granicus.if.org Git - postgresql/blob - src/backend/access/heap/heapam.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * heapam.c
   4  *        heap access method code
   5  *
   6  * Portions Copyright (c) 1996-2014, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        src/backend/access/heap/heapam.c
  12  *
  13  *
  14  * INTERFACE ROUTINES
  15  *              relation_open   - open any relation by relation OID
  16  *              relation_openrv - open any relation specified by a RangeVar
  17  *              relation_close  - close any relation
  18  *              heap_open               - open a heap relation by relation OID
  19  *              heap_openrv             - open a heap relation specified by a RangeVar
  20  *              heap_close              - (now just a macro for relation_close)
  21  *              heap_beginscan  - begin relation scan
  22  *              heap_rescan             - restart a relation scan
  23  *              heap_endscan    - end relation scan
  24  *              heap_getnext    - retrieve next tuple in scan
  25  *              heap_fetch              - retrieve tuple with given tid
  26  *              heap_insert             - insert tuple into a relation
  27  *              heap_multi_insert - insert multiple tuples into a relation
  28  *              heap_delete             - delete a tuple from a relation
  29  *              heap_update             - replace a tuple in a relation with another tuple
  30  *              heap_markpos    - mark scan position
  31  *              heap_restrpos   - restore position to marked location
  32  *              heap_sync               - sync heap, for when no WAL has been written
  33  *
  34  * NOTES
  35  *        This file contains the heap_ routines which implement
  36  *        the POSTGRES heap access method used for all POSTGRES
  37  *        relations.
  38  *
  39  *-------------------------------------------------------------------------
  40  */
  41 #include "postgres.h"
  42
  43 #include "access/heapam.h"
  44 #include "access/heapam_xlog.h"
  45 #include "access/hio.h"
  46 #include "access/multixact.h"
  47 #include "access/relscan.h"
  48 #include "access/sysattr.h"
  49 #include "access/transam.h"
  50 #include "access/tuptoaster.h"
  51 #include "access/valid.h"
  52 #include "access/visibilitymap.h"
  53 #include "access/xact.h"
  54 #include "access/xlog.h"
  55 #include "access/xloginsert.h"
  56 #include "access/xlogutils.h"
  57 #include "catalog/catalog.h"
  58 #include "catalog/namespace.h"
  59 #include "miscadmin.h"
  60 #include "pgstat.h"
  61 #include "storage/bufmgr.h"
  62 #include "storage/freespace.h"
  63 #include "storage/lmgr.h"
  64 #include "storage/predicate.h"
  65 #include "storage/procarray.h"
  66 #include "storage/smgr.h"
  67 #include "storage/standby.h"
  68 #include "utils/datum.h"
  69 #include "utils/inval.h"
  70 #include "utils/lsyscache.h"
  71 #include "utils/relcache.h"
  72 #include "utils/snapmgr.h"
  73 #include "utils/syscache.h"
  74 #include "utils/tqual.h"
  75
  76
  77 /* GUC variable */
  78 bool            synchronize_seqscans = true;
  79
  80
  81 static HeapScanDesc heap_beginscan_internal(Relation relation,
  82                                                 Snapshot snapshot,
  83                                                 int nkeys, ScanKey key,
  84                                                 bool allow_strat, bool allow_sync,
  85                                                 bool is_bitmapscan, bool temp_snap);
  86 static HeapTuple heap_prepare_insert(Relation relation, HeapTuple tup,
  87                                         TransactionId xid, CommandId cid, int options);
  88 static XLogRecPtr log_heap_update(Relation reln, Buffer oldbuf,
  89                                 Buffer newbuf, HeapTuple oldtup,
  90                                 HeapTuple newtup, HeapTuple old_key_tup,
  91                                 bool all_visible_cleared, bool new_all_visible_cleared);
  92 static void HeapSatisfiesHOTandKeyUpdate(Relation relation,
  93                                                          Bitmapset *hot_attrs,
  94                                                          Bitmapset *key_attrs, Bitmapset *id_attrs,
  95                                                          bool *satisfies_hot, bool *satisfies_key,
  96                                                          bool *satisfies_id,
  97                                                          HeapTuple oldtup, HeapTuple newtup);
  98 static void compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
  99                                                   uint16 old_infomask2, TransactionId add_to_xmax,
 100                                                   LockTupleMode mode, bool is_update,
 101                                                   TransactionId *result_xmax, uint16 *result_infomask,
 102                                                   uint16 *result_infomask2);
 103 static HTSU_Result heap_lock_updated_tuple(Relation rel, HeapTuple tuple,
 104                                                 ItemPointer ctid, TransactionId xid,
 105                                                 LockTupleMode mode);
 106 static void GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
 107                                            uint16 *new_infomask2);
 108 static TransactionId MultiXactIdGetUpdateXid(TransactionId xmax,
 109                                                 uint16 t_infomask);
 110 static void MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
 111                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
 112                                 int *remaining);
 113 static bool ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
 114                                                    uint16 infomask, Relation rel, int *remaining);
 115 static XLogRecPtr log_heap_new_cid(Relation relation, HeapTuple tup);
 116 static HeapTuple ExtractReplicaIdentity(Relation rel, HeapTuple tup, bool key_modified,
 117                                            bool *copy);
 118
 119
 120 /*
 121  * Each tuple lock mode has a corresponding heavyweight lock, and one or two
 122  * corresponding MultiXactStatuses (one to merely lock tuples, another one to
 123  * update them).  This table (and the macros below) helps us determine the
 124  * heavyweight lock mode and MultiXactStatus values to use for any particular
 125  * tuple lock strength.
