
3354                                         (errmsg("WAL file is from different system"),
3355                                          errdetail("Incorrect XLOG_SEG_SIZE in page header.")));
3356                         return false;
3357                 }
3358                 if (longhdr->xlp_xlog_blcksz != XLOG_BLCKSZ)
3359                 {
3360                         ereport(emode,
3361                                         (errmsg("WAL file is from different system"),
3362                                          errdetail("Incorrect XLOG_BLCKSZ in page header.")));
3363                         return false;
3364                 }
3365         }
3366         else if (readOff == 0)
3367         {
3368                 /* hmm, first page of file doesn't have a long header? */
3369                 ereport(emode,
3370                                 (errmsg("invalid info bits %04X in log file %u, segment %u, offset %u",
3371                                                 hdr->xlp_info, readId, readSeg, readOff)));
3372                 return false;
3373         }
3374
3375         recaddr.xlogid = readId;
3376         recaddr.xrecoff = readSeg * XLogSegSize + readOff;
3377         if (!XLByteEQ(hdr->xlp_pageaddr, recaddr))
3378         {
3379                 ereport(emode,
3380                                 (errmsg("unexpected pageaddr %X/%X in log file %u, segment %u, offset %u",
3381                                                 hdr->xlp_pageaddr.xlogid, hdr->xlp_pageaddr.xrecoff,
3382                                                 readId, readSeg, readOff)));
3383                 return false;
3384         }
3385
3386         /*
3387          * Check page TLI is one of the expected values.
3388          */
3389         if (!list_member_int(expectedTLIs, (int) hdr->xlp_tli))
3390         {
3391                 ereport(emode,
3392                                 (errmsg("unexpected timeline ID %u in log file %u, segment %u, offset %u",
3393                                                 hdr->xlp_tli,
3394                                                 readId, readSeg, readOff)));
3395                 return false;
3396         }
3397
3398         /*
3399          * Since child timelines are always assigned a TLI greater than their
3400          * immediate parent's TLI, we should never see TLI go backwards across
3401          * successive pages of a consistent WAL sequence.
3402          *
3403          * Of course this check should only be applied when advancing sequentially
3404          * across pages; therefore ReadRecord resets lastPageTLI to zero when
3405          * going to a random page.
3406          */
3407         if (hdr->xlp_tli < lastPageTLI)
3408         {
3409                 ereport(emode,
3410                                 (errmsg("out-of-sequence timeline ID %u (after %u) in log file %u, segment %u, offset %u",
3411                                                 hdr->xlp_tli, lastPageTLI,
3412                                                 readId, readSeg, readOff)));
3413                 return false;
3414         }
3415         lastPageTLI = hdr->xlp_tli;
3416         return true;
3417 }
3418
3419 /*
3420  * Try to read a timeline's history file.
3421  *
3422  * If successful, return the list of component TLIs (the given TLI followed by
3423  * its ancestor TLIs).  If we can't find the history file, assume that the
3424  * timeline has no parents, and return a list of just the specified timeline
3425  * ID.
3426  */
3427 static List *
3428 readTimeLineHistory(TimeLineID targetTLI)
3429 {
3430         List       *result;
3431         char            path[MAXPGPATH];
3432         char            histfname[MAXFNAMELEN];
3433         char            fline[MAXPGPATH];
3434         FILE       *fd;
3435
3436         if (InArchiveRecovery)
3437         {
3438                 TLHistoryFileName(histfname, targetTLI);
3439                 RestoreArchivedFile(path, histfname, "RECOVERYHISTORY", 0);
3440         }
3441         else
3442                 TLHistoryFilePath(path, targetTLI);
3443
3444         fd = AllocateFile(path, "r");
3445         if (fd == NULL)
3446         {
3447                 if (errno != ENOENT)
3448                         ereport(FATAL,
3449                                         (errcode_for_file_access(),
3450                                          errmsg("could not open file \"%s\": %m", path)));
3451                 /* Not there, so assume no parents */
3452                 return list_make1_int((int) targetTLI);
3453         }
3454
3455         result = NIL;
3456
3457         /*
3458          * Parse the file...
3459          */
3460         while (fgets(fline, sizeof(fline), fd) != NULL)
3461         {
3462                 /* skip leading whitespace and check for # comment */
3463                 char       *ptr;
3464                 char       *endptr;
3465                 TimeLineID      tli;
3466
3467                 for (ptr = fline; *ptr; ptr++)
3468                 {
3469                         if (!isspace((unsigned char) *ptr))
3470                                 break;
3471                 }
3472                 if (*ptr == '\0' || *ptr == '#')
3473                         continue;
3474
3475                 /* expect a numeric timeline ID as first field of line */
3476                 tli = (TimeLineID) strtoul(ptr, &endptr, 0);
3477                 if (endptr == ptr)
3478                         ereport(FATAL,
3479                                         (errmsg("syntax error in history file: %s", fline),
3480                                          errhint("Expected a numeric timeline ID.")));
3481
3482                 if (result &&
3483                         tli <= (TimeLineID) linitial_int(result))
3484                         ereport(FATAL,
3485                                         (errmsg("invalid data in history file: %s", fline),
3486                                    errhint("Timeline IDs must be in increasing sequence.")));
3487
3488                 /* Build list with newest item first */
3489                 result = lcons_int((int) tli, result);
3490
3491                 /* we ignore the remainder of each line */
3492         }
3493
3494         FreeFile(fd);
3495
3496         if (result &&
3497                 targetTLI <= (TimeLineID) linitial_int(result))
3498                 ereport(FATAL,
3499                                 (errmsg("invalid data in history file \"%s\"", path),
3500                         errhint("Timeline IDs must be less than child timeline's ID.")));
3501
3502         result = lcons_int((int) targetTLI, result);
3503
3504         ereport(DEBUG3,
3505                         (errmsg_internal("history of timeline %u is %s",
3506                                                          targetTLI, nodeToString(result))));
3507
3508         return result;
3509 }
3510
3511 /*
3512  * Probe whether a timeline history file exists for the given timeline ID
3513  */
3514 static bool
3515 existsTimeLineHistory(TimeLineID probeTLI)
3516 {
3517         char            path[MAXPGPATH];
3518         char            histfname[MAXFNAMELEN];
3519         FILE       *fd;
3520
3521         if (InArchiveRecovery)
3522         {
3523                 TLHistoryFileName(histfname, probeTLI);
3524                 RestoreArchivedFile(path, histfname, "RECOVERYHISTORY", 0);
3525         }
3526         else
3527                 TLHistoryFilePath(path, probeTLI);
3528
3529         fd = AllocateFile(path, "r");
3530         if (fd != NULL)
3531         {
3532                 FreeFile(fd);
3533                 return true;
3534         }
3535         else
3536         {
3537                 if (errno != ENOENT)
3538                         ereport(FATAL,
3539                                         (errcode_for_file_access(),
3540                                          errmsg("could not open file \"%s\": %m", path)));
3541                 return false;
3542         }
3543 }
3544
3545 /*
3546  * Find the newest existing timeline, assuming that startTLI exists.
3547  *
3548  * Note: while this is somewhat heuristic, it does positively guarantee
3549  * that (result + 1) is not a known timeline, and therefore it should
3550  * be safe to assign that ID to a new timeline.
3551  */
3552 static TimeLineID
3553 findNewestTimeLine(TimeLineID startTLI)
3554 {
3555         TimeLineID      newestTLI;
3556         TimeLineID      probeTLI;
3557
3558         /*
3559          * The algorithm is just to probe for the existence of timeline history
3560          * files.  XXX is it useful to allow gaps in the sequence?
3561          */
3562         newestTLI = startTLI;
3563
3564         for (probeTLI = startTLI + 1;; probeTLI++)
3565         {
3566                 if (existsTimeLineHistory(probeTLI))
3567                 {
3568                         newestTLI = probeTLI;           /* probeTLI exists */
3569                 }
3570                 else
3571                 {
3572                         /* doesn't exist, assume we're done */
3573                         break;
3574                 }
3575         }
3576
3577         return newestTLI;
3578 }
3579
3580 /*
3581  * Create a new timeline history file.
3582  *
3583  *      newTLI: ID of the new timeline
3584  *      parentTLI: ID of its immediate parent
3585  *      endTLI et al: ID of the last used WAL file, for annotation purposes
3586  *
3587  * Currently this is only used during recovery, and so there are no locking
3588  * considerations.      But we should be just as tense as XLogFileInit to avoid
3589  * emplacing a bogus file.
3590  */
3591 static void
3592 writeTimeLineHistory(TimeLineID newTLI, TimeLineID parentTLI,
3593                                          TimeLineID endTLI, uint32 endLogId, uint32 endLogSeg)
3594 {
3595         char            path[MAXPGPATH];
3596         char            tmppath[MAXPGPATH];
3597         char            histfname[MAXFNAMELEN];
3598         char            xlogfname[MAXFNAMELEN];
3599         char            buffer[BLCKSZ];
3600         int                     srcfd;
3601         int                     fd;
3602         int                     nbytes;
3603
3604         Assert(newTLI > parentTLI); /* else bad selection of newTLI */
3605
3606         /*
3607          * Write into a temp file name.
3608          */
3609         snprintf(tmppath, MAXPGPATH, XLOGDIR "/xlogtemp.%d", (int) getpid());
3610
3611         unlink(tmppath);
3612
3613         /* do not use XLOG_SYNC_BIT here --- want to fsync only at end of fill */
3614         fd = BasicOpenFile(tmppath, O_RDWR | O_CREAT | O_EXCL,
3615                                            S_IRUSR | S_IWUSR);
3616         if (fd < 0)
3617                 ereport(ERROR,
3618                                 (errcode_for_file_access(),
3619                                  errmsg("could not create file \"%s\": %m", tmppath)));
3620
3621         /*
3622          * If a history file exists for the parent, copy it verbatim
3623          */
3624         if (InArchiveRecovery)
3625         {
3626                 TLHistoryFileName(histfname, parentTLI);
3627                 RestoreArchivedFile(path, histfname, "RECOVERYHISTORY", 0);
3628         }
3629         else
3630                 TLHistoryFilePath(path, parentTLI);
3631
3632         srcfd = BasicOpenFile(path, O_RDONLY, 0);
3633         if (srcfd < 0)
3634         {
3635                 if (errno != ENOENT)
3636                         ereport(ERROR,
3637                                         (errcode_for_file_access(),
3638                                          errmsg("could not open file \"%s\": %m", path)));
3639                 /* Not there, so assume parent has no parents */
3640         }
3641         else
3642         {
3643                 for (;;)
3644                 {
3645                         errno = 0;
3646                         nbytes = (int) read(srcfd, buffer, sizeof(buffer));
3647                         if (nbytes < 0 || errno != 0)
3648                                 ereport(ERROR,
3649                                                 (errcode_for_file_access(),
3650                                                  errmsg("could not read file \"%s\": %m", path)));
3651                         if (nbytes == 0)
3652                                 break;
3653                         errno = 0;
3654                         if ((int) write(fd, buffer, nbytes) != nbytes)
3655                         {
3656                                 int                     save_errno = errno;
3657
3658                                 /*
3659                                  * If we fail to make the file, delete it to release disk
3660                                  * space
3661                                  */
3662                                 unlink(tmppath);
3663
3664                                 /*
3665                                  * if write didn't set errno, assume problem is no disk space
3666                                  */
3667                                 errno = save_errno ? save_errno : ENOSPC;
3668
3669                                 ereport(ERROR,
3670                                                 (errcode_for_file_access(),
3671                                          errmsg("could not write to file \"%s\": %m", tmppath)));
3672                         }
3673                 }
3674                 close(srcfd);
3675         }
3676
3677         /*
3678          * Append one line with the details of this timeline split.
3679          *
3680          * If we did have a parent file, insert an extra newline just in case the
3681          * parent file failed to end with one.
3682          */
3683         XLogFileName(xlogfname, endTLI, endLogId, endLogSeg);
3684
3685         snprintf(buffer, sizeof(buffer),
3686                          "%s%u\t%s\t%s transaction %u at %s\n",
3687                          (srcfd < 0) ? "" : "\n",
3688                          parentTLI,
3689                          xlogfname,
3690                          recoveryStopAfter ? "after" : "before",
3691                          recoveryStopXid,
3692                          timestamptz_to_str(recoveryStopTime));
3693
3694         nbytes = strlen(buffer);
3695         errno = 0;
3696         if ((int) write(fd, buffer, nbytes) != nbytes)
3697         {
3698                 int                     save_errno = errno;
3699
3700                 /*
3701                  * If we fail to make the file, delete it to release disk space
3702                  */
3703                 unlink(tmppath);
3704                 /* if write didn't set errno, assume problem is no disk space */
3705                 errno = save_errno ? save_errno : ENOSPC;
3706
3707                 ereport(ERROR,
3708                                 (errcode_for_file_access(),
3709                                  errmsg("could not write to file \"%s\": %m", tmppath)));
3710         }
3711
3712         if (pg_fsync(fd) != 0)
3713                 ereport(ERROR,
3714                                 (errcode_for_file_access(),
3715                                  errmsg("could not fsync file \"%s\": %m", tmppath)));
3716
3717         if (close(fd))
3718                 ereport(ERROR,
3719                                 (errcode_for_file_access(),
3720                                  errmsg("could not close file \"%s\": %m", tmppath)));
3721
3722
3723         /*
3724          * Now move the completed history file into place with its final name.
3725          */
3726         TLHistoryFilePath(path, newTLI);
3727
3728         /*
3729          * Prefer link() to rename() here just to be really sure that we don't
3730          * overwrite an existing logfile.  However, there shouldn't be one, so
3731          * rename() is an acceptable substitute except for the truly paranoid.
3732          */
3733 #if HAVE_WORKING_LINK
3734         if (link(tmppath, path) < 0)
3735                 ereport(ERROR,
3736                                 (errcode_for_file_access(),
3737                                  errmsg("could not link file \"%s\" to \"%s\": %m",
3738                                                 tmppath, path)));
3739         unlink(tmppath);
3740 #else
3741         if (rename(tmppath, path) < 0)
3742                 ereport(ERROR,
3743                                 (errcode_for_file_access(),
3744                                  errmsg("could not rename file \"%s\" to \"%s\": %m",
3745                                                 tmppath, path)));
3746 #endif
3747
3748         /* The history file can be archived immediately. */
3749         TLHistoryFileName(histfname, newTLI);
3750         XLogArchiveNotify(histfname);
3751 }
3752
3753 /*
3754  * I/O routines for pg_control
3755  *
3756  * *ControlFile is a buffer in shared memory that holds an image of the
3757  * contents of pg_control.      WriteControlFile() initializes pg_control
3758  * given a preloaded buffer, ReadControlFile() loads the buffer from
3759  * the pg_control file (during postmaster or standalone-backend startup),
3760  * and UpdateControlFile() rewrites pg_control after we modify xlog state.
3761  *
3762  * For simplicity, WriteControlFile() initializes the fields of pg_control
3763  * that are related to checking backend/database compatibility, and
3764  * ReadControlFile() verifies they are correct.  We could split out the
3765  * I/O and compatibility-check functions, but there seems no need currently.
3766  */
3767 static void
3768 WriteControlFile(void)
3769 {
3770         int                     fd;
3771         char            buffer[PG_CONTROL_SIZE];                /* need not be aligned */
3772         char       *localeptr;
3773
3774         /*
3775          * Initialize version and compatibility-check fields
3776          */
3777         ControlFile->pg_control_version = PG_CONTROL_VERSION;
3778         ControlFile->catalog_version_no = CATALOG_VERSION_NO;
3779
3780         ControlFile->maxAlign = MAXIMUM_ALIGNOF;
3781         ControlFile->floatFormat = FLOATFORMAT_VALUE;
3782
3783         ControlFile->blcksz = BLCKSZ;
3784         ControlFile->relseg_size = RELSEG_SIZE;
3785         ControlFile->xlog_blcksz = XLOG_BLCKSZ;
3786         ControlFile->xlog_seg_size = XLOG_SEG_SIZE;
3787
3788         ControlFile->nameDataLen = NAMEDATALEN;
3789         ControlFile->indexMaxKeys = INDEX_MAX_KEYS;
3790
3791         ControlFile->toast_max_chunk_size = TOAST_MAX_CHUNK_SIZE;
3792
3793 #ifdef HAVE_INT64_TIMESTAMP
3794         ControlFile->enableIntTimes = true;
3795 #else
3796         ControlFile->enableIntTimes = false;
3797 #endif
3798         ControlFile->float4ByVal = FLOAT4PASSBYVAL;
3799         ControlFile->float8ByVal = FLOAT8PASSBYVAL;
3800
3801         ControlFile->localeBuflen = LOCALE_NAME_BUFLEN;
3802         localeptr = setlocale(LC_COLLATE, NULL);
3803         if (!localeptr)
3804                 ereport(PANIC,
3805                                 (errmsg("invalid LC_COLLATE setting")));
3806         StrNCpy(ControlFile->lc_collate, localeptr, LOCALE_NAME_BUFLEN);
3807         localeptr = setlocale(LC_CTYPE, NULL);
3808         if (!localeptr)
3809                 ereport(PANIC,
3810                                 (errmsg("invalid LC_CTYPE setting")));
3811         StrNCpy(ControlFile->lc_ctype, localeptr, LOCALE_NAME_BUFLEN);
3812
3813         /* Contents are protected with a CRC */
3814         INIT_CRC32(ControlFile->crc);
3815         COMP_CRC32(ControlFile->crc,
3816                            (char *) ControlFile,
3817                            offsetof(ControlFileData, crc));
3818         FIN_CRC32(ControlFile->crc);
3819
3820         /*
3821          * We write out PG_CONTROL_SIZE bytes into pg_control, zero-padding the
3822          * excess over sizeof(ControlFileData).  This reduces the odds of
3823          * premature-EOF errors when reading pg_control.  We'll still fail when we
3824          * check the contents of the file, but hopefully with a more specific
3825          * error than "couldn't read pg_control".
3826          */
3827         if (sizeof(ControlFileData) > PG_CONTROL_SIZE)
3828                 elog(PANIC, "sizeof(ControlFileData) is larger than PG_CONTROL_SIZE; fix either one");
3829
3830         memset(buffer, 0, PG_CONTROL_SIZE);
3831         memcpy(buffer, ControlFile, sizeof(ControlFileData));
3832
3833         fd = BasicOpenFile(XLOG_CONTROL_FILE,
3834                                            O_RDWR | O_CREAT | O_EXCL | PG_BINARY,
3835                                            S_IRUSR | S_IWUSR);
3836         if (fd < 0)
3837                 ereport(PANIC,
3838                                 (errcode_for_file_access(),
3839                                  errmsg("could not create control file \"%s\": %m",
3840                                                 XLOG_CONTROL_FILE)));
3841
3842         errno = 0;
3843         if (write(fd, buffer, PG_CONTROL_SIZE) != PG_CONTROL_SIZE)
3844         {
3845                 /* if write didn't set errno, assume problem is no disk space */
3846                 if (errno == 0)
3847                         errno = ENOSPC;
3848                 ereport(PANIC,
3849                                 (errcode_for_file_access(),
3850                                  errmsg("could not write to control file: %m")));
3851         }
3852
3853         if (pg_fsync(fd) != 0)
3854                 ereport(PANIC,
3855                                 (errcode_for_file_access(),
3856                                  errmsg("could not fsync control file: %m")));
3857
3858         if (close(fd))
3859                 ereport(PANIC,
3860                                 (errcode_for_file_access(),
3861                                  errmsg("could not close control file: %m")));
3862 }
3863
3864 static void
3865 ReadControlFile(void)
3866 {
3867         pg_crc32        crc;
3868         int                     fd;
3869
3870         /*
3871          * Read data...