 126  *
 127  * Don't look at lockstatus/updstatus directly!  Use get_mxact_status_for_lock
 128  * instead.
 129  */
 130 static const struct
 131 {
 132         LOCKMODE        hwlock;
 133         int                     lockstatus;
 134         int                     updstatus;
 135 }
 136
 137                         tupleLockExtraInfo[MaxLockTupleMode + 1] =
 138 {
 139         {                                                       /* LockTupleKeyShare */
 140                 AccessShareLock,
 141                 MultiXactStatusForKeyShare,
 142                 -1                                              /* KeyShare does not allow updating tuples */
 143         },
 144         {                                                       /* LockTupleShare */
 145                 RowShareLock,
 146                 MultiXactStatusForShare,
 147                 -1                                              /* Share does not allow updating tuples */
 148         },
 149         {                                                       /* LockTupleNoKeyExclusive */
 150                 ExclusiveLock,
 151                 MultiXactStatusForNoKeyUpdate,
 152                 MultiXactStatusNoKeyUpdate
 153         },
 154         {                                                       /* LockTupleExclusive */
 155                 AccessExclusiveLock,
 156                 MultiXactStatusForUpdate,
 157                 MultiXactStatusUpdate
 158         }
 159 };
 160
 161 /* Get the LOCKMODE for a given MultiXactStatus */
 162 #define LOCKMODE_from_mxstatus(status) \
 163                         (tupleLockExtraInfo[TUPLOCK_from_mxstatus((status))].hwlock)
 164
 165 /*
 166  * Acquire heavyweight locks on tuples, using a LockTupleMode strength value.
 167  * This is more readable than having every caller translate it to lock.h's
 168  * LOCKMODE.
 169  */
 170 #define LockTupleTuplock(rel, tup, mode) \
 171         LockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 172 #define UnlockTupleTuplock(rel, tup, mode) \
 173         UnlockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 174 #define ConditionalLockTupleTuplock(rel, tup, mode) \
 175         ConditionalLockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 176
 177 /*
 178  * This table maps tuple lock strength values for each particular
 179  * MultiXactStatus value.
 180  */
 181 static const int MultiXactStatusLock[MaxMultiXactStatus + 1] =
 182 {
 183         LockTupleKeyShare,                      /* ForKeyShare */
 184         LockTupleShare,                         /* ForShare */
 185         LockTupleNoKeyExclusive,        /* ForNoKeyUpdate */
 186         LockTupleExclusive,                     /* ForUpdate */
 187         LockTupleNoKeyExclusive,        /* NoKeyUpdate */
 188         LockTupleExclusive                      /* Update */
 189 };
 190
 191 /* Get the LockTupleMode for a given MultiXactStatus */
 192 #define TUPLOCK_from_mxstatus(status) \
 193                         (MultiXactStatusLock[(status)])
 194
 195 /* ----------------------------------------------------------------
 196  *                                               heap support routines
 197  * ----------------------------------------------------------------
 198  */
 199
 200 /* ----------------
 201  *              initscan - scan code common to heap_beginscan and heap_rescan
 202  * ----------------
 203  */
 204 static void
 205 initscan(HeapScanDesc scan, ScanKey key, bool is_rescan)
 206 {
 207         bool            allow_strat;
 208         bool            allow_sync;
 209
 210         /*
 211          * Determine the number of blocks we have to scan.
 212          *
 213          * It is sufficient to do this once at scan start, since any tuples added
 214          * while the scan is in progress will be invisible to my snapshot anyway.
 215          * (That is not true when using a non-MVCC snapshot.  However, we couldn't
 216          * guarantee to return tuples added after scan start anyway, since they
 217          * might go into pages we already scanned.  To guarantee consistent
 218          * results for a non-MVCC snapshot, the caller must hold some higher-level
 219          * lock that ensures the interesting tuple(s) won't change.)
 220          */
 221         scan->rs_nblocks = RelationGetNumberOfBlocks(scan->rs_rd);
 222
 223         /*
 224          * If the table is large relative to NBuffers, use a bulk-read access
 225          * strategy and enable synchronized scanning (see syncscan.c).  Although
 226          * the thresholds for these features could be different, we make them the
 227          * same so that there are only two behaviors to tune rather than four.
 228          * (However, some callers need to be able to disable one or both of these
 229          * behaviors, independently of the size of the table; also there is a GUC
 230          * variable that can disable synchronized scanning.)
 231          *
 232          * During a rescan, don't make a new strategy object if we don't have to.
 233          */
 234         if (!RelationUsesLocalBuffers(scan->rs_rd) &&
 235                 scan->rs_nblocks > NBuffers / 4)
 236         {
 237                 allow_strat = scan->rs_allow_strat;
 238                 allow_sync = scan->rs_allow_sync;
 239         }
 240         else
 241                 allow_strat = allow_sync = false;
 242
 243         if (allow_strat)
 244         {
 245                 if (scan->rs_strategy == NULL)
 246                         scan->rs_strategy = GetAccessStrategy(BAS_BULKREAD);
 247         }
 248         else
 249         {
 250                 if (scan->rs_strategy != NULL)
 251                         FreeAccessStrategy(scan->rs_strategy);
 252                 scan->rs_strategy = NULL;
 253         }
 254
 255         if (is_rescan)
 256         {
 257                 /*
 258                  * If rescan, keep the previous startblock setting so that rewinding a
 259                  * cursor doesn't generate surprising results.  Reset the syncscan
 260                  * setting, though.
 261                  */
 262                 scan->rs_syncscan = (allow_sync && synchronize_seqscans);
 263         }
 264         else if (allow_sync && synchronize_seqscans)
 265         {
 266                 scan->rs_syncscan = true;
 267                 scan->rs_startblock = ss_get_location(scan->rs_rd, scan->rs_nblocks);
 268         }
 269         else
 270         {
 271                 scan->rs_syncscan = false;
 272                 scan->rs_startblock = 0;
 273         }
 274
 275         scan->rs_initblock = 0;
 276         scan->rs_numblocks = InvalidBlockNumber;
 277         scan->rs_inited = false;
 278         scan->rs_ctup.t_data = NULL;
 279         ItemPointerSetInvalid(&scan->rs_ctup.t_self);
 280         scan->rs_cbuf = InvalidBuffer;
 281         scan->rs_cblock = InvalidBlockNumber;
 282
 283         /* we don't have a marked position... */
 284         ItemPointerSetInvalid(&(scan->rs_mctid));
 285
 286         /* page-at-a-time fields are always invalid when not rs_inited */
 287
 288         /*
 289          * copy the scan key, if appropriate
 290          */
 291         if (key != NULL)
 292                 memcpy(scan->rs_key, key, scan->rs_nkeys * sizeof(ScanKeyData));
 293
 294         /*
 295          * Currently, we don't have a stats counter for bitmap heap scans (but the
 296          * underlying bitmap index scans will be counted).
 297          */
 298         if (!scan->rs_bitmapscan)
 299                 pgstat_count_heap_scan(scan->rs_rd);
 300 }
 301
 302 void
 303 heap_setscanlimits(HeapScanDesc scan, BlockNumber startBlk, BlockNumber numBlks)
 304 {
 305         scan->rs_startblock = startBlk;
 306         scan->rs_initblock = startBlk;
 307         scan->rs_numblocks = numBlks;
 308 }
 309
 310 /*
 311  * heapgetpage - subroutine for heapgettup()
 312  *
 313  * This routine reads and pins the specified page of the relation.
 314  * In page-at-a-time mode it performs additional work, namely determining
 315  * which tuples on the page are visible.
 316  */
 317 static void
 318 heapgetpage(HeapScanDesc scan, BlockNumber page)
 319 {
 320         Buffer          buffer;
 321         Snapshot        snapshot;
 322         Page            dp;
 323         int                     lines;
 324         int                     ntup;
 325         OffsetNumber lineoff;
 326         ItemId          lpp;
 327         bool            all_visible;
 328
 329         Assert(page < scan->rs_nblocks);
 330
 331         /* release previous scan buffer, if any */
 332         if (BufferIsValid(scan->rs_cbuf))
 333         {
 334                 ReleaseBuffer(scan->rs_cbuf);
 335                 scan->rs_cbuf = InvalidBuffer;
 336         }
 337
 338         /*
 339          * Be sure to check for interrupts at least once per page.  Checks at
 340          * higher code levels won't be able to stop a seqscan that encounters many
 341          * pages' worth of consecutive dead tuples.