3872          */
3873         fd = BasicOpenFile(XLOG_CONTROL_FILE,
3874                                            O_RDWR | PG_BINARY,
3875                                            S_IRUSR | S_IWUSR);
3876         if (fd < 0)
3877                 ereport(PANIC,
3878                                 (errcode_for_file_access(),
3879                                  errmsg("could not open control file \"%s\": %m",
3880                                                 XLOG_CONTROL_FILE)));
3881
3882         if (read(fd, ControlFile, sizeof(ControlFileData)) != sizeof(ControlFileData))
3883                 ereport(PANIC,
3884                                 (errcode_for_file_access(),
3885                                  errmsg("could not read from control file: %m")));
3886
3887         close(fd);
3888
3889         /*
3890          * Check for expected pg_control format version.  If this is wrong, the
3891          * CRC check will likely fail because we'll be checking the wrong number
3892          * of bytes.  Complaining about wrong version will probably be more
3893          * enlightening than complaining about wrong CRC.
3894          */
3895
3896         if (ControlFile->pg_control_version != PG_CONTROL_VERSION && ControlFile->pg_control_version % 65536 == 0 && ControlFile->pg_control_version / 65536 != 0)
3897                 ereport(FATAL,
3898                                 (errmsg("database files are incompatible with server"),
3899                                  errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d (0x%08x),"
3900                                                    " but the server was compiled with PG_CONTROL_VERSION %d (0x%08x).",
3901                                                    ControlFile->pg_control_version, ControlFile->pg_control_version,
3902                                                    PG_CONTROL_VERSION, PG_CONTROL_VERSION),
3903                                  errhint("This could be a problem of mismatched byte ordering.  It looks like you need to initdb.")));
3904
3905         if (ControlFile->pg_control_version != PG_CONTROL_VERSION)
3906                 ereport(FATAL,
3907                                 (errmsg("database files are incompatible with server"),
3908                                  errdetail("The database cluster was initialized with PG_CONTROL_VERSION %d,"
3909                                   " but the server was compiled with PG_CONTROL_VERSION %d.",
3910                                                 ControlFile->pg_control_version, PG_CONTROL_VERSION),
3911                                  errhint("It looks like you need to initdb.")));
3912
3913         /* Now check the CRC. */
3914         INIT_CRC32(crc);
3915         COMP_CRC32(crc,
3916                            (char *) ControlFile,
3917                            offsetof(ControlFileData, crc));
3918         FIN_CRC32(crc);
3919
3920         if (!EQ_CRC32(crc, ControlFile->crc))
3921                 ereport(FATAL,
3922                                 (errmsg("incorrect checksum in control file")));
3923
3924         /*
3925          * Do compatibility checking immediately.  We do this here for 2 reasons:
3926          *
3927          * (1) if the database isn't compatible with the backend executable, we
3928          * want to abort before we can possibly do any damage;
3929          *
3930          * (2) this code is executed in the postmaster, so the setlocale() will
3931          * propagate to forked backends, which aren't going to read this file for
3932          * themselves.  (These locale settings are considered critical
3933          * compatibility items because they can affect sort order of indexes.)
3934          */
3935         if (ControlFile->catalog_version_no != CATALOG_VERSION_NO)
3936                 ereport(FATAL,
3937                                 (errmsg("database files are incompatible with server"),
3938                                  errdetail("The database cluster was initialized with CATALOG_VERSION_NO %d,"
3939                                   " but the server was compiled with CATALOG_VERSION_NO %d.",
3940                                                 ControlFile->catalog_version_no, CATALOG_VERSION_NO),
3941                                  errhint("It looks like you need to initdb.")));
3942         if (ControlFile->maxAlign != MAXIMUM_ALIGNOF)
3943                 ereport(FATAL,
3944                                 (errmsg("database files are incompatible with server"),
3945                    errdetail("The database cluster was initialized with MAXALIGN %d,"
3946                                          " but the server was compiled with MAXALIGN %d.",
3947                                          ControlFile->maxAlign, MAXIMUM_ALIGNOF),
3948                                  errhint("It looks like you need to initdb.")));
3949         if (ControlFile->floatFormat != FLOATFORMAT_VALUE)
3950                 ereport(FATAL,
3951                                 (errmsg("database files are incompatible with server"),
3952                                  errdetail("The database cluster appears to use a different floating-point number format than the server executable."),
3953                                  errhint("It looks like you need to initdb.")));
3954         if (ControlFile->blcksz != BLCKSZ)
3955                 ereport(FATAL,
3956                                 (errmsg("database files are incompatible with server"),
3957                          errdetail("The database cluster was initialized with BLCKSZ %d,"
3958                                            " but the server was compiled with BLCKSZ %d.",
3959                                            ControlFile->blcksz, BLCKSZ),
3960                                  errhint("It looks like you need to recompile or initdb.")));
3961         if (ControlFile->relseg_size != RELSEG_SIZE)
3962                 ereport(FATAL,
3963                                 (errmsg("database files are incompatible with server"),
3964                 errdetail("The database cluster was initialized with RELSEG_SIZE %d,"
3965                                   " but the server was compiled with RELSEG_SIZE %d.",
3966                                   ControlFile->relseg_size, RELSEG_SIZE),
3967                                  errhint("It looks like you need to recompile or initdb.")));
3968         if (ControlFile->xlog_blcksz != XLOG_BLCKSZ)
3969                 ereport(FATAL,
3970                                 (errmsg("database files are incompatible with server"),
3971                 errdetail("The database cluster was initialized with XLOG_BLCKSZ %d,"
3972                                   " but the server was compiled with XLOG_BLCKSZ %d.",
3973                                   ControlFile->xlog_blcksz, XLOG_BLCKSZ),
3974                                  errhint("It looks like you need to recompile or initdb.")));
3975         if (ControlFile->xlog_seg_size != XLOG_SEG_SIZE)
3976                 ereport(FATAL,
3977                                 (errmsg("database files are incompatible with server"),
3978                                  errdetail("The database cluster was initialized with XLOG_SEG_SIZE %d,"
3979                                            " but the server was compiled with XLOG_SEG_SIZE %d.",
3980                                                    ControlFile->xlog_seg_size, XLOG_SEG_SIZE),
3981                                  errhint("It looks like you need to recompile or initdb.")));
3982         if (ControlFile->nameDataLen != NAMEDATALEN)
3983                 ereport(FATAL,
3984                                 (errmsg("database files are incompatible with server"),
3985                 errdetail("The database cluster was initialized with NAMEDATALEN %d,"
3986                                   " but the server was compiled with NAMEDATALEN %d.",
3987                                   ControlFile->nameDataLen, NAMEDATALEN),
3988                                  errhint("It looks like you need to recompile or initdb.")));
3989         if (ControlFile->indexMaxKeys != INDEX_MAX_KEYS)
3990                 ereport(FATAL,
3991                                 (errmsg("database files are incompatible with server"),
3992                                  errdetail("The database cluster was initialized with INDEX_MAX_KEYS %d,"
3993                                           " but the server was compiled with INDEX_MAX_KEYS %d.",
3994                                                    ControlFile->indexMaxKeys, INDEX_MAX_KEYS),
3995                                  errhint("It looks like you need to recompile or initdb.")));
3996         if (ControlFile->toast_max_chunk_size != TOAST_MAX_CHUNK_SIZE)
3997                 ereport(FATAL,
3998                                 (errmsg("database files are incompatible with server"),
3999                                  errdetail("The database cluster was initialized with TOAST_MAX_CHUNK_SIZE %d,"
4000                                 " but the server was compiled with TOAST_MAX_CHUNK_SIZE %d.",
4001                           ControlFile->toast_max_chunk_size, (int) TOAST_MAX_CHUNK_SIZE),
4002                                  errhint("It looks like you need to recompile or initdb.")));
4003
4004 #ifdef HAVE_INT64_TIMESTAMP
4005         if (ControlFile->enableIntTimes != true)
4006                 ereport(FATAL,
4007                                 (errmsg("database files are incompatible with server"),
4008                                  errdetail("The database cluster was initialized without HAVE_INT64_TIMESTAMP"
4009                                   " but the server was compiled with HAVE_INT64_TIMESTAMP."),
4010                                  errhint("It looks like you need to recompile or initdb.")));
4011 #else
4012         if (ControlFile->enableIntTimes != false)
4013                 ereport(FATAL,
4014                                 (errmsg("database files are incompatible with server"),
4015                                  errdetail("The database cluster was initialized with HAVE_INT64_TIMESTAMP"
4016                            " but the server was compiled without HAVE_INT64_TIMESTAMP."),
4017                                  errhint("It looks like you need to recompile or initdb.")));
4018 #endif
4019
4020 #ifdef USE_FLOAT4_BYVAL
4021         if (ControlFile->float4ByVal != true)
4022                 ereport(FATAL,
4023                                 (errmsg("database files are incompatible with server"),
4024                                  errdetail("The database cluster was initialized without USE_FLOAT4_BYVAL"
4025                                                    " but the server was compiled with USE_FLOAT4_BYVAL."),
4026                                  errhint("It looks like you need to recompile or initdb.")));
4027 #else
4028         if (ControlFile->float4ByVal != false)
4029                 ereport(FATAL,
4030                                 (errmsg("database files are incompatible with server"),
4031                                  errdetail("The database cluster was initialized with USE_FLOAT4_BYVAL"
4032                                                    " but the server was compiled without USE_FLOAT4_BYVAL."),
4033                                  errhint("It looks like you need to recompile or initdb.")));
4034 #endif
4035
4036 #ifdef USE_FLOAT8_BYVAL
4037         if (ControlFile->float8ByVal != true)
4038                 ereport(FATAL,
4039                                 (errmsg("database files are incompatible with server"),
4040                                  errdetail("The database cluster was initialized without USE_FLOAT8_BYVAL"
4041                                                    " but the server was compiled with USE_FLOAT8_BYVAL."),
4042                                  errhint("It looks like you need to recompile or initdb.")));
4043 #else
4044         if (ControlFile->float8ByVal != false)
4045                 ereport(FATAL,
4046                                 (errmsg("database files are incompatible with server"),
4047                                  errdetail("The database cluster was initialized with USE_FLOAT8_BYVAL"
4048                                                    " but the server was compiled without USE_FLOAT8_BYVAL."),
4049                                  errhint("It looks like you need to recompile or initdb.")));
4050 #endif
4051
4052         if (ControlFile->localeBuflen != LOCALE_NAME_BUFLEN)
4053                 ereport(FATAL,
4054                                 (errmsg("database files are incompatible with server"),
4055                                  errdetail("The database cluster was initialized with LOCALE_NAME_BUFLEN %d,"
4056                                   " but the server was compiled with LOCALE_NAME_BUFLEN %d.",
4057                                                    ControlFile->localeBuflen, LOCALE_NAME_BUFLEN),
4058                                  errhint("It looks like you need to recompile or initdb.")));
4059         if (pg_perm_setlocale(LC_COLLATE, ControlFile->lc_collate) == NULL)
4060                 ereport(FATAL,
4061                         (errmsg("database files are incompatible with operating system"),
4062                          errdetail("The database cluster was initialized with LC_COLLATE \"%s\","
4063                                            " which is not recognized by setlocale().",
4064                                            ControlFile->lc_collate),
4065                          errhint("It looks like you need to initdb or install locale support.")));
4066         if (pg_perm_setlocale(LC_CTYPE, ControlFile->lc_ctype) == NULL)
4067                 ereport(FATAL,
4068                         (errmsg("database files are incompatible with operating system"),
4069                 errdetail("The database cluster was initialized with LC_CTYPE \"%s\","
4070                                   " which is not recognized by setlocale().",
4071                                   ControlFile->lc_ctype),
4072                          errhint("It looks like you need to initdb or install locale support.")));
4073
4074         /* Make the fixed locale settings visible as GUC variables, too */
4075         SetConfigOption("lc_collate", ControlFile->lc_collate,
4076                                         PGC_INTERNAL, PGC_S_OVERRIDE);
4077         SetConfigOption("lc_ctype", ControlFile->lc_ctype,
4078                                         PGC_INTERNAL, PGC_S_OVERRIDE);
4079 }
4080
4081 void
4082 UpdateControlFile(void)
4083 {
4084         int                     fd;
4085
4086         INIT_CRC32(ControlFile->crc);
4087         COMP_CRC32(ControlFile->crc,
4088                            (char *) ControlFile,
4089                            offsetof(ControlFileData, crc));
4090         FIN_CRC32(ControlFile->crc);
4091
4092         fd = BasicOpenFile(XLOG_CONTROL_FILE,
4093                                            O_RDWR | PG_BINARY,
4094                                            S_IRUSR | S_IWUSR);
4095         if (fd < 0)
4096                 ereport(PANIC,
4097                                 (errcode_for_file_access(),
4098                                  errmsg("could not open control file \"%s\": %m",
4099                                                 XLOG_CONTROL_FILE)));
4100
4101         errno = 0;
4102         if (write(fd, ControlFile, sizeof(ControlFileData)) != sizeof(ControlFileData))
4103         {
4104                 /* if write didn't set errno, assume problem is no disk space */
4105                 if (errno == 0)
4106                         errno = ENOSPC;
4107                 ereport(PANIC,
4108                                 (errcode_for_file_access(),
4109                                  errmsg("could not write to control file: %m")));
4110         }
4111
4112         if (pg_fsync(fd) != 0)
4113                 ereport(PANIC,
4114                                 (errcode_for_file_access(),
4115                                  errmsg("could not fsync control file: %m")));
4116
4117         if (close(fd))
4118                 ereport(PANIC,
4119                                 (errcode_for_file_access(),
4120                                  errmsg("could not close control file: %m")));
4121 }
4122
4123 /*
4124  * Initialization of shared memory for XLOG
4125  */
4126 Size
4127 XLOGShmemSize(void)
4128 {
4129         Size            size;
4130
4131         /* XLogCtl */
4132         size = sizeof(XLogCtlData);
4133         /* xlblocks array */
4134         size = add_size(size, mul_size(sizeof(XLogRecPtr), XLOGbuffers));
4135         /* extra alignment padding for XLOG I/O buffers */
4136         size = add_size(size, ALIGNOF_XLOG_BUFFER);
4137         /* and the buffers themselves */
4138         size = add_size(size, mul_size(XLOG_BLCKSZ, XLOGbuffers));
4139
4140         /*
4141          * Note: we don't count ControlFileData, it comes out of the "slop factor"
4142          * added by CreateSharedMemoryAndSemaphores.  This lets us use this
4143          * routine again below to compute the actual allocation size.
4144          */
4145
4146         return size;
4147 }
4148
4149 void
4150 XLOGShmemInit(void)
4151 {
4152         bool            foundCFile,
4153                                 foundXLog;
4154         char       *allocptr;
4155
4156         ControlFile = (ControlFileData *)
4157                 ShmemInitStruct("Control File", sizeof(ControlFileData), &foundCFile);
4158         XLogCtl = (XLogCtlData *)
4159                 ShmemInitStruct("XLOG Ctl", XLOGShmemSize(), &foundXLog);
4160
4161         if (foundCFile || foundXLog)
4162         {
4163                 /* both should be present or neither */
4164                 Assert(foundCFile && foundXLog);
4165                 return;
4166         }
4167
4168         memset(XLogCtl, 0, sizeof(XLogCtlData));
4169
4170         /*
4171          * Since XLogCtlData contains XLogRecPtr fields, its sizeof should be a
4172          * multiple of the alignment for same, so no extra alignment padding is
4173          * needed here.
4174          */
4175         allocptr = ((char *) XLogCtl) + sizeof(XLogCtlData);
4176         XLogCtl->xlblocks = (XLogRecPtr *) allocptr;
4177         memset(XLogCtl->xlblocks, 0, sizeof(XLogRecPtr) * XLOGbuffers);
4178         allocptr += sizeof(XLogRecPtr) * XLOGbuffers;
4179
4180         /*
4181          * Align the start of the page buffers to an ALIGNOF_XLOG_BUFFER boundary.
4182          */
4183         allocptr = (char *) TYPEALIGN(ALIGNOF_XLOG_BUFFER, allocptr);
4184         XLogCtl->pages = allocptr;
4185         memset(XLogCtl->pages, 0, (Size) XLOG_BLCKSZ * XLOGbuffers);
4186
4187         /*
4188          * Do basic initialization of XLogCtl shared data. (StartupXLOG will fill
4189          * in additional info.)
4190          */
4191         XLogCtl->XLogCacheBlck = XLOGbuffers - 1;
4192         XLogCtl->Insert.currpage = (XLogPageHeader) (XLogCtl->pages);
4193         SpinLockInit(&XLogCtl->info_lck);
4194
4195         /*
4196          * If we are not in bootstrap mode, pg_control should already exist. Read
4197          * and validate it immediately (see comments in ReadControlFile() for the
4198          * reasons why).
4199          */
4200         if (!IsBootstrapProcessingMode())
4201                 ReadControlFile();
4202 }
4203
4204 /*
4205  * This func must be called ONCE on system install.  It creates pg_control
4206  * and the initial XLOG segment.
4207  */
4208 void
4209 BootStrapXLOG(void)
4210 {
4211         CheckPoint      checkPoint;
4212         char       *buffer;
4213         XLogPageHeader page;
4214         XLogLongPageHeader longpage;
4215         XLogRecord *record;
4216         bool            use_existent;
4217         uint64          sysidentifier;
4218         struct timeval tv;
4219         pg_crc32        crc;
4220
4221         /*
4222          * Select a hopefully-unique system identifier code for this installation.
4223          * We use the result of gettimeofday(), including the fractional seconds
4224          * field, as being about as unique as we can easily get.  (Think not to
4225          * use random(), since it hasn't been seeded and there's no portable way
4226          * to seed it other than the system clock value...)  The upper half of the
4227          * uint64 value is just the tv_sec part, while the lower half is the XOR
4228          * of tv_sec and tv_usec.  This is to ensure that we don't lose uniqueness
4229          * unnecessarily if "uint64" is really only 32 bits wide.  A person
4230          * knowing this encoding can determine the initialization time of the
4231          * installation, which could perhaps be useful sometimes.