 342          */
 343         CHECK_FOR_INTERRUPTS();
 344
 345         /* read page using selected strategy */
 346         scan->rs_cbuf = ReadBufferExtended(scan->rs_rd, MAIN_FORKNUM, page,
 347                                                                            RBM_NORMAL, scan->rs_strategy);
 348         scan->rs_cblock = page;
 349
 350         if (!scan->rs_pageatatime)
 351                 return;
 352
 353         buffer = scan->rs_cbuf;
 354         snapshot = scan->rs_snapshot;
 355
 356         /*
 357          * Prune and repair fragmentation for the whole page, if possible.
 358          */
 359         heap_page_prune_opt(scan->rs_rd, buffer);
 360
 361         /*
 362          * We must hold share lock on the buffer content while examining tuple
 363          * visibility.  Afterwards, however, the tuples we have found to be
 364          * visible are guaranteed good as long as we hold the buffer pin.
 365          */
 366         LockBuffer(buffer, BUFFER_LOCK_SHARE);
 367
 368         dp = (Page) BufferGetPage(buffer);
 369         lines = PageGetMaxOffsetNumber(dp);
 370         ntup = 0;
 371
 372         /*
 373          * If the all-visible flag indicates that all tuples on the page are
 374          * visible to everyone, we can skip the per-tuple visibility tests.
 375          *
 376          * Note: In hot standby, a tuple that's already visible to all
 377          * transactions in the master might still be invisible to a read-only
 378          * transaction in the standby. We partly handle this problem by tracking
 379          * the minimum xmin of visible tuples as the cut-off XID while marking a
 380          * page all-visible on master and WAL log that along with the visibility
 381          * map SET operation. In hot standby, we wait for (or abort) all
 382          * transactions that can potentially may not see one or more tuples on the
 383          * page. That's how index-only scans work fine in hot standby. A crucial
 384          * difference between index-only scans and heap scans is that the
 385          * index-only scan completely relies on the visibility map where as heap
 386          * scan looks at the page-level PD_ALL_VISIBLE flag. We are not sure if
 387          * the page-level flag can be trusted in the same way, because it might
 388          * get propagated somehow without being explicitly WAL-logged, e.g. via a
 389          * full page write. Until we can prove that beyond doubt, let's check each
 390          * tuple for visibility the hard way.
 391          */
 392         all_visible = PageIsAllVisible(dp) && !snapshot->takenDuringRecovery;
 393
 394         for (lineoff = FirstOffsetNumber, lpp = PageGetItemId(dp, lineoff);
 395                  lineoff <= lines;
 396                  lineoff++, lpp++)
 397         {
 398                 if (ItemIdIsNormal(lpp))
 399                 {
 400                         HeapTupleData loctup;
 401                         bool            valid;
 402
 403                         loctup.t_tableOid = RelationGetRelid(scan->rs_rd);
 404                         loctup.t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 405                         loctup.t_len = ItemIdGetLength(lpp);
 406                         ItemPointerSet(&(loctup.t_self), page, lineoff);
 407
 408                         if (all_visible)
 409                                 valid = true;
 410                         else
 411                                 valid = HeapTupleSatisfiesVisibility(&loctup, snapshot, buffer);
 412
 413                         CheckForSerializableConflictOut(valid, scan->rs_rd, &loctup,
 414                                                                                         buffer, snapshot);
 415
 416                         if (valid)
 417                                 scan->rs_vistuples[ntup++] = lineoff;
 418                 }
 419         }
 420
 421         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
 422
 423         Assert(ntup <= MaxHeapTuplesPerPage);
 424         scan->rs_ntuples = ntup;
 425 }
 426
 427 /* ----------------
 428  *              heapgettup - fetch next heap tuple
 429  *
 430  *              Initialize the scan if not already done; then advance to the next
 431  *              tuple as indicated by "dir"; return the next tuple in scan->rs_ctup,
 432  *              or set scan->rs_ctup.t_data = NULL if no more tuples.
 433  *
 434  * dir == NoMovementScanDirection means "re-fetch the tuple indicated
 435  * by scan->rs_ctup".
 436  *
 437  * Note: the reason nkeys/key are passed separately, even though they are
 438  * kept in the scan descriptor, is that the caller may not want us to check
 439  * the scankeys.
 440  *
 441  * Note: when we fall off the end of the scan in either direction, we
 442  * reset rs_inited.  This means that a further request with the same
 443  * scan direction will restart the scan, which is a bit odd, but a
 444  * request with the opposite scan direction will start a fresh scan
 445  * in the proper direction.  The latter is required behavior for cursors,
 446  * while the former case is generally undefined behavior in Postgres
 447  * so we don't care too much.