4232          */
4233         gettimeofday(&tv, NULL);
4234         sysidentifier = ((uint64) tv.tv_sec) << 32;
4235         sysidentifier |= (uint32) (tv.tv_sec | tv.tv_usec);
4236
4237         /* First timeline ID is always 1 */
4238         ThisTimeLineID = 1;
4239
4240         /* page buffer must be aligned suitably for O_DIRECT */
4241         buffer = (char *) palloc(XLOG_BLCKSZ + ALIGNOF_XLOG_BUFFER);
4242         page = (XLogPageHeader) TYPEALIGN(ALIGNOF_XLOG_BUFFER, buffer);
4243         memset(page, 0, XLOG_BLCKSZ);
4244
4245         /* Set up information for the initial checkpoint record */
4246         checkPoint.redo.xlogid = 0;
4247         checkPoint.redo.xrecoff = SizeOfXLogLongPHD;
4248         checkPoint.ThisTimeLineID = ThisTimeLineID;
4249         checkPoint.nextXidEpoch = 0;
4250         checkPoint.nextXid = FirstNormalTransactionId;
4251         checkPoint.nextOid = FirstBootstrapObjectId;
4252         checkPoint.nextMulti = FirstMultiXactId;
4253         checkPoint.nextMultiOffset = 0;
4254         checkPoint.time = (pg_time_t) time(NULL);
4255
4256         ShmemVariableCache->nextXid = checkPoint.nextXid;
4257         ShmemVariableCache->nextOid = checkPoint.nextOid;
4258         ShmemVariableCache->oidCount = 0;
4259         MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
4260
4261         /* Set up the XLOG page header */
4262         page->xlp_magic = XLOG_PAGE_MAGIC;
4263         page->xlp_info = XLP_LONG_HEADER;
4264         page->xlp_tli = ThisTimeLineID;
4265         page->xlp_pageaddr.xlogid = 0;
4266         page->xlp_pageaddr.xrecoff = 0;
4267         longpage = (XLogLongPageHeader) page;
4268         longpage->xlp_sysid = sysidentifier;
4269         longpage->xlp_seg_size = XLogSegSize;
4270         longpage->xlp_xlog_blcksz = XLOG_BLCKSZ;
4271
4272         /* Insert the initial checkpoint record */
4273         record = (XLogRecord *) ((char *) page + SizeOfXLogLongPHD);
4274         record->xl_prev.xlogid = 0;
4275         record->xl_prev.xrecoff = 0;
4276         record->xl_xid = InvalidTransactionId;
4277         record->xl_tot_len = SizeOfXLogRecord + sizeof(checkPoint);
4278         record->xl_len = sizeof(checkPoint);
4279         record->xl_info = XLOG_CHECKPOINT_SHUTDOWN;
4280         record->xl_rmid = RM_XLOG_ID;
4281         memcpy(XLogRecGetData(record), &checkPoint, sizeof(checkPoint));
4282
4283         INIT_CRC32(crc);
4284         COMP_CRC32(crc, &checkPoint, sizeof(checkPoint));
4285         COMP_CRC32(crc, (char *) record + sizeof(pg_crc32),
4286                            SizeOfXLogRecord - sizeof(pg_crc32));
4287         FIN_CRC32(crc);
4288         record->xl_crc = crc;
4289
4290         /* Create first XLOG segment file */
4291         use_existent = false;
4292         openLogFile = XLogFileInit(0, 0, &use_existent, false);
4293
4294         /* Write the first page with the initial record */
4295         errno = 0;
4296         if (write(openLogFile, page, XLOG_BLCKSZ) != XLOG_BLCKSZ)
4297         {
4298                 /* if write didn't set errno, assume problem is no disk space */
4299                 if (errno == 0)
4300                         errno = ENOSPC;
4301                 ereport(PANIC,
4302                                 (errcode_for_file_access(),
4303                           errmsg("could not write bootstrap transaction log file: %m")));
4304         }
4305
4306         if (pg_fsync(openLogFile) != 0)
4307                 ereport(PANIC,
4308                                 (errcode_for_file_access(),
4309                           errmsg("could not fsync bootstrap transaction log file: %m")));
4310
4311         if (close(openLogFile))
4312                 ereport(PANIC,
4313                                 (errcode_for_file_access(),
4314                           errmsg("could not close bootstrap transaction log file: %m")));
4315
4316         openLogFile = -1;
4317
4318         /* Now create pg_control */
4319
4320         memset(ControlFile, 0, sizeof(ControlFileData));
4321         /* Initialize pg_control status fields */
4322         ControlFile->system_identifier = sysidentifier;
4323         ControlFile->state = DB_SHUTDOWNED;
4324         ControlFile->time = checkPoint.time;
4325         ControlFile->checkPoint = checkPoint.redo;
4326         ControlFile->checkPointCopy = checkPoint;
4327         /* some additional ControlFile fields are set in WriteControlFile() */
4328
4329         WriteControlFile();
4330
4331         /* Bootstrap the commit log, too */
4332         BootStrapCLOG();
4333         BootStrapSUBTRANS();
4334         BootStrapMultiXact();
4335
4336         pfree(buffer);
4337 }
4338
4339 static char *
4340 str_time(pg_time_t tnow)
4341 {
4342         static char buf[128];
4343
4344         pg_strftime(buf, sizeof(buf),
4345                                 "%Y-%m-%d %H:%M:%S %Z",
4346                                 pg_localtime(&tnow, log_timezone));
4347
4348         return buf;
4349 }
4350
4351 /*
4352  * See if there is a recovery command file (recovery.conf), and if so
4353  * read in parameters for archive recovery.
4354  *
4355  * XXX longer term intention is to expand this to
4356  * cater for additional parameters and controls
4357  * possibly use a flex lexer similar to the GUC one
4358  */
4359 static void
4360 readRecoveryCommandFile(void)
4361 {
4362         FILE       *fd;
4363         char            cmdline[MAXPGPATH];
4364         TimeLineID      rtli = 0;
4365         bool            rtliGiven = false;
4366         bool            syntaxError = false;
4367
4368         fd = AllocateFile(RECOVERY_COMMAND_FILE, "r");
4369         if (fd == NULL)
4370         {
4371                 if (errno == ENOENT)
4372                         return;                         /* not there, so no archive recovery */
4373                 ereport(FATAL,
4374                                 (errcode_for_file_access(),
4375                                  errmsg("could not open recovery command file \"%s\": %m",
4376                                                 RECOVERY_COMMAND_FILE)));
4377         }
4378
4379         ereport(LOG,
4380                         (errmsg("starting archive recovery")));
4381
4382         /*
4383          * Parse the file...
4384          */
4385         while (fgets(cmdline, sizeof(cmdline), fd) != NULL)
4386         {
4387                 /* skip leading whitespace and check for # comment */
4388                 char       *ptr;
4389                 char       *tok1;
4390                 char       *tok2;
4391
4392                 for (ptr = cmdline; *ptr; ptr++)
4393                 {
4394                         if (!isspace((unsigned char) *ptr))
4395                                 break;
4396                 }
4397                 if (*ptr == '\0' || *ptr == '#')
4398                         continue;
4399
4400                 /* identify the quoted parameter value */
4401                 tok1 = strtok(ptr, "'");
4402                 if (!tok1)
4403                 {
4404                         syntaxError = true;
4405                         break;
4406                 }
4407                 tok2 = strtok(NULL, "'");
4408                 if (!tok2)
4409                 {
4410                         syntaxError = true;
4411                         break;
4412                 }
4413                 /* reparse to get just the parameter name */
4414                 tok1 = strtok(ptr, " \t=");
4415                 if (!tok1)
4416                 {
4417                         syntaxError = true;
4418                         break;
4419                 }
4420
4421                 if (strcmp(tok1, "restore_command") == 0)
4422                 {
4423                         recoveryRestoreCommand = pstrdup(tok2);
4424                         ereport(LOG,
4425                                         (errmsg("restore_command = '%s'",
4426                                                         recoveryRestoreCommand)));
4427                 }
4428                 else if (strcmp(tok1, "recovery_target_timeline") == 0)
4429                 {
4430                         rtliGiven = true;
4431                         if (strcmp(tok2, "latest") == 0)
4432                                 rtli = 0;
4433                         else
4434                         {
4435                                 errno = 0;
4436                                 rtli = (TimeLineID) strtoul(tok2, NULL, 0);
4437                                 if (errno == EINVAL || errno == ERANGE)
4438                                         ereport(FATAL,
4439                                                         (errmsg("recovery_target_timeline is not a valid number: \"%s\"",
4440                                                                         tok2)));
4441                         }
4442                         if (rtli)
4443                                 ereport(LOG,
4444                                                 (errmsg("recovery_target_timeline = %u", rtli)));
4445                         else
4446                                 ereport(LOG,
4447                                                 (errmsg("recovery_target_timeline = latest")));
4448                 }
4449                 else if (strcmp(tok1, "recovery_target_xid") == 0)
4450                 {
4451                         errno = 0;
4452                         recoveryTargetXid = (TransactionId) strtoul(tok2, NULL, 0);
4453                         if (errno == EINVAL || errno == ERANGE)
4454                                 ereport(FATAL,
4455                                  (errmsg("recovery_target_xid is not a valid number: \"%s\"",
4456                                                  tok2)));
4457                         ereport(LOG,
4458                                         (errmsg("recovery_target_xid = %u",
4459                                                         recoveryTargetXid)));
4460                         recoveryTarget = true;
4461                         recoveryTargetExact = true;
4462                 }
4463                 else if (strcmp(tok1, "recovery_target_time") == 0)
4464                 {
4465                         /*
4466                          * if recovery_target_xid specified, then this overrides
4467                          * recovery_target_time
4468                          */
4469                         if (recoveryTargetExact)
4470                                 continue;
4471                         recoveryTarget = true;
4472                         recoveryTargetExact = false;
4473
4474                         /*
4475                          * Convert the time string given by the user to TimestampTz form.
4476                          */
4477                         recoveryTargetTime =
4478                                 DatumGetTimestampTz(DirectFunctionCall3(timestamptz_in,
4479                                                                                                                 CStringGetDatum(tok2),
4480                                                                                                 ObjectIdGetDatum(InvalidOid),
4481                                                                                                                 Int32GetDatum(-1)));
4482                         ereport(LOG,
4483                                         (errmsg("recovery_target_time = '%s'",
4484                                                         timestamptz_to_str(recoveryTargetTime))));
4485                 }
4486                 else if (strcmp(tok1, "recovery_target_inclusive") == 0)
4487                 {
4488                         /*
4489                          * does nothing if a recovery_target is not also set
4490                          */
4491                         if (strcmp(tok2, "true") == 0)
4492                                 recoveryTargetInclusive = true;
4493                         else
4494                         {
4495                                 recoveryTargetInclusive = false;
4496                                 tok2 = "false";
4497                         }
4498                         ereport(LOG,
4499                                         (errmsg("recovery_target_inclusive = %s", tok2)));
4500                 }
4501                 else if (strcmp(tok1, "log_restartpoints") == 0)
4502                 {
4503                         /*
4504                          * does nothing if a recovery_target is not also set
4505                          */
4506                         if (strcmp(tok2, "true") == 0)
4507                                 recoveryLogRestartpoints = true;
4508                         else
4509                         {
4510                                 recoveryLogRestartpoints = false;
4511                                 tok2 = "false";
4512                         }
4513                         ereport(LOG,
4514                                         (errmsg("log_restartpoints = %s", tok2)));
4515                 }
4516                 else
4517                         ereport(FATAL,
4518                                         (errmsg("unrecognized recovery parameter \"%s\"",
4519                                                         tok1)));
4520         }
4521
4522         FreeFile(fd);
4523
4524         if (syntaxError)
4525                 ereport(FATAL,
4526                                 (errmsg("syntax error in recovery command file: %s",
4527                                                 cmdline),
4528                           errhint("Lines should have the format parameter = 'value'.")));
4529
4530         /* Check that required parameters were supplied */
4531         if (recoveryRestoreCommand == NULL)
4532                 ereport(FATAL,
4533                                 (errmsg("recovery command file \"%s\" did not specify restore_command",
4534                                                 RECOVERY_COMMAND_FILE)));
4535
4536         /* Enable fetching from archive recovery area */
4537         InArchiveRecovery = true;
4538
4539         /*
4540          * If user specified recovery_target_timeline, validate it or compute the
4541          * "latest" value.      We can't do this until after we've gotten the restore
4542          * command and set InArchiveRecovery, because we need to fetch timeline
4543          * history files from the archive.
4544          */
4545         if (rtliGiven)
4546         {
4547                 if (rtli)
4548                 {
4549                         /* Timeline 1 does not have a history file, all else should */
4550                         if (rtli != 1 && !existsTimeLineHistory(rtli))
4551                                 ereport(FATAL,
4552                                                 (errmsg("recovery target timeline %u does not exist",
4553                                                                 rtli)));
4554                         recoveryTargetTLI = rtli;
4555                 }
4556                 else
4557                 {
4558                         /* We start the "latest" search from pg_control's timeline */
4559                         recoveryTargetTLI = findNewestTimeLine(recoveryTargetTLI);
4560                 }
4561         }
4562 }
4563
4564 /*
4565  * Exit archive-recovery state
4566  */
4567 static void
4568 exitArchiveRecovery(TimeLineID endTLI, uint32 endLogId, uint32 endLogSeg)
4569 {
4570         char            recoveryPath[MAXPGPATH];
4571         char            xlogpath[MAXPGPATH];
4572
4573         /*
4574          * We are no longer in archive recovery state.
4575          */
4576         InArchiveRecovery = false;
4577
4578         /*
4579          * We should have the ending log segment currently open.  Verify, and then
4580          * close it (to avoid problems on Windows with trying to rename or delete
4581          * an open file).
4582          */
4583         Assert(readFile >= 0);
4584         Assert(readId == endLogId);
4585         Assert(readSeg == endLogSeg);
4586
4587         close(readFile);
4588         readFile = -1;
4589
4590         /*
4591          * If the segment was fetched from archival storage, we want to replace
4592          * the existing xlog segment (if any) with the archival version.  This is
4593          * because whatever is in XLOGDIR is very possibly older than what we have
4594          * from the archives, since it could have come from restoring a PGDATA
4595          * backup.      In any case, the archival version certainly is more
4596          * descriptive of what our current database state is, because that is what
4597          * we replayed from.
4598          *
4599          * Note that if we are establishing a new timeline, ThisTimeLineID is
4600          * already set to the new value, and so we will create a new file instead
4601          * of overwriting any existing file.  (This is, in fact, always the case
4602          * at present.)
4603          */
4604         snprintf(recoveryPath, MAXPGPATH, XLOGDIR "/RECOVERYXLOG");
4605         XLogFilePath(xlogpath, ThisTimeLineID, endLogId, endLogSeg);
4606
4607         if (restoredFromArchive)
4608         {
4609                 ereport(DEBUG3,
4610                                 (errmsg_internal("moving last restored xlog to \"%s\"",
4611                                                                  xlogpath)));
4612                 unlink(xlogpath);               /* might or might not exist */
4613                 if (rename(recoveryPath, xlogpath) != 0)
4614                         ereport(FATAL,
4615                                         (errcode_for_file_access(),
4616                                          errmsg("could not rename file \"%s\" to \"%s\": %m",
4617                                                         recoveryPath, xlogpath)));
4618                 /* XXX might we need to fix permissions on the file? */
4619         }
4620         else
4621         {
4622                 /*
4623                  * If the latest segment is not archival, but there's still a
4624                  * RECOVERYXLOG laying about, get rid of it.
4625                  */
4626                 unlink(recoveryPath);   /* ignore any error */
4627
4628                 /*
4629                  * If we are establishing a new timeline, we have to copy data from
4630                  * the last WAL segment of the old timeline to create a starting WAL
4631                  * segment for the new timeline.
4632                  */
4633                 if (endTLI != ThisTimeLineID)
4634                         XLogFileCopy(endLogId, endLogSeg,
4635                                                  endTLI, endLogId, endLogSeg);
4636         }
4637
4638         /*
4639          * Let's just make real sure there are not .ready or .done flags posted
4640          * for the new segment.
4641          */
4642         XLogFileName(xlogpath, ThisTimeLineID, endLogId, endLogSeg);
4643         XLogArchiveCleanup(xlogpath);
4644
4645         /* Get rid of any remaining recovered timeline-history file, too */
4646         snprintf(recoveryPath, MAXPGPATH, XLOGDIR "/RECOVERYHISTORY");
4647         unlink(recoveryPath);           /* ignore any error */
4648
4649         /*
4650          * Rename the config file out of the way, so that we don't accidentally
4651          * re-enter archive recovery mode in a subsequent crash.
4652          */
4653         unlink(RECOVERY_COMMAND_DONE);
4654         if (rename(RECOVERY_COMMAND_FILE, RECOVERY_COMMAND_DONE) != 0)
4655                 ereport(FATAL,
4656                                 (errcode_for_file_access(),
4657                                  errmsg("could not rename file \"%s\" to \"%s\": %m",
4658                                                 RECOVERY_COMMAND_FILE, RECOVERY_COMMAND_DONE)));
4659
4660         ereport(LOG,
4661                         (errmsg("archive recovery complete")));
4662 }
4663
4664 /*
4665  * For point-in-time recovery, this function decides whether we want to
4666  * stop applying the XLOG at or after the current record.
4667  *
4668  * Returns TRUE if we are stopping, FALSE otherwise.  On TRUE return,
4669  * *includeThis is set TRUE if we should apply this record before stopping.
4670  * Also, some information is saved in recoveryStopXid et al for use in
4671  * annotating the new timeline's history file.
4672  */
4673 static bool
4674 recoveryStopsHere(XLogRecord *record, bool *includeThis)
4675 {
4676         bool            stopsHere;
4677         uint8           record_info;
4678         TimestampTz recordXtime;
4679
4680         /* We only consider stopping at COMMIT or ABORT records */
4681         if (record->xl_rmid != RM_XACT_ID)
4682                 return false;
4683         record_info = record->xl_info & ~XLR_INFO_MASK;
4684         if (record_info == XLOG_XACT_COMMIT)
4685         {
4686                 xl_xact_commit *recordXactCommitData;
4687
4688                 recordXactCommitData = (xl_xact_commit *) XLogRecGetData(record);
4689                 recordXtime = recordXactCommitData->xact_time;
4690         }
4691         else if (record_info == XLOG_XACT_ABORT)
4692         {
4693                 xl_xact_abort *recordXactAbortData;
4694
4695                 recordXactAbortData = (xl_xact_abort *) XLogRecGetData(record);
4696                 recordXtime = recordXactAbortData->xact_time;
4697         }
4698         else
4699                 return false;
4700
4701         /* Remember the most recent COMMIT/ABORT time for logging purposes */
4702         recoveryLastXTime = recordXtime;
4703
4704         /* Do we have a PITR target at all? */
4705         if (!recoveryTarget)
4706                 return false;
4707
4708         if (recoveryTargetExact)
4709         {
4710                 /*
4711                  * there can be only one transaction end record with this exact
4712                  * transactionid
4713                  *
4714                  * when testing for an xid, we MUST test for equality only, since
4715                  * transactions are numbered in the order they start, not the order
4716                  * they complete. A higher numbered xid will complete before you about
4717                  * 50% of the time...
4718                  */
4719                 stopsHere = (record->xl_xid == recoveryTargetXid);
4720                 if (stopsHere)
4721                         *includeThis = recoveryTargetInclusive;
4722         }
4723         else
4724         {
4725                 /*
4726                  * there can be many transactions that share the same commit time, so
4727                  * we stop after the last one, if we are inclusive, or stop at the
4728                  * first one if we are exclusive
4729                  */
4730                 if (recoveryTargetInclusive)
4731                         stopsHere = (recordXtime > recoveryTargetTime);
4732                 else
4733                         stopsHere = (recordXtime >= recoveryTargetTime);
4734                 if (stopsHere)
4735                         *includeThis = false;
4736         }
4737
4738         if (stopsHere)
4739         {
4740                 recoveryStopXid = record->xl_xid;
4741                 recoveryStopTime = recordXtime;
4742                 recoveryStopAfter = *includeThis;
4743
4744                 if (record_info == XLOG_XACT_COMMIT)
4745                 {
4746                         if (recoveryStopAfter)
4747                                 ereport(LOG,
4748                                                 (errmsg("recovery stopping after commit of transaction %u, time %s",
4749                                                                 recoveryStopXid,
4750                                                                 timestamptz_to_str(recoveryStopTime))));
4751                         else
4752                                 ereport(LOG,
4753                                                 (errmsg("recovery stopping before commit of transaction %u, time %s",
4754                                                                 recoveryStopXid,
4755                                                                 timestamptz_to_str(recoveryStopTime))));
4756                 }
4757                 else
4758                 {
4759                         if (recoveryStopAfter)
4760                                 ereport(LOG,
4761                                                 (errmsg("recovery stopping after abort of transaction %u, time %s",
4762                                                                 recoveryStopXid,
4763                                                                 timestamptz_to_str(recoveryStopTime))));
4764                         else
4765                                 ereport(LOG,
4766                                                 (errmsg("recovery stopping before abort of transaction %u, time %s",
4767                                                                 recoveryStopXid,
4768                                                                 timestamptz_to_str(recoveryStopTime))));
4769                 }
4770         }
4771
4772         return stopsHere;
4773 }
4774
4775 /*
4776  * This must be called ONCE during postmaster or standalone-backend startup
4777  */
4778 void
4779 StartupXLOG(void)
4780 {
4781         XLogCtlInsert *Insert;
4782         CheckPoint      checkPoint;
4783         bool            wasShutdown;
4784         bool            reachedStopPoint = false;
4785         bool            haveBackupLabel = false;
4786         XLogRecPtr      RecPtr,
4787                                 LastRec,
4788                                 checkPointLoc,
4789                                 minRecoveryLoc,
4790                                 EndOfLog;
4791         uint32          endLogId;
4792         uint32          endLogSeg;
4793         XLogRecord *record;
4794         uint32          freespace;
4795         TransactionId oldestActiveXID;
4796
4797         /*
4798          * Read control file and check XLOG status looks valid.
4799          *
4800          * Note: in most control paths, *ControlFile is already valid and we need
4801          * not do ReadControlFile() here, but might as well do it to be sure.
4802          */
4803         ReadControlFile();
4804
4805         if (ControlFile->state < DB_SHUTDOWNED ||
4806                 ControlFile->state > DB_IN_PRODUCTION ||
4807                 !XRecOffIsValid(ControlFile->checkPoint.xrecoff))
4808                 ereport(FATAL,
4809                                 (errmsg("control file contains invalid data")));
4810
4811         if (ControlFile->state == DB_SHUTDOWNED)
4812                 ereport(LOG,
4813                                 (errmsg("database system was shut down at %s",
4814                                                 str_time(ControlFile->time))));
4815         else if (ControlFile->state == DB_SHUTDOWNING)
4816                 ereport(LOG,
4817                                 (errmsg("database system shutdown was interrupted; last known up at %s",
4818                                                 str_time(ControlFile->time))));
4819         else if (ControlFile->state == DB_IN_CRASH_RECOVERY)
4820                 ereport(LOG,
4821                    (errmsg("database system was interrupted while in recovery at %s",
4822                                    str_time(ControlFile->time)),
4823                         errhint("This probably means that some data is corrupted and"
4824                                         " you will have to use the last backup for recovery.")));
4825         else if (ControlFile->state == DB_IN_ARCHIVE_RECOVERY)
4826                 ereport(LOG,
4827                                 (errmsg("database system was interrupted while in recovery at log time %s",
4828                                                 str_time(ControlFile->checkPointCopy.time)),
4829                                  errhint("If this has occurred more than once some data might be corrupted"
4830                           " and you might need to choose an earlier recovery target.")));
4831         else if (ControlFile->state == DB_IN_PRODUCTION)
4832                 ereport(LOG,
4833                           (errmsg("database system was interrupted; last known up at %s",
4834                                           str_time(ControlFile->time))));
4835
4836         /* This is just to allow attaching to startup process with a debugger */
4837 #ifdef XLOG_REPLAY_DELAY
4838         if (ControlFile->state != DB_SHUTDOWNED)
4839                 pg_usleep(60000000L);
4840 #endif
4841
4842         /*
4843          * Initialize on the assumption we want to recover to the same timeline
4844          * that's active according to pg_control.
4845          */
4846         recoveryTargetTLI = ControlFile->checkPointCopy.ThisTimeLineID;
4847
4848         /*
4849          * Check for recovery control file, and if so set up state for offline
4850          * recovery
4851          */
4852         readRecoveryCommandFile();
4853
4854         /* Now we can determine the list of expected TLIs */
4855         expectedTLIs = readTimeLineHistory(recoveryTargetTLI);
4856
4857         /*
4858          * If pg_control's timeline is not in expectedTLIs, then we cannot
4859          * proceed: the backup is not part of the history of the requested
4860          * timeline.
4861          */
4862         if (!list_member_int(expectedTLIs,
4863                                                  (int) ControlFile->checkPointCopy.ThisTimeLineID))
4864                 ereport(FATAL,
4865                                 (errmsg("requested timeline %u is not a child of database system timeline %u",
4866                                                 recoveryTargetTLI,
4867                                                 ControlFile->checkPointCopy.ThisTimeLineID)));
4868
4869         if (read_backup_label(&checkPointLoc, &minRecoveryLoc))
4870         {
4871                 /*
4872                  * When a backup_label file is present, we want to roll forward from
4873                  * the checkpoint it identifies, rather than using pg_control.
4874                  */
4875                 record = ReadCheckpointRecord(checkPointLoc, 0);
4876                 if (record != NULL)
4877                 {
4878                         ereport(DEBUG1,
4879                                         (errmsg("checkpoint record is at %X/%X",
4880                                                         checkPointLoc.xlogid, checkPointLoc.xrecoff)));
4881                         InRecovery = true;      /* force recovery even if SHUTDOWNED */
4882                 }
4883                 else
4884                 {
4885                         ereport(PANIC,
4886                                         (errmsg("could not locate required checkpoint record"),
4887                                          errhint("If you are not restoring from a backup, try removing the file \"%s/backup_label\".", DataDir)));
4888                 }
4889                 /* set flag to delete it later */
4890                 haveBackupLabel = true;
4891         }
4892         else
4893         {
4894                 /*
4895                  * Get the last valid checkpoint record.  If the latest one according
4896                  * to pg_control is broken, try the next-to-last one.
4897                  */
4898                 checkPointLoc = ControlFile->checkPoint;
4899                 record = ReadCheckpointRecord(checkPointLoc, 1);
4900                 if (record != NULL)
4901                 {
4902                         ereport(DEBUG1,
4903                                         (errmsg("checkpoint record is at %X/%X",
4904                                                         checkPointLoc.xlogid, checkPointLoc.xrecoff)));
4905                 }
4906                 else
4907                 {
4908                         checkPointLoc = ControlFile->prevCheckPoint;
4909                         record = ReadCheckpointRecord(checkPointLoc, 2);
4910                         if (record != NULL)
4911                         {
4912                                 ereport(LOG,
4913                                                 (errmsg("using previous checkpoint record at %X/%X",
4914                                                           checkPointLoc.xlogid, checkPointLoc.xrecoff)));
4915                                 InRecovery = true;              /* force recovery even if SHUTDOWNED */
4916                         }
4917                         else
4918                                 ereport(PANIC,
4919                                          (errmsg("could not locate a valid checkpoint record")));
4920                 }
4921         }
4922
4923         LastRec = RecPtr = checkPointLoc;
4924         memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
4925         wasShutdown = (record->xl_info == XLOG_CHECKPOINT_SHUTDOWN);
4926
4927         ereport(DEBUG1,
4928                         (errmsg("redo record is at %X/%X; shutdown %s",
4929                                         checkPoint.redo.xlogid, checkPoint.redo.xrecoff,
4930                                         wasShutdown ? "TRUE" : "FALSE")));
4931         ereport(DEBUG1,
4932                         (errmsg("next transaction ID: %u/%u; next OID: %u",
4933                                         checkPoint.nextXidEpoch, checkPoint.nextXid,
4934                                         checkPoint.nextOid)));
4935         ereport(DEBUG1,
4936                         (errmsg("next MultiXactId: %u; next MultiXactOffset: %u",
4937                                         checkPoint.nextMulti, checkPoint.nextMultiOffset)));
4938         if (!TransactionIdIsNormal(checkPoint.nextXid))
4939                 ereport(PANIC,
4940                                 (errmsg("invalid next transaction ID")));
4941
4942         ShmemVariableCache->nextXid = checkPoint.nextXid;
4943         ShmemVariableCache->nextOid = checkPoint.nextOid;
4944         ShmemVariableCache->oidCount = 0;
4945         MultiXactSetNextMXact(checkPoint.nextMulti, checkPoint.nextMultiOffset);
4946
4947         /*
4948          * We must replay WAL entries using the same TimeLineID they were created
4949          * under, so temporarily adopt the TLI indicated by the checkpoint (see
4950          * also xlog_redo()).
4951          */
4952         ThisTimeLineID = checkPoint.ThisTimeLineID;
4953
4954         RedoRecPtr = XLogCtl->Insert.RedoRecPtr = checkPoint.redo;
4955
4956         if (XLByteLT(RecPtr, checkPoint.redo))
4957                 ereport(PANIC,
4958                                 (errmsg("invalid redo in checkpoint record")));
4959
4960         /*
4961          * Check whether we need to force recovery from WAL.  If it appears to
4962          * have been a clean shutdown and we did not have a recovery.conf file,
4963          * then assume no recovery needed.
4964          */
4965         if (XLByteLT(checkPoint.redo, RecPtr))
4966         {
4967                 if (wasShutdown)
4968                         ereport(PANIC,
4969                                         (errmsg("invalid redo record in shutdown checkpoint")));
4970                 InRecovery = true;
4971         }
4972         else if (ControlFile->state != DB_SHUTDOWNED)
4973                 InRecovery = true;
4974         else if (InArchiveRecovery)
4975         {
4976                 /* force recovery due to presence of recovery.conf */
4977                 InRecovery = true;
4978         }
4979
4980         /* REDO */
4981         if (InRecovery)
4982         {
4983                 int                     rmid;
4984
4985                 /*
4986                  * Update pg_control to show that we are recovering and to show the
4987                  * selected checkpoint as the place we are starting from. We also mark
4988                  * pg_control with any minimum recovery stop point obtained from a
4989                  * backup history file.
4990                  */
4991                 if (InArchiveRecovery)
4992                 {
4993                         ereport(LOG,
4994                                         (errmsg("automatic recovery in progress")));
4995                         ControlFile->state = DB_IN_ARCHIVE_RECOVERY;
4996                 }
4997                 else
4998                 {
4999                         ereport(LOG,
5000                                         (errmsg("database system was not properly shut down; "
5001                                                         "automatic recovery in progress")));
5002                         ControlFile->state = DB_IN_CRASH_RECOVERY;
5003                 }
5004                 ControlFile->prevCheckPoint = ControlFile->checkPoint;
5005                 ControlFile->checkPoint = checkPointLoc;
5006                 ControlFile->checkPointCopy = checkPoint;
5007                 if (minRecoveryLoc.xlogid != 0 || minRecoveryLoc.xrecoff != 0)
5008                         ControlFile->minRecoveryPoint = minRecoveryLoc;
5009                 ControlFile->time = (pg_time_t) time(NULL);
5010                 UpdateControlFile();
5011
5012                 /*
5013                  * If there was a backup label file, it's done its job and the info
5014                  * has now been propagated into pg_control.  We must get rid of the
5015                  * label file so that if we crash during recovery, we'll pick up at
5016                  * the latest recovery restartpoint instead of going all the way back
5017                  * to the backup start point.  It seems prudent though to just rename
5018                  * the file out of the way rather than delete it completely.
5019                  */
5020                 if (haveBackupLabel)
5021                 {
5022                         unlink(BACKUP_LABEL_OLD);
5023                         if (rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD) != 0)
5024                                 ereport(FATAL,
5025                                                 (errcode_for_file_access(),
5026                                                  errmsg("could not rename file \"%s\" to \"%s\": %m",
5027                                                                 BACKUP_LABEL_FILE, BACKUP_LABEL_OLD)));
5028                 }
5029
5030                 /* Start up the recovery environment */
5031                 XLogInitRelationCache();
5032
5033                 for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5034                 {
5035                         if (RmgrTable[rmid].rm_startup != NULL)
5036                                 RmgrTable[rmid].rm_startup();
5037                 }
5038
5039                 /*
5040                  * Find the first record that logically follows the checkpoint --- it
5041                  * might physically precede it, though.
5042                  */
5043                 if (XLByteLT(checkPoint.redo, RecPtr))
5044                 {
5045                         /* back up to find the record */
5046                         record = ReadRecord(&(checkPoint.redo), PANIC);
5047                 }
5048                 else
5049                 {
5050                         /* just have to read next record after CheckPoint */
5051                         record = ReadRecord(NULL, LOG);
5052                 }
5053
5054                 if (record != NULL)
5055                 {
5056                         bool            recoveryContinue = true;
5057                         bool            recoveryApply = true;
5058                         ErrorContextCallback errcontext;
5059
5060                         InRedo = true;
5061                         ereport(LOG,
5062                                         (errmsg("redo starts at %X/%X",
5063                                                         ReadRecPtr.xlogid, ReadRecPtr.xrecoff)));
5064
5065                         /*
5066                          * main redo apply loop
5067                          */
5068                         do
5069                         {
5070 #ifdef WAL_DEBUG
5071                                 if (XLOG_DEBUG)
5072                                 {
5073                                         StringInfoData buf;
5074
5075                                         initStringInfo(&buf);
5076                                         appendStringInfo(&buf, "REDO @ %X/%X; LSN %X/%X: ",
5077                                                                          ReadRecPtr.xlogid, ReadRecPtr.xrecoff,
5078                                                                          EndRecPtr.xlogid, EndRecPtr.xrecoff);
5079                                         xlog_outrec(&buf, record);
5080                                         appendStringInfo(&buf, " - ");
5081                                         RmgrTable[record->xl_rmid].rm_desc(&buf,
5082                                                                                                            record->xl_info,
5083                                                                                                          XLogRecGetData(record));
5084                                         elog(LOG, "%s", buf.data);
5085                                         pfree(buf.data);
5086                                 }
5087 #endif
5088
5089                                 /*
5090                                  * Have we reached our recovery target?
5091                                  */
5092                                 if (recoveryStopsHere(record, &recoveryApply))
5093                                 {
5094                                         reachedStopPoint = true;        /* see below */
5095                                         recoveryContinue = false;
5096                                         if (!recoveryApply)
5097                                                 break;
5098                                 }
5099
5100                                 /* Setup error traceback support for ereport() */
5101                                 errcontext.callback = rm_redo_error_callback;
5102                                 errcontext.arg = (void *) record;
5103                                 errcontext.previous = error_context_stack;
5104                                 error_context_stack = &errcontext;
5105
5106                                 /* nextXid must be beyond record's xid */
5107                                 if (TransactionIdFollowsOrEquals(record->xl_xid,
5108                                                                                                  ShmemVariableCache->nextXid))
5109                                 {
5110                                         ShmemVariableCache->nextXid = record->xl_xid;
5111                                         TransactionIdAdvance(ShmemVariableCache->nextXid);
5112                                 }
5113
5114                                 if (record->xl_info & XLR_BKP_BLOCK_MASK)
5115                                         RestoreBkpBlocks(record, EndRecPtr);
5116
5117                                 RmgrTable[record->xl_rmid].rm_redo(EndRecPtr, record);
5118
5119                                 /* Pop the error context stack */
5120                                 error_context_stack = errcontext.previous;
5121
5122                                 LastRec = ReadRecPtr;
5123
5124                                 record = ReadRecord(NULL, LOG);
5125                         } while (record != NULL && recoveryContinue);
5126
5127                         /*
5128                          * end of main redo apply loop
5129                          */
5130
5131                         ereport(LOG,
5132                                         (errmsg("redo done at %X/%X",
5133                                                         ReadRecPtr.xlogid, ReadRecPtr.xrecoff)));
5134                         if (recoveryLastXTime)
5135                                 ereport(LOG,
5136                                          (errmsg("last completed transaction was at log time %s",
5137                                                          timestamptz_to_str(recoveryLastXTime))));
5138                         InRedo = false;
5139                 }
5140                 else
5141                 {
5142                         /* there are no WAL records following the checkpoint */
5143                         ereport(LOG,
5144                                         (errmsg("redo is not required")));
5145                 }
5146         }
5147
5148         /*
5149          * Re-fetch the last valid or last applied record, so we can identify the
5150          * exact endpoint of what we consider the valid portion of WAL.
5151          */
5152         record = ReadRecord(&LastRec, PANIC);
5153         EndOfLog = EndRecPtr;
5154         XLByteToPrevSeg(EndOfLog, endLogId, endLogSeg);
5155
5156         /*
5157          * Complain if we did not roll forward far enough to render the backup
5158          * dump consistent.
5159          */
5160         if (XLByteLT(EndOfLog, ControlFile->minRecoveryPoint))
5161         {
5162                 if (reachedStopPoint)   /* stopped because of stop request */
5163                         ereport(FATAL,
5164                                         (errmsg("requested recovery stop point is before end time of backup dump")));
5165                 else    /* ran off end of WAL */
5166                         ereport(FATAL,
5167                                         (errmsg("WAL ends before end time of backup dump")));
5168         }
5169
5170         /*
5171          * Consider whether we need to assign a new timeline ID.
5172          *
5173          * If we are doing an archive recovery, we always assign a new ID.      This
5174          * handles a couple of issues.  If we stopped short of the end of WAL
5175          * during recovery, then we are clearly generating a new timeline and must
5176          * assign it a unique new ID.  Even if we ran to the end, modifying the
5177          * current last segment is problematic because it may result in trying to
5178          * overwrite an already-archived copy of that segment, and we encourage
5179          * DBAs to make their archive_commands reject that.  We can dodge the
5180          * problem by making the new active segment have a new timeline ID.
5181          *
5182          * In a normal crash recovery, we can just extend the timeline we were in.
5183          */
5184         if (InArchiveRecovery)
5185         {
5186                 ThisTimeLineID = findNewestTimeLine(recoveryTargetTLI) + 1;
5187                 ereport(LOG,
5188                                 (errmsg("selected new timeline ID: %u", ThisTimeLineID)));
5189                 writeTimeLineHistory(ThisTimeLineID, recoveryTargetTLI,
5190                                                          curFileTLI, endLogId, endLogSeg);
5191         }
5192
5193         /* Save the selected TimeLineID in shared memory, too */
5194         XLogCtl->ThisTimeLineID = ThisTimeLineID;
5195
5196         /*
5197          * We are now done reading the old WAL.  Turn off archive fetching if it
5198          * was active, and make a writable copy of the last WAL segment. (Note
5199          * that we also have a copy of the last block of the old WAL in readBuf;
5200          * we will use that below.)
5201          */
5202         if (InArchiveRecovery)
5203                 exitArchiveRecovery(curFileTLI, endLogId, endLogSeg);
5204
5205         /*
5206          * Prepare to write WAL starting at EndOfLog position, and init xlog
5207          * buffer cache using the block containing the last record from the
5208          * previous incarnation.