 448  * ----------------
 449  */
 450 static void
 451 heapgettup(HeapScanDesc scan,
 452                    ScanDirection dir,
 453                    int nkeys,
 454                    ScanKey key)
 455 {
 456         HeapTuple       tuple = &(scan->rs_ctup);
 457         Snapshot        snapshot = scan->rs_snapshot;
 458         bool            backward = ScanDirectionIsBackward(dir);
 459         BlockNumber page;
 460         bool            finished;
 461         Page            dp;
 462         int                     lines;
 463         OffsetNumber lineoff;
 464         int                     linesleft;
 465         ItemId          lpp;
 466
 467         /*
 468          * calculate next starting lineoff, given scan direction
 469          */
 470         if (ScanDirectionIsForward(dir))
 471         {
 472                 if (!scan->rs_inited)
 473                 {
 474                         /*
 475                          * return null immediately if relation is empty
 476                          */
 477                         if (scan->rs_nblocks == 0)
 478                         {
 479                                 Assert(!BufferIsValid(scan->rs_cbuf));
 480                                 tuple->t_data = NULL;
 481                                 return;
 482                         }
 483                         page = scan->rs_startblock; /* first page */
 484                         heapgetpage(scan, page);
 485                         lineoff = FirstOffsetNumber;            /* first offnum */
 486                         scan->rs_inited = true;
 487                 }
 488                 else
 489                 {
 490                         /* continue from previously returned page/tuple */
 491                         page = scan->rs_cblock;         /* current page */
 492                         lineoff =                       /* next offnum */
 493                                 OffsetNumberNext(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 494                 }
 495
 496                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 497
 498                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 499                 lines = PageGetMaxOffsetNumber(dp);
 500                 /* page and lineoff now reference the physically next tid */
 501
 502                 linesleft = lines - lineoff + 1;
 503         }
 504         else if (backward)
 505         {
 506                 if (!scan->rs_inited)
 507                 {
 508                         /*
 509                          * return null immediately if relation is empty
 510                          */
 511                         if (scan->rs_nblocks == 0)
 512                         {
 513                                 Assert(!BufferIsValid(scan->rs_cbuf));
 514                                 tuple->t_data = NULL;
 515                                 return;
 516                         }
 517
 518                         /*
 519                          * Disable reporting to syncscan logic in a backwards scan; it's
 520                          * not very likely anyone else is doing the same thing at the same
 521                          * time, and much more likely that we'll just bollix things for
 522                          * forward scanners.
 523                          */
 524                         scan->rs_syncscan = false;
 525                         /* start from last page of the scan */
 526                         if (scan->rs_startblock > 0)
 527                                 page = scan->rs_startblock - 1;
 528                         else
 529                                 page = scan->rs_nblocks - 1;
 530                         heapgetpage(scan, page);
 531                 }
 532                 else
 533                 {
 534                         /* continue from previously returned page/tuple */
 535                         page = scan->rs_cblock;         /* current page */
 536                 }
 537
 538                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 539
 540                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 541                 lines = PageGetMaxOffsetNumber(dp);
 542
 543                 if (!scan->rs_inited)
 544                 {
 545                         lineoff = lines;        /* final offnum */
 546                         scan->rs_inited = true;
 547                 }
 548                 else
 549                 {
 550                         lineoff =                       /* previous offnum */
 551                                 OffsetNumberPrev(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 552                 }
 553                 /* page and lineoff now reference the physically previous tid */
 554
 555                 linesleft = lineoff;
 556         }
 557         else
 558         {
 559                 /*
 560                  * ``no movement'' scan direction: refetch prior tuple
 561                  */
 562                 if (!scan->rs_inited)
 563                 {
 564                         Assert(!BufferIsValid(scan->rs_cbuf));
 565                         tuple->t_data = NULL;
 566                         return;
 567                 }
 568
 569                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 570                 if (page != scan->rs_cblock)
 571                         heapgetpage(scan, page);
 572
 573                 /* Since the tuple was previously fetched, needn't lock page here */
 574                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 575                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 576                 lpp = PageGetItemId(dp, lineoff);
 577                 Assert(ItemIdIsNormal(lpp));
 578
 579                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 580                 tuple->t_len = ItemIdGetLength(lpp);
 581
 582                 return;
 583         }
 584
 585         /*
 586          * advance the scan until we find a qualifying tuple or run out of stuff
 587          * to scan
 588          */
 589         lpp = PageGetItemId(dp, lineoff);
 590         for (;;)
 591         {
 592                 while (linesleft > 0)
 593                 {
 594                         if (ItemIdIsNormal(lpp))
 595                         {
 596                                 bool            valid;
 597
 598                                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 599                                 tuple->t_len = ItemIdGetLength(lpp);
 600                                 ItemPointerSet(&(tuple->t_self), page, lineoff);
 601
 602                                 /*
 603                                  * if current tuple qualifies, return it.