5209          */
5210         openLogId = endLogId;
5211         openLogSeg = endLogSeg;
5212         openLogFile = XLogFileOpen(openLogId, openLogSeg);
5213         openLogOff = 0;
5214         Insert = &XLogCtl->Insert;
5215         Insert->PrevRecord = LastRec;
5216         XLogCtl->xlblocks[0].xlogid = openLogId;
5217         XLogCtl->xlblocks[0].xrecoff =
5218                 ((EndOfLog.xrecoff - 1) / XLOG_BLCKSZ + 1) * XLOG_BLCKSZ;
5219
5220         /*
5221          * Tricky point here: readBuf contains the *last* block that the LastRec
5222          * record spans, not the one it starts in.      The last block is indeed the
5223          * one we want to use.
5224          */
5225         Assert(readOff == (XLogCtl->xlblocks[0].xrecoff - XLOG_BLCKSZ) % XLogSegSize);
5226         memcpy((char *) Insert->currpage, readBuf, XLOG_BLCKSZ);
5227         Insert->currpos = (char *) Insert->currpage +
5228                 (EndOfLog.xrecoff + XLOG_BLCKSZ - XLogCtl->xlblocks[0].xrecoff);
5229
5230         LogwrtResult.Write = LogwrtResult.Flush = EndOfLog;
5231
5232         XLogCtl->Write.LogwrtResult = LogwrtResult;
5233         Insert->LogwrtResult = LogwrtResult;
5234         XLogCtl->LogwrtResult = LogwrtResult;
5235
5236         XLogCtl->LogwrtRqst.Write = EndOfLog;
5237         XLogCtl->LogwrtRqst.Flush = EndOfLog;
5238
5239         freespace = INSERT_FREESPACE(Insert);
5240         if (freespace > 0)
5241         {
5242                 /* Make sure rest of page is zero */
5243                 MemSet(Insert->currpos, 0, freespace);
5244                 XLogCtl->Write.curridx = 0;
5245         }
5246         else
5247         {
5248                 /*
5249                  * Whenever Write.LogwrtResult points to exactly the end of a page,
5250                  * Write.curridx must point to the *next* page (see XLogWrite()).
5251                  *
5252                  * Note: it might seem we should do AdvanceXLInsertBuffer() here, but
5253                  * this is sufficient.  The first actual attempt to insert a log
5254                  * record will advance the insert state.
5255                  */
5256                 XLogCtl->Write.curridx = NextBufIdx(0);
5257         }
5258
5259         /* Pre-scan prepared transactions to find out the range of XIDs present */
5260         oldestActiveXID = PrescanPreparedTransactions();
5261
5262         if (InRecovery)
5263         {
5264                 int                     rmid;
5265
5266                 /*
5267                  * Allow resource managers to do any required cleanup.
5268                  */
5269                 for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
5270                 {
5271                         if (RmgrTable[rmid].rm_cleanup != NULL)
5272                                 RmgrTable[rmid].rm_cleanup();
5273                 }
5274
5275                 /*
5276                  * Check to see if the XLOG sequence contained any unresolved
5277                  * references to uninitialized pages.
5278                  */
5279                 XLogCheckInvalidPages();
5280
5281                 /*
5282                  * Reset pgstat data, because it may be invalid after recovery.
5283                  */
5284                 pgstat_reset_all();
5285
5286                 /*
5287                  * Perform a checkpoint to update all our recovery activity to disk.
5288                  *
5289                  * Note that we write a shutdown checkpoint rather than an on-line
5290                  * one. This is not particularly critical, but since we may be
5291                  * assigning a new TLI, using a shutdown checkpoint allows us to have
5292                  * the rule that TLI only changes in shutdown checkpoints, which
5293                  * allows some extra error checking in xlog_redo.
5294                  */
5295                 CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
5296
5297                 /*
5298                  * Close down recovery environment
5299                  */
5300                 XLogCloseRelationCache();
5301         }
5302
5303         /*
5304          * Preallocate additional log files, if wanted.
5305          */
5306         PreallocXlogFiles(EndOfLog);
5307
5308         /*
5309          * Okay, we're officially UP.
5310          */
5311         InRecovery = false;
5312
5313         ControlFile->state = DB_IN_PRODUCTION;
5314         ControlFile->time = (pg_time_t) time(NULL);
5315         UpdateControlFile();
5316
5317         /* start the archive_timeout timer running */
5318         XLogCtl->Write.lastSegSwitchTime = ControlFile->time;
5319
5320         /* initialize shared-memory copy of latest checkpoint XID/epoch */
5321         XLogCtl->ckptXidEpoch = ControlFile->checkPointCopy.nextXidEpoch;
5322         XLogCtl->ckptXid = ControlFile->checkPointCopy.nextXid;
5323
5324         /* also initialize latestCompletedXid, to nextXid - 1 */
5325         ShmemVariableCache->latestCompletedXid = ShmemVariableCache->nextXid;
5326         TransactionIdRetreat(ShmemVariableCache->latestCompletedXid);
5327
5328         /* Start up the commit log and related stuff, too */
5329         StartupCLOG();
5330         StartupSUBTRANS(oldestActiveXID);
5331         StartupMultiXact();
5332
5333         /* Reload shared-memory state for prepared transactions */
5334         RecoverPreparedTransactions();
5335
5336         /* Shut down readFile facility, free space */
5337         if (readFile >= 0)
5338         {
5339                 close(readFile);
5340                 readFile = -1;
5341         }
5342         if (readBuf)
5343         {
5344                 free(readBuf);
5345                 readBuf = NULL;
5346         }
5347         if (readRecordBuf)
5348         {
5349                 free(readRecordBuf);
5350                 readRecordBuf = NULL;
5351                 readRecordBufSize = 0;
5352         }
5353 }
5354
5355 /*
5356  * Subroutine to try to fetch and validate a prior checkpoint record.
5357  *
5358  * whichChkpt identifies the checkpoint (merely for reporting purposes).
5359  * 1 for "primary", 2 for "secondary", 0 for "other" (backup_label)
5360  */
5361 static XLogRecord *
5362 ReadCheckpointRecord(XLogRecPtr RecPtr, int whichChkpt)
5363 {
5364         XLogRecord *record;
5365
5366         if (!XRecOffIsValid(RecPtr.xrecoff))
5367         {
5368                 switch (whichChkpt)
5369                 {
5370                         case 1:
5371                                 ereport(LOG,
5372                                 (errmsg("invalid primary checkpoint link in control file")));
5373                                 break;
5374                         case 2:
5375                                 ereport(LOG,
5376                                                 (errmsg("invalid secondary checkpoint link in control file")));
5377                                 break;
5378                         default:
5379                                 ereport(LOG,
5380                                    (errmsg("invalid checkpoint link in backup_label file")));
5381                                 break;
5382                 }
5383                 return NULL;
5384         }
5385
5386         record = ReadRecord(&RecPtr, LOG);
5387
5388         if (record == NULL)
5389         {
5390                 switch (whichChkpt)
5391                 {
5392                         case 1:
5393                                 ereport(LOG,
5394                                                 (errmsg("invalid primary checkpoint record")));
5395                                 break;
5396                         case 2:
5397                                 ereport(LOG,
5398                                                 (errmsg("invalid secondary checkpoint record")));
5399                                 break;
5400                         default:
5401                                 ereport(LOG,
5402                                                 (errmsg("invalid checkpoint record")));
5403                                 break;
5404                 }
5405                 return NULL;
5406         }
5407         if (record->xl_rmid != RM_XLOG_ID)
5408         {
5409                 switch (whichChkpt)
5410                 {
5411                         case 1:
5412                                 ereport(LOG,
5413                                                 (errmsg("invalid resource manager ID in primary checkpoint record")));
5414                                 break;
5415                         case 2:
5416                                 ereport(LOG,
5417                                                 (errmsg("invalid resource manager ID in secondary checkpoint record")));
5418                                 break;
5419                         default:
5420                                 ereport(LOG,
5421                                 (errmsg("invalid resource manager ID in checkpoint record")));
5422                                 break;
5423                 }
5424                 return NULL;
5425         }
5426         if (record->xl_info != XLOG_CHECKPOINT_SHUTDOWN &&
5427                 record->xl_info != XLOG_CHECKPOINT_ONLINE)
5428         {
5429                 switch (whichChkpt)
5430                 {
5431                         case 1:
5432                                 ereport(LOG,
5433                                    (errmsg("invalid xl_info in primary checkpoint record")));
5434                                 break;
5435                         case 2:
5436                                 ereport(LOG,
5437                                  (errmsg("invalid xl_info in secondary checkpoint record")));
5438                                 break;
5439                         default:
5440                                 ereport(LOG,
5441                                                 (errmsg("invalid xl_info in checkpoint record")));
5442                                 break;
5443                 }
5444                 return NULL;
5445         }
5446         if (record->xl_len != sizeof(CheckPoint) ||
5447                 record->xl_tot_len != SizeOfXLogRecord + sizeof(CheckPoint))
5448         {
5449                 switch (whichChkpt)
5450                 {
5451                         case 1:
5452                                 ereport(LOG,
5453                                         (errmsg("invalid length of primary checkpoint record")));
5454                                 break;
5455                         case 2:
5456                                 ereport(LOG,
5457                                   (errmsg("invalid length of secondary checkpoint record")));
5458                                 break;
5459                         default:
5460                                 ereport(LOG,
5461                                                 (errmsg("invalid length of checkpoint record")));
5462                                 break;
5463                 }
5464                 return NULL;
5465         }
5466         return record;
5467 }
5468
5469 /*
5470  * This must be called during startup of a backend process, except that
5471  * it need not be called in a standalone backend (which does StartupXLOG
5472  * instead).  We need to initialize the local copies of ThisTimeLineID and
5473  * RedoRecPtr.
5474  *
5475  * Note: before Postgres 8.0, we went to some effort to keep the postmaster
5476  * process's copies of ThisTimeLineID and RedoRecPtr valid too.  This was
5477  * unnecessary however, since the postmaster itself never touches XLOG anyway.
5478  */
5479 void
5480 InitXLOGAccess(void)
5481 {
5482         /* ThisTimeLineID doesn't change so we need no lock to copy it */
5483         ThisTimeLineID = XLogCtl->ThisTimeLineID;
5484         /* Use GetRedoRecPtr to copy the RedoRecPtr safely */
5485         (void) GetRedoRecPtr();
5486 }
5487
5488 /*
5489  * Once spawned, a backend may update its local RedoRecPtr from
5490  * XLogCtl->Insert.RedoRecPtr; it must hold the insert lock or info_lck
5491  * to do so.  This is done in XLogInsert() or GetRedoRecPtr().
5492  */
5493 XLogRecPtr
5494 GetRedoRecPtr(void)
5495 {
5496         /* use volatile pointer to prevent code rearrangement */
5497         volatile XLogCtlData *xlogctl = XLogCtl;
5498
5499         SpinLockAcquire(&xlogctl->info_lck);
5500         Assert(XLByteLE(RedoRecPtr, xlogctl->Insert.RedoRecPtr));
5501         RedoRecPtr = xlogctl->Insert.RedoRecPtr;
5502         SpinLockRelease(&xlogctl->info_lck);
5503
5504         return RedoRecPtr;
5505 }
5506
5507 /*
5508  * GetInsertRecPtr -- Returns the current insert position.
5509  *
5510  * NOTE: The value *actually* returned is the position of the last full
5511  * xlog page. It lags behind the real insert position by at most 1 page.
5512  * For that, we don't need to acquire WALInsertLock which can be quite
5513  * heavily contended, and an approximation is enough for the current
5514  * usage of this function.
5515  */
5516 XLogRecPtr
5517 GetInsertRecPtr(void)
5518 {
5519         /* use volatile pointer to prevent code rearrangement */
5520         volatile XLogCtlData *xlogctl = XLogCtl;
5521         XLogRecPtr      recptr;
5522
5523         SpinLockAcquire(&xlogctl->info_lck);
5524         recptr = xlogctl->LogwrtRqst.Write;
5525         SpinLockRelease(&xlogctl->info_lck);
5526
5527         return recptr;
5528 }
5529
5530 /*
5531  * Get the time of the last xlog segment switch
5532  */
5533 pg_time_t
5534 GetLastSegSwitchTime(void)
5535 {
5536         pg_time_t       result;
5537
5538         /* Need WALWriteLock, but shared lock is sufficient */
5539         LWLockAcquire(WALWriteLock, LW_SHARED);
5540         result = XLogCtl->Write.lastSegSwitchTime;
5541         LWLockRelease(WALWriteLock);
5542
5543         return result;
5544 }
5545
5546 /*
5547  * GetNextXidAndEpoch - get the current nextXid value and associated epoch
5548  *
5549  * This is exported for use by code that would like to have 64-bit XIDs.
5550  * We don't really support such things, but all XIDs within the system
5551  * can be presumed "close to" the result, and thus the epoch associated
5552  * with them can be determined.
5553  */
5554 void
5555 GetNextXidAndEpoch(TransactionId *xid, uint32 *epoch)
5556 {
5557         uint32          ckptXidEpoch;
5558         TransactionId ckptXid;
5559         TransactionId nextXid;
5560
5561         /* Must read checkpoint info first, else have race condition */
5562         {
5563                 /* use volatile pointer to prevent code rearrangement */
5564                 volatile XLogCtlData *xlogctl = XLogCtl;
5565
5566                 SpinLockAcquire(&xlogctl->info_lck);
5567                 ckptXidEpoch = xlogctl->ckptXidEpoch;
5568                 ckptXid = xlogctl->ckptXid;
5569                 SpinLockRelease(&xlogctl->info_lck);
5570         }
5571
5572         /* Now fetch current nextXid */
5573         nextXid = ReadNewTransactionId();
5574
5575         /*
5576          * nextXid is certainly logically later than ckptXid.  So if it's
5577          * numerically less, it must have wrapped into the next epoch.
5578          */
5579         if (nextXid < ckptXid)
5580                 ckptXidEpoch++;
5581
5582         *xid = nextXid;
5583         *epoch = ckptXidEpoch;
5584 }
5585
5586 /*
5587  * This must be called ONCE during postmaster or standalone-backend shutdown
5588  */
5589 void
5590 ShutdownXLOG(int code, Datum arg)
5591 {
5592         ereport(LOG,
5593                         (errmsg("shutting down")));
5594
5595         CreateCheckPoint(CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_IMMEDIATE);
5596         ShutdownCLOG();
5597         ShutdownSUBTRANS();
5598         ShutdownMultiXact();
5599
5600         ereport(LOG,
5601                         (errmsg("database system is shut down")));
5602 }
5603
5604 /*
5605  * Log start of a checkpoint.
5606  */
5607 static void
5608 LogCheckpointStart(int flags)
5609 {
5610         elog(LOG, "checkpoint starting:%s%s%s%s%s%s",
5611                  (flags & CHECKPOINT_IS_SHUTDOWN) ? " shutdown" : "",
5612                  (flags & CHECKPOINT_IMMEDIATE) ? " immediate" : "",
5613                  (flags & CHECKPOINT_FORCE) ? " force" : "",
5614                  (flags & CHECKPOINT_WAIT) ? " wait" : "",
5615                  (flags & CHECKPOINT_CAUSE_XLOG) ? " xlog" : "",
5616                  (flags & CHECKPOINT_CAUSE_TIME) ? " time" : "");
5617 }
5618
5619 /*
5620  * Log end of a checkpoint.
5621  */
5622 static void
5623 LogCheckpointEnd(void)
5624 {
5625         long            write_secs,
5626                                 sync_secs,
5627                                 total_secs;
5628         int                     write_usecs,
5629                                 sync_usecs,
5630                                 total_usecs;
5631
5632         CheckpointStats.ckpt_end_t = GetCurrentTimestamp();
5633
5634         TimestampDifference(CheckpointStats.ckpt_start_t,
5635                                                 CheckpointStats.ckpt_end_t,
5636                                                 &total_secs, &total_usecs);
5637
5638         TimestampDifference(CheckpointStats.ckpt_write_t,
5639                                                 CheckpointStats.ckpt_sync_t,
5640                                                 &write_secs, &write_usecs);
5641
5642         TimestampDifference(CheckpointStats.ckpt_sync_t,
5643                                                 CheckpointStats.ckpt_sync_end_t,
5644                                                 &sync_secs, &sync_usecs);
5645
5646         elog(LOG, "checkpoint complete: wrote %d buffers (%.1f%%); "
5647                  "%d transaction log file(s) added, %d removed, %d recycled; "
5648                  "write=%ld.%03d s, sync=%ld.%03d s, total=%ld.%03d s",
5649                  CheckpointStats.ckpt_bufs_written,
5650                  (double) CheckpointStats.ckpt_bufs_written * 100 / NBuffers,
5651                  CheckpointStats.ckpt_segs_added,
5652                  CheckpointStats.ckpt_segs_removed,
5653                  CheckpointStats.ckpt_segs_recycled,
5654                  write_secs, write_usecs / 1000,
5655                  sync_secs, sync_usecs / 1000,
5656                  total_secs, total_usecs / 1000);
5657 }
5658
5659 /*
5660  * Perform a checkpoint --- either during shutdown, or on-the-fly
5661  *
5662  * flags is a bitwise OR of the following:
5663  *      CHECKPOINT_IS_SHUTDOWN: checkpoint is for database shutdown.
5664  *      CHECKPOINT_IMMEDIATE: finish the checkpoint ASAP,
5665  *              ignoring checkpoint_completion_target parameter.
5666  *      CHECKPOINT_FORCE: force a checkpoint even if no XLOG activity has occured
5667  *              since the last one (implied by CHECKPOINT_IS_SHUTDOWN).
5668  *
5669  * Note: flags contains other bits, of interest here only for logging purposes.
5670  * In particular note that this routine is synchronous and does not pay
5671  * attention to CHECKPOINT_WAIT.
5672  */
5673 void
5674 CreateCheckPoint(int flags)
5675 {
5676         bool            shutdown = (flags & CHECKPOINT_IS_SHUTDOWN) != 0;
5677         CheckPoint      checkPoint;
5678         XLogRecPtr      recptr;
5679         XLogCtlInsert *Insert = &XLogCtl->Insert;
5680         XLogRecData rdata;
5681         uint32          freespace;
5682         uint32          _logId;
5683         uint32          _logSeg;
5684         TransactionId *inCommitXids;
5685         int                     nInCommit;
5686
5687         /*
5688          * Acquire CheckpointLock to ensure only one checkpoint happens at a time.
5689          * (This is just pro forma, since in the present system structure there is
5690          * only one process that is allowed to issue checkpoints at any given
5691          * time.)
5692          */
5693         LWLockAcquire(CheckpointLock, LW_EXCLUSIVE);
5694
5695         /*
5696          * Prepare to accumulate statistics.
5697          *
5698          * Note: because it is possible for log_checkpoints to change while a
5699          * checkpoint proceeds, we always accumulate stats, even if
5700          * log_checkpoints is currently off.
5701          */
5702         MemSet(&CheckpointStats, 0, sizeof(CheckpointStats));
5703         CheckpointStats.ckpt_start_t = GetCurrentTimestamp();
5704
5705         /*
5706          * Use a critical section to force system panic if we have trouble.
5707          */
5708         START_CRIT_SECTION();
5709
5710         if (shutdown)
5711         {
5712                 ControlFile->state = DB_SHUTDOWNING;
5713                 ControlFile->time = (pg_time_t) time(NULL);
5714                 UpdateControlFile();
5715         }
5716
5717         /*
5718          * Let smgr prepare for checkpoint; this has to happen before we determine
5719          * the REDO pointer.  Note that smgr must not do anything that'd have to
5720          * be undone if we decide no checkpoint is needed.
5721          */
5722         smgrpreckpt();
5723
5724         /* Begin filling in the checkpoint WAL record */
5725         MemSet(&checkPoint, 0, sizeof(checkPoint));
5726         checkPoint.ThisTimeLineID = ThisTimeLineID;
5727         checkPoint.time = (pg_time_t) time(NULL);
5728
5729         /*
5730          * We must hold WALInsertLock while examining insert state to determine
5731          * the checkpoint REDO pointer.
5732          */
5733         LWLockAcquire(WALInsertLock, LW_EXCLUSIVE);
5734
5735         /*
5736          * If this isn't a shutdown or forced checkpoint, and we have not inserted
5737          * any XLOG records since the start of the last checkpoint, skip the
5738          * checkpoint.  The idea here is to avoid inserting duplicate checkpoints
5739          * when the system is idle. That wastes log space, and more importantly it
5740          * exposes us to possible loss of both current and previous checkpoint
5741          * records if the machine crashes just as we're writing the update.
5742          * (Perhaps it'd make even more sense to checkpoint only when the previous
5743          * checkpoint record is in a different xlog page?)
5744          *
5745          * We have to make two tests to determine that nothing has happened since
5746          * the start of the last checkpoint: current insertion point must match
5747          * the end of the last checkpoint record, and its redo pointer must point
5748          * to itself.
5749          */
5750         if ((flags & (CHECKPOINT_IS_SHUTDOWN | CHECKPOINT_FORCE)) == 0)
5751         {
5752                 XLogRecPtr      curInsert;
5753
5754                 INSERT_RECPTR(curInsert, Insert, Insert->curridx);
5755                 if (curInsert.xlogid == ControlFile->checkPoint.xlogid &&
5756                         curInsert.xrecoff == ControlFile->checkPoint.xrecoff +
5757                         MAXALIGN(SizeOfXLogRecord + sizeof(CheckPoint)) &&
5758                         ControlFile->checkPoint.xlogid ==
5759                         ControlFile->checkPointCopy.redo.xlogid &&
5760                         ControlFile->checkPoint.xrecoff ==
5761                         ControlFile->checkPointCopy.redo.xrecoff)
5762                 {
5763                         LWLockRelease(WALInsertLock);
5764                         LWLockRelease(CheckpointLock);
5765                         END_CRIT_SECTION();
5766                         return;
5767                 }
5768         }
5769
5770         /*
5771          * Compute new REDO record ptr = location of next XLOG record.
5772          *
5773          * NB: this is NOT necessarily where the checkpoint record itself will be,
5774          * since other backends may insert more XLOG records while we're off doing
5775          * the buffer flush work.  Those XLOG records are logically after the
5776          * checkpoint, even though physically before it.  Got that?
5777          */
5778         freespace = INSERT_FREESPACE(Insert);
5779         if (freespace < SizeOfXLogRecord)
5780         {
5781                 (void) AdvanceXLInsertBuffer(false);
5782                 /* OK to ignore update return flag, since we will do flush anyway */
5783                 freespace = INSERT_FREESPACE(Insert);
5784         }
5785         INSERT_RECPTR(checkPoint.redo, Insert, Insert->curridx);
5786
5787         /*
5788          * Here we update the shared RedoRecPtr for future XLogInsert calls; this
5789          * must be done while holding the insert lock AND the info_lck.
5790          *
5791          * Note: if we fail to complete the checkpoint, RedoRecPtr will be left
5792          * pointing past where it really needs to point.  This is okay; the only
5793          * consequence is that XLogInsert might back up whole buffers that it
5794          * didn't really need to.  We can't postpone advancing RedoRecPtr because
5795          * XLogInserts that happen while we are dumping buffers must assume that
5796          * their buffer changes are not included in the checkpoint.
5797          */
5798         {
5799                 /* use volatile pointer to prevent code rearrangement */
5800                 volatile XLogCtlData *xlogctl = XLogCtl;
5801
5802                 SpinLockAcquire(&xlogctl->info_lck);
5803                 RedoRecPtr = xlogctl->Insert.RedoRecPtr = checkPoint.redo;
5804                 SpinLockRelease(&xlogctl->info_lck);
5805         }
5806
5807         /*
5808          * Now we can release WAL insert lock, allowing other xacts to proceed
5809          * while we are flushing disk buffers.
5810          */
5811         LWLockRelease(WALInsertLock);
5812
5813         /*
5814          * If enabled, log checkpoint start.  We postpone this until now so as not
5815          * to log anything if we decided to skip the checkpoint.
5816          */
5817         if (log_checkpoints)
5818                 LogCheckpointStart(flags);
5819
5820         /*
5821          * Before flushing data, we must wait for any transactions that are
5822          * currently in their commit critical sections.  If an xact inserted its
5823          * commit record into XLOG just before the REDO point, then a crash
5824          * restart from the REDO point would not replay that record, which means
5825          * that our flushing had better include the xact's update of pg_clog.  So
5826          * we wait till he's out of his commit critical section before proceeding.
5827          * See notes in RecordTransactionCommit().
5828          *
5829          * Because we've already released WALInsertLock, this test is a bit fuzzy:
5830          * it is possible that we will wait for xacts we didn't really need to
5831          * wait for.  But the delay should be short and it seems better to make
5832          * checkpoint take a bit longer than to hold locks longer than necessary.
5833          * (In fact, the whole reason we have this issue is that xact.c does
5834          * commit record XLOG insertion and clog update as two separate steps
5835          * protected by different locks, but again that seems best on grounds of
5836          * minimizing lock contention.)
5837          *
5838          * A transaction that has not yet set inCommit when we look cannot be at
5839          * risk, since he's not inserted his commit record yet; and one that's
5840          * already cleared it is not at risk either, since he's done fixing clog
5841          * and we will correctly flush the update below.  So we cannot miss any
5842          * xacts we need to wait for.
5843          */
5844         nInCommit = GetTransactionsInCommit(&inCommitXids);
5845         if (nInCommit > 0)
5846         {
5847                 do
5848                 {
5849                         pg_usleep(10000L);      /* wait for 10 msec */
5850                 } while (HaveTransactionsInCommit(inCommitXids, nInCommit));
5851         }
5852         pfree(inCommitXids);
5853
5854         /*
5855          * Get the other info we need for the checkpoint record.
5856          */
5857         LWLockAcquire(XidGenLock, LW_SHARED);
5858         checkPoint.nextXid = ShmemVariableCache->nextXid;
5859         LWLockRelease(XidGenLock);
5860
5861         /* Increase XID epoch if we've wrapped around since last checkpoint */
5862         checkPoint.nextXidEpoch = ControlFile->checkPointCopy.nextXidEpoch;
5863         if (checkPoint.nextXid < ControlFile->checkPointCopy.nextXid)
5864                 checkPoint.nextXidEpoch++;
5865
5866         LWLockAcquire(OidGenLock, LW_SHARED);
5867         checkPoint.nextOid = ShmemVariableCache->nextOid;
5868         if (!shutdown)
5869                 checkPoint.nextOid += ShmemVariableCache->oidCount;
5870         LWLockRelease(OidGenLock);
5871
5872         MultiXactGetCheckptMulti(shutdown,
5873                                                          &checkPoint.nextMulti,
5874                                                          &checkPoint.nextMultiOffset);
5875
5876         /*
5877          * Having constructed the checkpoint record, ensure all shmem disk buffers
5878          * and commit-log buffers are flushed to disk.
5879          *
5880          * This I/O could fail for various reasons.  If so, we will fail to
5881          * complete the checkpoint, but there is no reason to force a system
5882          * panic. Accordingly, exit critical section while doing it.
5883          */
5884         END_CRIT_SECTION();
5885
5886         CheckPointGuts(checkPoint.redo, flags);
5887
5888         START_CRIT_SECTION();
5889
5890         /*
5891          * Now insert the checkpoint record into XLOG.
5892          */
5893         rdata.data = (char *) (&checkPoint);
5894         rdata.len = sizeof(checkPoint);
5895         rdata.buffer = InvalidBuffer;
5896         rdata.next = NULL;
5897
5898         recptr = XLogInsert(RM_XLOG_ID,
5899                                                 shutdown ? XLOG_CHECKPOINT_SHUTDOWN :
5900                                                 XLOG_CHECKPOINT_ONLINE,
5901                                                 &rdata);
5902
5903         XLogFlush(recptr);
5904
5905         /*
5906          * We now have ProcLastRecPtr = start of actual checkpoint record, recptr
5907          * = end of actual checkpoint record.
5908          */
5909         if (shutdown && !XLByteEQ(checkPoint.redo, ProcLastRecPtr))
5910                 ereport(PANIC,
5911                                 (errmsg("concurrent transaction log activity while database system is shutting down")));
5912
5913         /*
5914          * Select point at which we can truncate the log, which we base on the
5915          * prior checkpoint's earliest info.
5916          */
5917         XLByteToSeg(ControlFile->checkPointCopy.redo, _logId, _logSeg);
5918
5919         /*
5920          * Update the control file.
5921          */
5922         LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
5923         if (shutdown)
5924                 ControlFile->state = DB_SHUTDOWNED;
5925         ControlFile->prevCheckPoint = ControlFile->checkPoint;
5926         ControlFile->checkPoint = ProcLastRecPtr;
5927         ControlFile->checkPointCopy = checkPoint;
5928         ControlFile->time = (pg_time_t) time(NULL);
5929         UpdateControlFile();
5930         LWLockRelease(ControlFileLock);
5931
5932         /* Update shared-memory copy of checkpoint XID/epoch */
5933         {
5934                 /* use volatile pointer to prevent code rearrangement */
5935                 volatile XLogCtlData *xlogctl = XLogCtl;
5936
5937                 SpinLockAcquire(&xlogctl->info_lck);
5938                 xlogctl->ckptXidEpoch = checkPoint.nextXidEpoch;
5939                 xlogctl->ckptXid = checkPoint.nextXid;
5940                 SpinLockRelease(&xlogctl->info_lck);
5941         }
5942
5943         /*
5944          * We are now done with critical updates; no need for system panic if we
5945          * have trouble while fooling with old log segments.
5946          */
5947         END_CRIT_SECTION();
5948
5949         /*
5950          * Let smgr do post-checkpoint cleanup (eg, deleting old files).
5951          */
5952         smgrpostckpt();
5953
5954         /*
5955          * Delete old log files (those no longer needed even for previous
5956          * checkpoint).
5957          */
5958         if (_logId || _logSeg)
5959         {
5960                 PrevLogSeg(_logId, _logSeg);
5961                 RemoveOldXlogFiles(_logId, _logSeg, recptr);
5962         }
5963
5964         /*
5965          * Make more log segments if needed.  (Do this after recycling old log
5966          * segments, since that may supply some of the needed files.)
5967          */
5968         if (!shutdown)
5969                 PreallocXlogFiles(recptr);
5970
5971         /*
5972          * Truncate pg_subtrans if possible.  We can throw away all data before
5973          * the oldest XMIN of any running transaction.  No future transaction will
5974          * attempt to reference any pg_subtrans entry older than that (see Asserts
5975          * in subtrans.c).      During recovery, though, we mustn't do this because
5976          * StartupSUBTRANS hasn't been called yet.
5977          */
5978         if (!InRecovery)
5979                 TruncateSUBTRANS(GetOldestXmin(true, false));
5980
5981         /* All real work is done, but log before releasing lock. */
5982         if (log_checkpoints)
5983                 LogCheckpointEnd();
5984
5985         LWLockRelease(CheckpointLock);
5986 }
5987
5988 /*
5989  * Flush all data in shared memory to disk, and fsync
5990  *
5991  * This is the common code shared between regular checkpoints and
5992  * recovery restartpoints.
5993  */
5994 static void
5995 CheckPointGuts(XLogRecPtr checkPointRedo, int flags)
5996 {
5997         CheckPointCLOG();
5998         CheckPointSUBTRANS();
5999         CheckPointMultiXact();
6000         CheckPointBuffers(flags);       /* performs all required fsyncs */
6001         /* We deliberately delay 2PC checkpointing as long as possible */
6002         CheckPointTwoPhase(checkPointRedo);
6003 }
6004
6005 /*
6006  * Set a recovery restart point if appropriate
6007  *
6008  * This is similar to CreateCheckPoint, but is used during WAL recovery
6009  * to establish a point from which recovery can roll forward without
6010  * replaying the entire recovery log.  This function is called each time
6011  * a checkpoint record is read from XLOG; it must determine whether a
6012  * restartpoint is needed or not.
6013  */
6014 static void
6015 RecoveryRestartPoint(const CheckPoint *checkPoint)
6016 {
6017         int                     elapsed_secs;
6018         int                     rmid;
6019
6020         /*
6021          * Do nothing if the elapsed time since the last restartpoint is less than
6022          * half of checkpoint_timeout.  (We use a value less than
6023          * checkpoint_timeout so that variations in the timing of checkpoints on
6024          * the master, or speed of transmission of WAL segments to a slave, won't
6025          * make the slave skip a restartpoint once it's synced with the master.)
6026          * Checking true elapsed time keeps us from doing restartpoints too often
6027          * while rapidly scanning large amounts of WAL.
6028          */
6029         elapsed_secs = (pg_time_t) time(NULL) - ControlFile->time;
6030         if (elapsed_secs < CheckPointTimeout / 2)
6031                 return;
6032
6033         /*
6034          * Is it safe to checkpoint?  We must ask each of the resource managers
6035          * whether they have any partial state information that might prevent a
6036          * correct restart from this point.  If so, we skip this opportunity, but
6037          * return at the next checkpoint record for another try.
6038          */
6039         for (rmid = 0; rmid <= RM_MAX_ID; rmid++)
6040         {
6041                 if (RmgrTable[rmid].rm_safe_restartpoint != NULL)
6042                         if (!(RmgrTable[rmid].rm_safe_restartpoint()))
6043                         {
6044                                 elog(DEBUG2, "RM %d not safe to record restart point at %X/%X",
6045                                          rmid,
6046                                          checkPoint->redo.xlogid,
6047                                          checkPoint->redo.xrecoff);
6048                                 return;
6049                         }
6050         }
6051
6052         /*
6053          * OK, force data out to disk
6054          */
6055         CheckPointGuts(checkPoint->redo, CHECKPOINT_IMMEDIATE);
6056
6057         /*
6058          * Update pg_control so that any subsequent crash will restart from this
6059          * checkpoint.  Note: ReadRecPtr gives the XLOG address of the checkpoint
6060          * record itself.
6061          */
6062         ControlFile->prevCheckPoint = ControlFile->checkPoint;
6063         ControlFile->checkPoint = ReadRecPtr;
6064         ControlFile->checkPointCopy = *checkPoint;
6065         ControlFile->time = (pg_time_t) time(NULL);
6066         UpdateControlFile();
6067
6068         ereport((recoveryLogRestartpoints ? LOG : DEBUG2),
6069                         (errmsg("recovery restart point at %X/%X",
6070                                         checkPoint->redo.xlogid, checkPoint->redo.xrecoff)));
6071         if (recoveryLastXTime)
6072                 ereport((recoveryLogRestartpoints ? LOG : DEBUG2),
6073                                 (errmsg("last completed transaction was at log time %s",
6074                                                 timestamptz_to_str(recoveryLastXTime))));
6075 }
6076
6077 /*
6078  * Write a NEXTOID log record
6079  */
6080 void
6081 XLogPutNextOid(Oid nextOid)
6082 {
6083         XLogRecData rdata;
6084
6085         rdata.data = (char *) (&nextOid);
6086         rdata.len = sizeof(Oid);
6087         rdata.buffer = InvalidBuffer;
6088         rdata.next = NULL;
6089         (void) XLogInsert(RM_XLOG_ID, XLOG_NEXTOID, &rdata);
6090
6091         /*
6092          * We need not flush the NEXTOID record immediately, because any of the
6093          * just-allocated OIDs could only reach disk as part of a tuple insert or
6094          * update that would have its own XLOG record that must follow the NEXTOID
6095          * record.      Therefore, the standard buffer LSN interlock applied to those
6096          * records will ensure no such OID reaches disk before the NEXTOID record
6097          * does.
6098          *
6099          * Note, however, that the above statement only covers state "within" the
6100          * database.  When we use a generated OID as a file or directory name, we
6101          * are in a sense violating the basic WAL rule, because that filesystem
6102          * change may reach disk before the NEXTOID WAL record does.  The impact
6103          * of this is that if a database crash occurs immediately afterward, we
6104          * might after restart re-generate the same OID and find that it conflicts
6105          * with the leftover file or directory.  But since for safety's sake we
6106          * always loop until finding a nonconflicting filename, this poses no real
6107          * problem in practice. See pgsql-hackers discussion 27-Sep-2006.
6108          */
6109 }
6110
6111 /*
6112  * Write an XLOG SWITCH record.
6113  *
6114  * Here we just blindly issue an XLogInsert request for the record.
6115  * All the magic happens inside XLogInsert.
6116  *
6117  * The return value is either the end+1 address of the switch record,
6118  * or the end+1 address of the prior segment if we did not need to
6119  * write a switch record because we are already at segment start.