 604                                  */
 605                                 valid = HeapTupleSatisfiesVisibility(tuple,
 606                                                                                                          snapshot,
 607                                                                                                          scan->rs_cbuf);
 608
 609                                 CheckForSerializableConflictOut(valid, scan->rs_rd, tuple,
 610                                                                                                 scan->rs_cbuf, snapshot);
 611
 612                                 if (valid && key != NULL)
 613                                         HeapKeyTest(tuple, RelationGetDescr(scan->rs_rd),
 614                                                                 nkeys, key, valid);
 615
 616                                 if (valid)
 617                                 {
 618                                         LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 619                                         return;
 620                                 }
 621                         }
 622
 623                         /*
 624                          * otherwise move to the next item on the page
 625                          */
 626                         --linesleft;
 627                         if (backward)
 628                         {
 629                                 --lpp;                  /* move back in this page's ItemId array */
 630                                 --lineoff;
 631                         }
 632                         else
 633                         {
 634                                 ++lpp;                  /* move forward in this page's ItemId array */
 635                                 ++lineoff;
 636                         }
 637                 }
 638
 639                 /*
 640                  * if we get here, it means we've exhausted the items on this page and
 641                  * it's time to move to the next.
 642                  */
 643                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 644
 645                 /*
 646                  * advance to next/prior page and detect end of scan
 647                  */
 648                 if (backward)
 649                 {
 650                         finished = (page == scan->rs_startblock) ||
 651                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks <= 0 : false);
 652                         if (page == 0)
 653                                 page = scan->rs_nblocks;
 654                         page--;
 655                 }
 656                 else
 657                 {
 658                         page++;
 659                         if (page >= scan->rs_nblocks)
 660                                 page = 0;
 661                         finished = (page == scan->rs_startblock) ||
 662                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks <= 0 : false);
 663
 664                         /*
 665                          * Report our new scan position for synchronization purposes. We
 666                          * don't do that when moving backwards, however. That would just
 667                          * mess up any other forward-moving scanners.
 668                          *
 669                          * Note: we do this before checking for end of scan so that the
 670                          * final state of the position hint is back at the start of the
 671                          * rel.  That's not strictly necessary, but otherwise when you run
 672                          * the same query multiple times the starting position would shift
 673                          * a little bit backwards on every invocation, which is confusing.
 674                          * We don't guarantee any specific ordering in general, though.
 675                          */
 676                         if (scan->rs_syncscan)
 677                                 ss_report_location(scan->rs_rd, page);
 678                 }
 679
 680                 /*
 681                  * return NULL if we've exhausted all the pages
 682                  */
 683                 if (finished)
 684                 {
 685                         if (BufferIsValid(scan->rs_cbuf))
 686                                 ReleaseBuffer(scan->rs_cbuf);
 687                         scan->rs_cbuf = InvalidBuffer;
 688                         scan->rs_cblock = InvalidBlockNumber;
 689                         tuple->t_data = NULL;
 690                         scan->rs_inited = false;
 691                         return;
 692                 }
 693
 694                 heapgetpage(scan, page);
 695
 696                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 697
 698                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 699                 lines = PageGetMaxOffsetNumber((Page) dp);
 700                 linesleft = lines;
 701                 if (backward)
 702                 {
 703                         lineoff = lines;
 704                         lpp = PageGetItemId(dp, lines);
 705                 }
 706                 else
 707                 {
 708                         lineoff = FirstOffsetNumber;
 709                         lpp = PageGetItemId(dp, FirstOffsetNumber);
 710                 }
 711         }
 712 }
 713
 714 /* ----------------
 715  *              heapgettup_pagemode - fetch next heap tuple in page-at-a-time mode
 716  *
 717  *              Same API as heapgettup, but used in page-at-a-time mode
 718  *
 719  * The internal logic is much the same as heapgettup's too, but there are some
 720  * differences: we do not take the buffer content lock (that only needs to
 721  * happen inside heapgetpage), and we iterate through just the tuples listed
 722  * in rs_vistuples[] rather than all tuples on the page.  Notice that
 723  * lineindex is 0-based, where the corresponding loop variable lineoff in
 724  * heapgettup is 1-based.
 725  * ----------------
 726  */
 727 static void
 728 heapgettup_pagemode(HeapScanDesc scan,
 729                                         ScanDirection dir,
 730                                         int nkeys,
 731                                         ScanKey key)
 732 {
 733         HeapTuple       tuple = &(scan->rs_ctup);
 734         bool            backward = ScanDirectionIsBackward(dir);
 735         BlockNumber page;
 736         bool            finished;
 737         Page            dp;
 738         int                     lines;
 739         int                     lineindex;
 740         OffsetNumber lineoff;
 741         int                     linesleft;
 742         ItemId          lpp;
 743
 744         /*
 745          * calculate next starting lineindex, given scan direction
 746          */
 747         if (ScanDirectionIsForward(dir))
 748         {
 749                 if (!scan->rs_inited)
 750                 {
 751                         /*
 752                          * return null immediately if relation is empty
 753                          */
 754                         if (scan->rs_nblocks == 0)
 755                         {
 756                                 Assert(!BufferIsValid(scan->rs_cbuf));
 757                                 tuple->t_data = NULL;
 758                                 return;
 759                         }
 760                         page = scan->rs_startblock; /* first page */
 761                         heapgetpage(scan, page);
 762                         lineindex = 0;
 763                         scan->rs_inited = true;
 764                 }
 765                 else
 766                 {
 767                         /* continue from previously returned page/tuple */
 768                         page = scan->rs_cblock;         /* current page */
 769                         lineindex = scan->rs_cindex + 1;
 770                 }
 771
 772                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 773                 lines = scan->rs_ntuples;
 774                 /* page and lineindex now reference the next visible tid */
 775
 776                 linesleft = lines - lineindex;
 777         }
 778         else if (backward)
 779         {
 780                 if (!scan->rs_inited)
 781                 {
 782                         /*
 783                          * return null immediately if relation is empty
 784                          */
 785                         if (scan->rs_nblocks == 0)
 786                         {
 787                                 Assert(!BufferIsValid(scan->rs_cbuf));
 788                                 tuple->t_data = NULL;
 789                                 return;
 790                         }
 791
 792                         /*
 793                          * Disable reporting to syncscan logic in a backwards scan; it's
 794                          * not very likely anyone else is doing the same thing at the same
 795                          * time, and much more likely that we'll just bollix things for
 796                          * forward scanners.
 797                          */
 798                         scan->rs_syncscan = false;
 799                         /* start from last page of the scan */
 800                         if (scan->rs_startblock > 0)
 801                                 page = scan->rs_startblock - 1;
 802                         else
 803                                 page = scan->rs_nblocks - 1;
 804                         heapgetpage(scan, page);
 805                 }
 806                 else
 807                 {
 808                         /* continue from previously returned page/tuple */
 809                         page = scan->rs_cblock;         /* current page */
 810                 }
 811
 812                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 813                 lines = scan->rs_ntuples;
 814
 815                 if (!scan->rs_inited)
 816                 {
 817                         lineindex = lines - 1;
 818                         scan->rs_inited = true;
 819                 }
 820                 else
 821                 {
 822                         lineindex = scan->rs_cindex - 1;
 823                 }
 824                 /* page and lineindex now reference the previous visible tid */
 825
 826                 linesleft = lineindex + 1;
 827         }
 828         else
 829         {
 830                 /*
 831                  * ``no movement'' scan direction: refetch prior tuple
 832                  */
 833                 if (!scan->rs_inited)
 834                 {
 835                         Assert(!BufferIsValid(scan->rs_cbuf));
 836                         tuple->t_data = NULL;
 837                         return;
 838                 }
 839
 840                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 841                 if (page != scan->rs_cblock)
 842                         heapgetpage(scan, page);
 843
 844                 /* Since the tuple was previously fetched, needn't lock page here */
 845                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 846                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 847                 lpp = PageGetItemId(dp, lineoff);
 848                 Assert(ItemIdIsNormal(lpp));
 849
 850                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 851                 tuple->t_len = ItemIdGetLength(lpp);
 852
 853                 /* check that rs_cindex is in sync */
 854                 Assert(scan->rs_cindex < scan->rs_ntuples);
 855                 Assert(lineoff == scan->rs_vistuples[scan->rs_cindex]);
 856
 857                 return;
 858         }
 859
 860         /*
 861          * advance the scan until we find a qualifying tuple or run out of stuff
 862          * to scan
 863          */
 864         for (;;)
 865         {
 866                 while (linesleft > 0)
 867                 {
 868                         lineoff = scan->rs_vistuples[lineindex];
 869                         lpp = PageGetItemId(dp, lineoff);
 870                         Assert(ItemIdIsNormal(lpp));
 871
 872                         tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 873                         tuple->t_len = ItemIdGetLength(lpp);
 874                         ItemPointerSet(&(tuple->t_self), page, lineoff);
 875
 876                         /*
 877                          * if current tuple qualifies, return it.