6120  */
6121 XLogRecPtr
6122 RequestXLogSwitch(void)
6123 {
6124         XLogRecPtr      RecPtr;
6125         XLogRecData rdata;
6126
6127         /* XLOG SWITCH, alone among xlog record types, has no data */
6128         rdata.buffer = InvalidBuffer;
6129         rdata.data = NULL;
6130         rdata.len = 0;
6131         rdata.next = NULL;
6132
6133         RecPtr = XLogInsert(RM_XLOG_ID, XLOG_SWITCH, &rdata);
6134
6135         return RecPtr;
6136 }
6137
6138 /*
6139  * XLOG resource manager's routines
6140  */
6141 void
6142 xlog_redo(XLogRecPtr lsn, XLogRecord *record)
6143 {
6144         uint8           info = record->xl_info & ~XLR_INFO_MASK;
6145
6146         if (info == XLOG_NEXTOID)
6147         {
6148                 Oid                     nextOid;
6149
6150                 memcpy(&nextOid, XLogRecGetData(record), sizeof(Oid));
6151                 if (ShmemVariableCache->nextOid < nextOid)
6152                 {
6153                         ShmemVariableCache->nextOid = nextOid;
6154                         ShmemVariableCache->oidCount = 0;
6155                 }
6156         }
6157         else if (info == XLOG_CHECKPOINT_SHUTDOWN)
6158         {
6159                 CheckPoint      checkPoint;
6160
6161                 memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
6162                 /* In a SHUTDOWN checkpoint, believe the counters exactly */
6163                 ShmemVariableCache->nextXid = checkPoint.nextXid;
6164                 ShmemVariableCache->nextOid = checkPoint.nextOid;
6165                 ShmemVariableCache->oidCount = 0;
6166                 MultiXactSetNextMXact(checkPoint.nextMulti,
6167                                                           checkPoint.nextMultiOffset);
6168
6169                 /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
6170                 ControlFile->checkPointCopy.nextXidEpoch = checkPoint.nextXidEpoch;
6171                 ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
6172
6173                 /*
6174                  * TLI may change in a shutdown checkpoint, but it shouldn't decrease
6175                  */
6176                 if (checkPoint.ThisTimeLineID != ThisTimeLineID)
6177                 {
6178                         if (checkPoint.ThisTimeLineID < ThisTimeLineID ||
6179                                 !list_member_int(expectedTLIs,
6180                                                                  (int) checkPoint.ThisTimeLineID))
6181                                 ereport(PANIC,
6182                                                 (errmsg("unexpected timeline ID %u (after %u) in checkpoint record",
6183                                                                 checkPoint.ThisTimeLineID, ThisTimeLineID)));
6184                         /* Following WAL records should be run with new TLI */
6185                         ThisTimeLineID = checkPoint.ThisTimeLineID;
6186                 }
6187
6188                 RecoveryRestartPoint(&checkPoint);
6189         }
6190         else if (info == XLOG_CHECKPOINT_ONLINE)
6191         {
6192                 CheckPoint      checkPoint;
6193
6194                 memcpy(&checkPoint, XLogRecGetData(record), sizeof(CheckPoint));
6195                 /* In an ONLINE checkpoint, treat the counters like NEXTOID */
6196                 if (TransactionIdPrecedes(ShmemVariableCache->nextXid,
6197                                                                   checkPoint.nextXid))
6198                         ShmemVariableCache->nextXid = checkPoint.nextXid;
6199                 if (ShmemVariableCache->nextOid < checkPoint.nextOid)
6200                 {
6201                         ShmemVariableCache->nextOid = checkPoint.nextOid;
6202                         ShmemVariableCache->oidCount = 0;
6203                 }
6204                 MultiXactAdvanceNextMXact(checkPoint.nextMulti,
6205                                                                   checkPoint.nextMultiOffset);
6206
6207                 /* ControlFile->checkPointCopy always tracks the latest ckpt XID */
6208                 ControlFile->checkPointCopy.nextXidEpoch = checkPoint.nextXidEpoch;
6209                 ControlFile->checkPointCopy.nextXid = checkPoint.nextXid;
6210
6211                 /* TLI should not change in an on-line checkpoint */
6212                 if (checkPoint.ThisTimeLineID != ThisTimeLineID)
6213                         ereport(PANIC,
6214                                         (errmsg("unexpected timeline ID %u (should be %u) in checkpoint record",
6215                                                         checkPoint.ThisTimeLineID, ThisTimeLineID)));
6216
6217                 RecoveryRestartPoint(&checkPoint);
6218         }
6219         else if (info == XLOG_NOOP)
6220         {
6221                 /* nothing to do here */
6222         }
6223         else if (info == XLOG_SWITCH)
6224         {
6225                 /* nothing to do here */
6226         }
6227 }
6228
6229 void
6230 xlog_desc(StringInfo buf, uint8 xl_info, char *rec)
6231 {
6232         uint8           info = xl_info & ~XLR_INFO_MASK;
6233
6234         if (info == XLOG_CHECKPOINT_SHUTDOWN ||
6235                 info == XLOG_CHECKPOINT_ONLINE)
6236         {
6237                 CheckPoint *checkpoint = (CheckPoint *) rec;
6238
6239                 appendStringInfo(buf, "checkpoint: redo %X/%X; "
6240                                                  "tli %u; xid %u/%u; oid %u; multi %u; offset %u; %s",
6241                                                  checkpoint->redo.xlogid, checkpoint->redo.xrecoff,
6242                                                  checkpoint->ThisTimeLineID,
6243                                                  checkpoint->nextXidEpoch, checkpoint->nextXid,
6244                                                  checkpoint->nextOid,
6245                                                  checkpoint->nextMulti,
6246                                                  checkpoint->nextMultiOffset,
6247                                  (info == XLOG_CHECKPOINT_SHUTDOWN) ? "shutdown" : "online");
6248         }
6249         else if (info == XLOG_NOOP)
6250         {
6251                 appendStringInfo(buf, "xlog no-op");
6252         }
6253         else if (info == XLOG_NEXTOID)
6254         {
6255                 Oid                     nextOid;
6256
6257                 memcpy(&nextOid, rec, sizeof(Oid));
6258                 appendStringInfo(buf, "nextOid: %u", nextOid);
6259         }
6260         else if (info == XLOG_SWITCH)
6261         {
6262                 appendStringInfo(buf, "xlog switch");
6263         }
6264         else
6265                 appendStringInfo(buf, "UNKNOWN");
6266 }
6267
6268 #ifdef WAL_DEBUG
6269
6270 static void
6271 xlog_outrec(StringInfo buf, XLogRecord *record)
6272 {
6273         int                     i;
6274
6275         appendStringInfo(buf, "prev %X/%X; xid %u",
6276                                          record->xl_prev.xlogid, record->xl_prev.xrecoff,
6277                                          record->xl_xid);
6278
6279         for (i = 0; i < XLR_MAX_BKP_BLOCKS; i++)
6280         {
6281                 if (record->xl_info & XLR_SET_BKP_BLOCK(i))
6282                         appendStringInfo(buf, "; bkpb%d", i + 1);
6283         }
6284
6285         appendStringInfo(buf, ": %s", RmgrTable[record->xl_rmid].rm_name);
6286 }
6287 #endif   /* WAL_DEBUG */
6288
6289
6290 /*
6291  * GUC support
6292  */
6293 const char *
6294 assign_xlog_sync_method(const char *method, bool doit, GucSource source)
6295 {
6296         int                     new_sync_method;
6297         int                     new_sync_bit;
6298
6299         if (pg_strcasecmp(method, "fsync") == 0)
6300         {
6301                 new_sync_method = SYNC_METHOD_FSYNC;
6302                 new_sync_bit = 0;
6303         }
6304 #ifdef HAVE_FSYNC_WRITETHROUGH
6305         else if (pg_strcasecmp(method, "fsync_writethrough") == 0)
6306         {
6307                 new_sync_method = SYNC_METHOD_FSYNC_WRITETHROUGH;
6308                 new_sync_bit = 0;
6309         }
6310 #endif
6311 #ifdef HAVE_FDATASYNC
6312         else if (pg_strcasecmp(method, "fdatasync") == 0)
6313         {
6314                 new_sync_method = SYNC_METHOD_FDATASYNC;
6315                 new_sync_bit = 0;
6316         }
6317 #endif
6318 #ifdef OPEN_SYNC_FLAG
6319         else if (pg_strcasecmp(method, "open_sync") == 0)
6320         {
6321                 new_sync_method = SYNC_METHOD_OPEN;
6322                 new_sync_bit = OPEN_SYNC_FLAG;
6323         }
6324 #endif
6325 #ifdef OPEN_DATASYNC_FLAG
6326         else if (pg_strcasecmp(method, "open_datasync") == 0)
6327         {
6328                 new_sync_method = SYNC_METHOD_OPEN;
6329                 new_sync_bit = OPEN_DATASYNC_FLAG;
6330         }
6331 #endif
6332         else
6333                 return NULL;
6334
6335         if (!doit)
6336                 return method;
6337
6338         if (sync_method != new_sync_method || open_sync_bit != new_sync_bit)
6339         {
6340                 /*
6341                  * To ensure that no blocks escape unsynced, force an fsync on the
6342                  * currently open log segment (if any).  Also, if the open flag is
6343                  * changing, close the log file so it will be reopened (with new flag
6344                  * bit) at next use.
6345                  */
6346                 if (openLogFile >= 0)
6347                 {
6348                         if (pg_fsync(openLogFile) != 0)
6349                                 ereport(PANIC,
6350                                                 (errcode_for_file_access(),
6351                                                  errmsg("could not fsync log file %u, segment %u: %m",
6352                                                                 openLogId, openLogSeg)));
6353                         if (open_sync_bit != new_sync_bit)
6354                                 XLogFileClose();
6355                 }
6356                 sync_method = new_sync_method;
6357                 open_sync_bit = new_sync_bit;
6358         }
6359
6360         return method;
6361 }
6362
6363
6364 /*
6365  * Issue appropriate kind of fsync (if any) on the current XLOG output file
6366  */
6367 static void
6368 issue_xlog_fsync(void)
6369 {
6370         switch (sync_method)
6371         {
6372                 case SYNC_METHOD_FSYNC:
6373                         if (pg_fsync_no_writethrough(openLogFile) != 0)
6374                                 ereport(PANIC,
6375                                                 (errcode_for_file_access(),
6376                                                  errmsg("could not fsync log file %u, segment %u: %m",
6377                                                                 openLogId, openLogSeg)));
6378                         break;
6379 #ifdef HAVE_FSYNC_WRITETHROUGH
6380                 case SYNC_METHOD_FSYNC_WRITETHROUGH:
6381                         if (pg_fsync_writethrough(openLogFile) != 0)
6382                                 ereport(PANIC,
6383                                                 (errcode_for_file_access(),
6384                                                  errmsg("could not fsync write-through log file %u, segment %u: %m",
6385                                                                 openLogId, openLogSeg)));
6386                         break;
6387 #endif
6388 #ifdef HAVE_FDATASYNC
6389                 case SYNC_METHOD_FDATASYNC:
6390                         if (pg_fdatasync(openLogFile) != 0)
6391                                 ereport(PANIC,
6392                                                 (errcode_for_file_access(),
6393                                         errmsg("could not fdatasync log file %u, segment %u: %m",
6394                                                    openLogId, openLogSeg)));
6395                         break;
6396 #endif
6397                 case SYNC_METHOD_OPEN:
6398                         /* write synced it already */
6399                         break;
6400                 default:
6401                         elog(PANIC, "unrecognized wal_sync_method: %d", sync_method);
6402                         break;
6403         }
6404 }
6405
6406
6407 /*
6408  * pg_start_backup: set up for taking an on-line backup dump
6409  *
6410  * Essentially what this does is to create a backup label file in $PGDATA,
6411  * where it will be archived as part of the backup dump.  The label file
6412  * contains the user-supplied label string (typically this would be used
6413  * to tell where the backup dump will be stored) and the starting time and
6414  * starting WAL location for the dump.
6415  */
6416 Datum
6417 pg_start_backup(PG_FUNCTION_ARGS)
6418 {
6419         text       *backupid = PG_GETARG_TEXT_P(0);
6420         char       *backupidstr;
6421         XLogRecPtr      checkpointloc;
6422         XLogRecPtr      startpoint;
6423         pg_time_t       stamp_time;
6424         char            strfbuf[128];
6425         char            xlogfilename[MAXFNAMELEN];
6426         uint32          _logId;
6427         uint32          _logSeg;
6428         struct stat stat_buf;
6429         FILE       *fp;
6430
6431         if (!superuser())
6432                 ereport(ERROR,
6433                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
6434                                  errmsg("must be superuser to run a backup")));
6435
6436         if (!XLogArchivingActive())
6437                 ereport(ERROR,
6438                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6439                                  errmsg("WAL archiving is not active"),
6440                                  errhint("archive_mode must be enabled at server start.")));
6441
6442         if (!XLogArchiveCommandSet())
6443                 ereport(ERROR,
6444                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6445                                  errmsg("WAL archiving is not active"),
6446                                  errhint("archive_command must be defined before "
6447                                                  "online backups can be made safely.")));
6448
6449         backupidstr = text_to_cstring(backupid);
6450
6451         /*
6452          * Mark backup active in shared memory.  We must do full-page WAL writes
6453          * during an on-line backup even if not doing so at other times, because
6454          * it's quite possible for the backup dump to obtain a "torn" (partially
6455          * written) copy of a database page if it reads the page concurrently with
6456          * our write to the same page.  This can be fixed as long as the first
6457          * write to the page in the WAL sequence is a full-page write. Hence, we
6458          * turn on forcePageWrites and then force a CHECKPOINT, to ensure there
6459          * are no dirty pages in shared memory that might get dumped while the
6460          * backup is in progress without having a corresponding WAL record.  (Once
6461          * the backup is complete, we need not force full-page writes anymore,
6462          * since we expect that any pages not modified during the backup interval
6463          * must have been correctly captured by the backup.)
6464          *
6465          * We must hold WALInsertLock to change the value of forcePageWrites, to
6466          * ensure adequate interlocking against XLogInsert().
6467          */
6468         LWLockAcquire(WALInsertLock, LW_EXCLUSIVE);
6469         if (XLogCtl->Insert.forcePageWrites)
6470         {
6471                 LWLockRelease(WALInsertLock);
6472                 ereport(ERROR,
6473                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6474                                  errmsg("a backup is already in progress"),
6475                                  errhint("Run pg_stop_backup() and try again.")));
6476         }
6477         XLogCtl->Insert.forcePageWrites = true;
6478         LWLockRelease(WALInsertLock);
6479
6480         /* Ensure we release forcePageWrites if fail below */
6481         PG_ENSURE_ERROR_CLEANUP(pg_start_backup_callback, (Datum) 0);
6482         {
6483                 /*
6484                  * Force a CHECKPOINT.  Aside from being necessary to prevent torn
6485                  * page problems, this guarantees that two successive backup runs will
6486                  * have different checkpoint positions and hence different history
6487                  * file names, even if nothing happened in between.
6488                  *
6489                  * We don't use CHECKPOINT_IMMEDIATE, hence this can take awhile.
6490                  */
6491                 RequestCheckpoint(CHECKPOINT_FORCE | CHECKPOINT_WAIT);
6492
6493                 /*
6494                  * Now we need to fetch the checkpoint record location, and also its
6495                  * REDO pointer.  The oldest point in WAL that would be needed to
6496                  * restore starting from the checkpoint is precisely the REDO pointer.
6497                  */
6498                 LWLockAcquire(ControlFileLock, LW_EXCLUSIVE);
6499                 checkpointloc = ControlFile->checkPoint;
6500                 startpoint = ControlFile->checkPointCopy.redo;
6501                 LWLockRelease(ControlFileLock);
6502
6503                 XLByteToSeg(startpoint, _logId, _logSeg);
6504                 XLogFileName(xlogfilename, ThisTimeLineID, _logId, _logSeg);
6505
6506                 /* Use the log timezone here, not the session timezone */
6507                 stamp_time = (pg_time_t) time(NULL);
6508                 pg_strftime(strfbuf, sizeof(strfbuf),
6509                                         "%Y-%m-%d %H:%M:%S %Z",
6510                                         pg_localtime(&stamp_time, log_timezone));
6511
6512                 /*
6513                  * Check for existing backup label --- implies a backup is already
6514                  * running.  (XXX given that we checked forcePageWrites above, maybe
6515                  * it would be OK to just unlink any such label file?)
6516                  */
6517                 if (stat(BACKUP_LABEL_FILE, &stat_buf) != 0)
6518                 {
6519                         if (errno != ENOENT)
6520                                 ereport(ERROR,
6521                                                 (errcode_for_file_access(),
6522                                                  errmsg("could not stat file \"%s\": %m",
6523                                                                 BACKUP_LABEL_FILE)));
6524                 }
6525                 else
6526                         ereport(ERROR,
6527                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6528                                          errmsg("a backup is already in progress"),
6529                                          errhint("If you're sure there is no backup in progress, remove file \"%s\" and try again.",
6530                                                          BACKUP_LABEL_FILE)));
6531
6532                 /*
6533                  * Okay, write the file
6534                  */
6535                 fp = AllocateFile(BACKUP_LABEL_FILE, "w");
6536                 if (!fp)
6537                         ereport(ERROR,
6538                                         (errcode_for_file_access(),
6539                                          errmsg("could not create file \"%s\": %m",
6540                                                         BACKUP_LABEL_FILE)));
6541                 fprintf(fp, "START WAL LOCATION: %X/%X (file %s)\n",
6542                                 startpoint.xlogid, startpoint.xrecoff, xlogfilename);
6543                 fprintf(fp, "CHECKPOINT LOCATION: %X/%X\n",
6544                                 checkpointloc.xlogid, checkpointloc.xrecoff);
6545                 fprintf(fp, "START TIME: %s\n", strfbuf);
6546                 fprintf(fp, "LABEL: %s\n", backupidstr);
6547                 if (fflush(fp) || ferror(fp) || FreeFile(fp))
6548                         ereport(ERROR,
6549                                         (errcode_for_file_access(),
6550                                          errmsg("could not write file \"%s\": %m",
6551                                                         BACKUP_LABEL_FILE)));
6552         }
6553         PG_END_ENSURE_ERROR_CLEANUP(pg_start_backup_callback, (Datum) 0);
6554
6555         /*
6556          * We're done.  As a convenience, return the starting WAL location.
6557          */
6558         snprintf(xlogfilename, sizeof(xlogfilename), "%X/%X",
6559                          startpoint.xlogid, startpoint.xrecoff);
6560         PG_RETURN_TEXT_P(cstring_to_text(xlogfilename));
6561 }
6562
6563 /* Error cleanup callback for pg_start_backup */
6564 static void
6565 pg_start_backup_callback(int code, Datum arg)
6566 {
6567         /* Turn off forcePageWrites on failure */
6568         LWLockAcquire(WALInsertLock, LW_EXCLUSIVE);
6569         XLogCtl->Insert.forcePageWrites = false;
6570         LWLockRelease(WALInsertLock);
6571 }
6572
6573 /*
6574  * pg_stop_backup: finish taking an on-line backup dump
6575  *
6576  * We remove the backup label file created by pg_start_backup, and instead
6577  * create a backup history file in pg_xlog (whence it will immediately be
6578  * archived).  The backup history file contains the same info found in
6579  * the label file, plus the backup-end time and WAL location.
6580  * Note: different from CancelBackup which just cancels online backup mode.
6581  */
6582 Datum
6583 pg_stop_backup(PG_FUNCTION_ARGS)
6584 {
6585         XLogRecPtr      startpoint;
6586         XLogRecPtr      stoppoint;
6587         pg_time_t       stamp_time;
6588         char            strfbuf[128];
6589         char            histfilepath[MAXPGPATH];
6590         char            startxlogfilename[MAXFNAMELEN];
6591         char            stopxlogfilename[MAXFNAMELEN];
6592         uint32          _logId;
6593         uint32          _logSeg;
6594         FILE       *lfp;
6595         FILE       *fp;
6596         char            ch;
6597         int                     ich;
6598         int                     seconds_before_warning;
6599         int                     waits = 0;
6600
6601         if (!superuser())
6602                 ereport(ERROR,
6603                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
6604                                  (errmsg("must be superuser to run a backup"))));
6605
6606         /*
6607          * OK to clear forcePageWrites
6608          */
6609         LWLockAcquire(WALInsertLock, LW_EXCLUSIVE);
6610         XLogCtl->Insert.forcePageWrites = false;
6611         LWLockRelease(WALInsertLock);
6612
6613         /*
6614          * Force a switch to a new xlog segment file, so that the backup is valid
6615          * as soon as archiver moves out the current segment file. We'll report
6616          * the end address of the XLOG SWITCH record as the backup stopping point.
6617          */
6618         stoppoint = RequestXLogSwitch();
6619
6620         XLByteToSeg(stoppoint, _logId, _logSeg);
6621         XLogFileName(stopxlogfilename, ThisTimeLineID, _logId, _logSeg);
6622
6623         /* Use the log timezone here, not the session timezone */
6624         stamp_time = (pg_time_t) time(NULL);
6625         pg_strftime(strfbuf, sizeof(strfbuf),
6626                                 "%Y-%m-%d %H:%M:%S %Z",
6627                                 pg_localtime(&stamp_time, log_timezone));
6628
6629         /*
6630          * Open the existing label file
6631          */
6632         lfp = AllocateFile(BACKUP_LABEL_FILE, "r");
6633         if (!lfp)
6634         {
6635                 if (errno != ENOENT)
6636                         ereport(ERROR,
6637                                         (errcode_for_file_access(),
6638                                          errmsg("could not read file \"%s\": %m",
6639                                                         BACKUP_LABEL_FILE)));
6640                 ereport(ERROR,
6641                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6642                                  errmsg("a backup is not in progress")));
6643         }
6644
6645         /*
6646          * Read and parse the START WAL LOCATION line (this code is pretty crude,
6647          * but we are not expecting any variability in the file format).
6648          */
6649         if (fscanf(lfp, "START WAL LOCATION: %X/%X (file %24s)%c",
6650                            &startpoint.xlogid, &startpoint.xrecoff, startxlogfilename,
6651                            &ch) != 4 || ch != '\n')
6652                 ereport(ERROR,
6653                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6654                                  errmsg("invalid data in file \"%s\"", BACKUP_LABEL_FILE)));
6655
6656         /*
6657          * Write the backup history file
6658          */
6659         XLByteToSeg(startpoint, _logId, _logSeg);
6660         BackupHistoryFilePath(histfilepath, ThisTimeLineID, _logId, _logSeg,
6661                                                   startpoint.xrecoff % XLogSegSize);
6662         fp = AllocateFile(histfilepath, "w");
6663         if (!fp)
6664                 ereport(ERROR,
6665                                 (errcode_for_file_access(),
6666                                  errmsg("could not create file \"%s\": %m",
6667                                                 histfilepath)));
6668         fprintf(fp, "START WAL LOCATION: %X/%X (file %s)\n",
6669                         startpoint.xlogid, startpoint.xrecoff, startxlogfilename);
6670         fprintf(fp, "STOP WAL LOCATION: %X/%X (file %s)\n",
6671                         stoppoint.xlogid, stoppoint.xrecoff, stopxlogfilename);
6672         /* transfer remaining lines from label to history file */
6673         while ((ich = fgetc(lfp)) != EOF)
6674                 fputc(ich, fp);
6675         fprintf(fp, "STOP TIME: %s\n", strfbuf);
6676         if (fflush(fp) || ferror(fp) || FreeFile(fp))
6677                 ereport(ERROR,
6678                                 (errcode_for_file_access(),
6679                                  errmsg("could not write file \"%s\": %m",
6680                                                 histfilepath)));
6681
6682         /*
6683          * Close and remove the backup label file
6684          */
6685         if (ferror(lfp) || FreeFile(lfp))
6686                 ereport(ERROR,
6687                                 (errcode_for_file_access(),
6688                                  errmsg("could not read file \"%s\": %m",
6689                                                 BACKUP_LABEL_FILE)));
6690         if (unlink(BACKUP_LABEL_FILE) != 0)
6691                 ereport(ERROR,
6692                                 (errcode_for_file_access(),
6693                                  errmsg("could not remove file \"%s\": %m",
6694                                                 BACKUP_LABEL_FILE)));
6695
6696         /*
6697          * Clean out any no-longer-needed history files.  As a side effect, this
6698          * will post a .ready file for the newly created history file, notifying
6699          * the archiver that history file may be archived immediately.
6700          */
6701         CleanupBackupHistory();
6702
6703         /*
6704          * Wait until the history file has been archived. We assume that the
6705          * alphabetic sorting property of the WAL files ensures the last WAL
6706          * file is guaranteed archived by the time the history file is archived.
6707          *
6708          * We wait forever, since archive_command is supposed to work and
6709          * we assume the admin wanted his backup to work completely. If you
6710          * don't wish to wait, you can SET statement_timeout = xx;
6711          *
6712          * If the status file is missing, we assume that is because it was
6713          * set to .ready before we slept, then while asleep it has been set
6714          * to .done and then removed by a concurrent checkpoint.
6715          */
6716         BackupHistoryFileName(histfilepath, ThisTimeLineID, _logId, _logSeg,
6717                                                   startpoint.xrecoff % XLogSegSize);
6718
6719         seconds_before_warning = 60;
6720         waits = 0;
6721
6722         while (!XLogArchiveCheckDone(histfilepath, false))
6723         {
6724                 CHECK_FOR_INTERRUPTS();
6725
6726                 pg_usleep(1000000L);
6727
6728                 if (++waits >= seconds_before_warning)
6729                 {
6730                         seconds_before_warning *= 2;     /* This wraps in >10 years... */
6731                         elog(WARNING, "pg_stop_backup() waiting for archive to complete "
6732                                                         "(%d seconds delay)", waits);
6733                 }
6734         }
6735
6736         /*
6737          * We're done.  As a convenience, return the ending WAL location.
6738          */
6739         snprintf(stopxlogfilename, sizeof(stopxlogfilename), "%X/%X",
6740                          stoppoint.xlogid, stoppoint.xrecoff);
6741         PG_RETURN_TEXT_P(cstring_to_text(stopxlogfilename));
6742 }
6743
6744 /*
6745  * pg_switch_xlog: switch to next xlog file
6746  */
6747 Datum
6748 pg_switch_xlog(PG_FUNCTION_ARGS)
6749 {
6750         XLogRecPtr      switchpoint;
6751         char            location[MAXFNAMELEN];
6752
6753         if (!superuser())
6754                 ereport(ERROR,
6755                                 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
6756                          (errmsg("must be superuser to switch transaction log files"))));
6757
6758         switchpoint = RequestXLogSwitch();
6759
6760         /*
6761          * As a convenience, return the WAL location of the switch record
6762          */
6763         snprintf(location, sizeof(location), "%X/%X",
6764                          switchpoint.xlogid, switchpoint.xrecoff);
6765         PG_RETURN_TEXT_P(cstring_to_text(location));
6766 }
6767
6768 /*
6769  * Report the current WAL write location (same format as pg_start_backup etc)
6770  *
6771  * This is useful for determining how much of WAL is visible to an external
6772  * archiving process.  Note that the data before this point is written out
6773  * to the kernel, but is not necessarily synced to disk.
6774  */
6775 Datum
6776 pg_current_xlog_location(PG_FUNCTION_ARGS)
6777 {
6778         char            location[MAXFNAMELEN];
6779
6780         /* Make sure we have an up-to-date local LogwrtResult */
6781         {
6782                 /* use volatile pointer to prevent code rearrangement */
6783                 volatile XLogCtlData *xlogctl = XLogCtl;
6784
6785                 SpinLockAcquire(&xlogctl->info_lck);
6786                 LogwrtResult = xlogctl->LogwrtResult;
6787                 SpinLockRelease(&xlogctl->info_lck);
6788         }
6789
6790         snprintf(location, sizeof(location), "%X/%X",
6791                          LogwrtResult.Write.xlogid, LogwrtResult.Write.xrecoff);
6792         PG_RETURN_TEXT_P(cstring_to_text(location));
6793 }
6794
6795 /*
6796  * Report the current WAL insert location (same format as pg_start_backup etc)
6797  *
6798  * This function is mostly for debugging purposes.
6799  */
6800 Datum
6801 pg_current_xlog_insert_location(PG_FUNCTION_ARGS)
6802 {
6803         XLogCtlInsert *Insert = &XLogCtl->Insert;
6804         XLogRecPtr      current_recptr;
6805         char            location[MAXFNAMELEN];
6806
6807         /*
6808          * Get the current end-of-WAL position ... shared lock is sufficient
6809          */
6810         LWLockAcquire(WALInsertLock, LW_SHARED);
6811         INSERT_RECPTR(current_recptr, Insert, Insert->curridx);
6812         LWLockRelease(WALInsertLock);
6813
6814         snprintf(location, sizeof(location), "%X/%X",
6815                          current_recptr.xlogid, current_recptr.xrecoff);
6816         PG_RETURN_TEXT_P(cstring_to_text(location));
6817 }
6818
6819 /*
6820  * Compute an xlog file name and decimal byte offset given a WAL location,
6821  * such as is returned by pg_stop_backup() or pg_xlog_switch().
6822  *
6823  * Note that a location exactly at a segment boundary is taken to be in
6824  * the previous segment.  This is usually the right thing, since the
6825  * expected usage is to determine which xlog file(s) are ready to archive.
6826  */
6827 Datum
6828 pg_xlogfile_name_offset(PG_FUNCTION_ARGS)
6829 {
6830         text       *location = PG_GETARG_TEXT_P(0);
6831         char       *locationstr;
6832         unsigned int uxlogid;
6833         unsigned int uxrecoff;
6834         uint32          xlogid;
6835         uint32          xlogseg;
6836         uint32          xrecoff;
6837         XLogRecPtr      locationpoint;
6838         char            xlogfilename[MAXFNAMELEN];
6839         Datum           values[2];
6840         bool            isnull[2];
6841         TupleDesc       resultTupleDesc;
6842         HeapTuple       resultHeapTuple;
6843         Datum           result;
6844
6845         /*
6846          * Read input and parse
6847          */
6848         locationstr = text_to_cstring(location);
6849
6850         if (sscanf(locationstr, "%X/%X", &uxlogid, &uxrecoff) != 2)
6851                 ereport(ERROR,
6852                                 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
6853                                  errmsg("could not parse transaction log location \"%s\"",
6854                                                 locationstr)));
6855
6856         locationpoint.xlogid = uxlogid;
6857         locationpoint.xrecoff = uxrecoff;
6858
6859         /*
6860          * Construct a tuple descriptor for the result row.  This must match this
6861          * function's pg_proc entry!
6862          */
6863         resultTupleDesc = CreateTemplateTupleDesc(2, false);
6864         TupleDescInitEntry(resultTupleDesc, (AttrNumber) 1, "file_name",
6865                                            TEXTOID, -1, 0);
6866         TupleDescInitEntry(resultTupleDesc, (AttrNumber) 2, "file_offset",
6867                                            INT4OID, -1, 0);
6868
6869         resultTupleDesc = BlessTupleDesc(resultTupleDesc);
6870
6871         /*
6872          * xlogfilename
6873          */
6874         XLByteToPrevSeg(locationpoint, xlogid, xlogseg);
6875         XLogFileName(xlogfilename, ThisTimeLineID, xlogid, xlogseg);
6876
6877         values[0] = CStringGetTextDatum(xlogfilename);
6878         isnull[0] = false;
6879
6880         /*
6881          * offset
6882          */
6883         xrecoff = locationpoint.xrecoff - xlogseg * XLogSegSize;
6884
6885         values[1] = UInt32GetDatum(xrecoff);
6886         isnull[1] = false;
6887
6888         /*
6889          * Tuple jam: Having first prepared your Datums, then squash together
6890          */
6891         resultHeapTuple = heap_form_tuple(resultTupleDesc, values, isnull);
6892
6893         result = HeapTupleGetDatum(resultHeapTuple);
6894
6895         PG_RETURN_DATUM(result);
6896 }
6897
6898 /*
6899  * Compute an xlog file name given a WAL location,
6900  * such as is returned by pg_stop_backup() or pg_xlog_switch().
6901  */
6902 Datum
6903 pg_xlogfile_name(PG_FUNCTION_ARGS)
6904 {
6905         text       *location = PG_GETARG_TEXT_P(0);
6906         char       *locationstr;
6907         unsigned int uxlogid;
6908         unsigned int uxrecoff;
6909         uint32          xlogid;
6910         uint32          xlogseg;
6911         XLogRecPtr      locationpoint;
6912         char            xlogfilename[MAXFNAMELEN];
6913
6914         locationstr = text_to_cstring(location);
6915
6916         if (sscanf(locationstr, "%X/%X", &uxlogid, &uxrecoff) != 2)
6917                 ereport(ERROR,
6918                                 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
6919                                  errmsg("could not parse transaction log location \"%s\"",
6920                                                 locationstr)));
6921
6922         locationpoint.xlogid = uxlogid;
6923         locationpoint.xrecoff = uxrecoff;
6924
6925         XLByteToPrevSeg(locationpoint, xlogid, xlogseg);
6926         XLogFileName(xlogfilename, ThisTimeLineID, xlogid, xlogseg);
6927
6928         PG_RETURN_TEXT_P(cstring_to_text(xlogfilename));
6929 }
6930
6931 /*
6932  * read_backup_label: check to see if a backup_label file is present
6933  *
6934  * If we see a backup_label during recovery, we assume that we are recovering
6935  * from a backup dump file, and we therefore roll forward from the checkpoint
6936  * identified by the label file, NOT what pg_control says.      This avoids the
6937  * problem that pg_control might have been archived one or more checkpoints
6938  * later than the start of the dump, and so if we rely on it as the start
6939  * point, we will fail to restore a consistent database state.
6940  *
6941  * We also attempt to retrieve the corresponding backup history file.
6942  * If successful, set *minRecoveryLoc to constrain valid PITR stopping
6943  * points.
6944  *
6945  * Returns TRUE if a backup_label was found (and fills the checkpoint
6946  * location into *checkPointLoc); returns FALSE if not.
6947  */
6948 static bool
6949 read_backup_label(XLogRecPtr *checkPointLoc, XLogRecPtr *minRecoveryLoc)
6950 {
6951         XLogRecPtr      startpoint;
6952         XLogRecPtr      stoppoint;
6953         char            histfilename[MAXFNAMELEN];
6954         char            histfilepath[MAXPGPATH];
6955         char            startxlogfilename[MAXFNAMELEN];
6956         char            stopxlogfilename[MAXFNAMELEN];
6957         TimeLineID      tli;
6958         uint32          _logId;
6959         uint32          _logSeg;
6960         FILE       *lfp;
6961         FILE       *fp;
6962         char            ch;
6963
6964         /* Default is to not constrain recovery stop point */
6965         minRecoveryLoc->xlogid = 0;
6966         minRecoveryLoc->xrecoff = 0;
6967
6968         /*
6969          * See if label file is present
6970          */
6971         lfp = AllocateFile(BACKUP_LABEL_FILE, "r");
6972         if (!lfp)
6973         {
6974                 if (errno != ENOENT)
6975                         ereport(FATAL,
6976                                         (errcode_for_file_access(),
6977                                          errmsg("could not read file \"%s\": %m",
6978                                                         BACKUP_LABEL_FILE)));
6979                 return false;                   /* it's not there, all is fine */
6980         }
6981
6982         /*
6983          * Read and parse the START WAL LOCATION and CHECKPOINT lines (this code
6984          * is pretty crude, but we are not expecting any variability in the file
6985          * format).
6986          */
6987         if (fscanf(lfp, "START WAL LOCATION: %X/%X (file %08X%16s)%c",
6988                            &startpoint.xlogid, &startpoint.xrecoff, &tli,
6989                            startxlogfilename, &ch) != 5 || ch != '\n')
6990                 ereport(FATAL,
6991                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6992                                  errmsg("invalid data in file \"%s\"", BACKUP_LABEL_FILE)));
6993         if (fscanf(lfp, "CHECKPOINT LOCATION: %X/%X%c",
6994                            &checkPointLoc->xlogid, &checkPointLoc->xrecoff,
6995                            &ch) != 3 || ch != '\n')
6996                 ereport(FATAL,
6997                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
6998                                  errmsg("invalid data in file \"%s\"", BACKUP_LABEL_FILE)));
6999         if (ferror(lfp) || FreeFile(lfp))
7000                 ereport(FATAL,
7001                                 (errcode_for_file_access(),
7002                                  errmsg("could not read file \"%s\": %m",
7003                                                 BACKUP_LABEL_FILE)));
7004
7005         /*
7006          * Try to retrieve the backup history file (no error if we can't)
7007          */
7008         XLByteToSeg(startpoint, _logId, _logSeg);
7009         BackupHistoryFileName(histfilename, tli, _logId, _logSeg,
7010                                                   startpoint.xrecoff % XLogSegSize);
7011
7012         if (InArchiveRecovery)
7013                 RestoreArchivedFile(histfilepath, histfilename, "RECOVERYHISTORY", 0);
7014         else
7015                 BackupHistoryFilePath(histfilepath, tli, _logId, _logSeg,
7016                                                           startpoint.xrecoff % XLogSegSize);
7017
7018         fp = AllocateFile(histfilepath, "r");
7019         if (fp)
7020         {
7021                 /*
7022                  * Parse history file to identify stop point.
7023                  */
7024                 if (fscanf(fp, "START WAL LOCATION: %X/%X (file %24s)%c",
7025                                    &startpoint.xlogid, &startpoint.xrecoff, startxlogfilename,
7026                                    &ch) != 4 || ch != '\n')
7027                         ereport(FATAL,
7028                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
7029                                          errmsg("invalid data in file \"%s\"", histfilename)));
7030                 if (fscanf(fp, "STOP WAL LOCATION: %X/%X (file %24s)%c",
7031                                    &stoppoint.xlogid, &stoppoint.xrecoff, stopxlogfilename,
7032                                    &ch) != 4 || ch != '\n')
7033                         ereport(FATAL,
7034                                         (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
7035                                          errmsg("invalid data in file \"%s\"", histfilename)));
7036                 *minRecoveryLoc = stoppoint;
7037                 if (ferror(fp) || FreeFile(fp))
7038                         ereport(FATAL,
7039                                         (errcode_for_file_access(),
7040                                          errmsg("could not read file \"%s\": %m",
7041                                                         histfilepath)));
7042         }
7043
7044         return true;
7045 }
7046
7047 /*
7048  * Error context callback for errors occurring during rm_redo().
7049  */
7050 static void
7051 rm_redo_error_callback(void *arg)
7052 {
7053         XLogRecord *record = (XLogRecord *) arg;
7054         StringInfoData buf;
7055
7056         initStringInfo(&buf);
7057         RmgrTable[record->xl_rmid].rm_desc(&buf,
7058                                                                            record->xl_info,
7059                                                                            XLogRecGetData(record));
7060
7061         /* don't bother emitting empty description */
7062         if (buf.len > 0)
7063                 errcontext("xlog redo %s", buf.data);
7064
7065         pfree(buf.data);
7066 }
7067
7068 /*
7069  * BackupInProgress: check if online backup mode is active
7070  *
7071  * This is done by checking for existence of the "backup_label" file.
7072  */
7073 bool
7074 BackupInProgress(void)
7075 {
7076         struct stat stat_buf;
7077
7078         return (stat(BACKUP_LABEL_FILE, &stat_buf) == 0);
7079 }
7080
7081 /*
7082  * CancelBackup: rename the "backup_label" file to cancel backup mode
7083  *
7084  * If the "backup_label" file exists, it will be renamed to "backup_label.old".
7085  * Note that this will render an online backup in progress useless.
7086  * To correctly finish an online backup, pg_stop_backup must be called.
7087  */
7088 void
7089 CancelBackup(void)
7090 {
7091         struct stat stat_buf;
7092
7093         /* if the file is not there, return */
7094         if (stat(BACKUP_LABEL_FILE, &stat_buf) < 0)
7095                 return;
7096
7097         /* remove leftover file from previously cancelled backup if it exists */
7098         unlink(BACKUP_LABEL_OLD);
7099
7100         if (rename(BACKUP_LABEL_FILE, BACKUP_LABEL_OLD) == 0)
7101         {
7102                 ereport(LOG,
7103                                 (errmsg("online backup mode cancelled"),
7104                                  errdetail("\"%s\" renamed to \"%s\"",
7105                                                 BACKUP_LABEL_FILE, BACKUP_LABEL_OLD)));
7106         }
7107         else
7108         {
7109                 ereport(WARNING,
7110                                 (errcode_for_file_access(),
7111                                  errmsg("could not rename \"%s\" to \"%s\", backup mode not cancelled: %m",
7112                                                 BACKUP_LABEL_FILE, BACKUP_LABEL_OLD)));
7113         }
7114 }
7115