 878                          */
 879                         if (key != NULL)
 880                         {
 881                                 bool            valid;
 882
 883                                 HeapKeyTest(tuple, RelationGetDescr(scan->rs_rd),
 884                                                         nkeys, key, valid);
 885                                 if (valid)
 886                                 {
 887                                         scan->rs_cindex = lineindex;
 888                                         return;
 889                                 }
 890                         }
 891                         else
 892                         {
 893                                 scan->rs_cindex = lineindex;
 894                                 return;
 895                         }
 896
 897                         /*
 898                          * otherwise move to the next item on the page
 899                          */
 900                         --linesleft;
 901                         if (backward)
 902                                 --lineindex;
 903                         else
 904                                 ++lineindex;
 905                 }
 906
 907                 /*
 908                  * if we get here, it means we've exhausted the items on this page and
 909                  * it's time to move to the next.
 910                  */
 911                 if (backward)
 912                 {
 913                         finished = (page == scan->rs_startblock) ||
 914                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks <= 0 : false);
 915                         if (page == 0)
 916                                 page = scan->rs_nblocks;
 917                         page--;
 918                 }
 919                 else
 920                 {
 921                         page++;
 922                         if (page >= scan->rs_nblocks)
 923                                 page = 0;
 924                         finished = (page == scan->rs_startblock) ||
 925                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks <= 0 : false);
 926
 927                         /*
 928                          * Report our new scan position for synchronization purposes. We
 929                          * don't do that when moving backwards, however. That would just
 930                          * mess up any other forward-moving scanners.
 931                          *
 932                          * Note: we do this before checking for end of scan so that the
 933                          * final state of the position hint is back at the start of the
 934                          * rel.  That's not strictly necessary, but otherwise when you run
 935                          * the same query multiple times the starting position would shift
 936                          * a little bit backwards on every invocation, which is confusing.
 937                          * We don't guarantee any specific ordering in general, though.
 938                          */
 939                         if (scan->rs_syncscan)
 940                                 ss_report_location(scan->rs_rd, page);
 941                 }
 942
 943                 /*
 944                  * return NULL if we've exhausted all the pages
 945                  */
 946                 if (finished)
 947                 {
 948                         if (BufferIsValid(scan->rs_cbuf))
 949                                 ReleaseBuffer(scan->rs_cbuf);
 950                         scan->rs_cbuf = InvalidBuffer;
 951                         scan->rs_cblock = InvalidBlockNumber;
 952                         tuple->t_data = NULL;
 953                         scan->rs_inited = false;
 954                         return;
 955                 }
 956
 957                 heapgetpage(scan, page);
 958
 959                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 960                 lines = scan->rs_ntuples;
 961                 linesleft = lines;
 962                 if (backward)
 963                         lineindex = lines - 1;
 964                 else
 965                         lineindex = 0;
 966         }
 967 }
 968
 969
 970 #if defined(DISABLE_COMPLEX_MACRO)
 971 /*
 972  * This is formatted so oddly so that the correspondence to the macro
 973  * definition in access/htup_details.h is maintained.
 974  */
 975 Datum
 976 fastgetattr(HeapTuple tup, int attnum, TupleDesc tupleDesc,
 977                         bool *isnull)
 978 {
 979         return (
 980                         (attnum) > 0 ?
 981                         (
 982                          (*(isnull) = false),
 983                          HeapTupleNoNulls(tup) ?
 984                          (
 985                           (tupleDesc)->attrs[(attnum) - 1]->attcacheoff >= 0 ?
 986                           (
 987                            fetchatt((tupleDesc)->attrs[(attnum) - 1],
 988                                                 (char *) (tup)->t_data + (tup)->t_data->t_hoff +
 989                                                 (tupleDesc)->attrs[(attnum) - 1]->attcacheoff)
 990                            )
 991                           :
 992                           nocachegetattr((tup), (attnum), (tupleDesc))
 993                           )
 994                          :
 995                          (
 996                           att_isnull((attnum) - 1, (tup)->t_data->t_bits) ?
 997                           (
 998                            (*(isnull) = true),
 999                            (Datum) NULL
1000                            )
1001                           :
1002                           (
1003                            nocachegetattr((tup), (attnum), (tupleDesc))
1004                            )
1005                           )
1006                          )
1007                         :
1008                         (
1009                          (Datum) NULL
1010                          )
1011                 );
1012 }
1013 #endif   /* defined(DISABLE_COMPLEX_MACRO) */
1014
1015
1016 /* ----------------------------------------------------------------
1017  *                                       heap access method interface
1018  * ----------------------------------------------------------------
1019  */
1020
1021 /* ----------------
1022  *              relation_open - open any relation by relation OID
1023  *
1024  *              If lockmode is not "NoLock", the specified kind of lock is
1025  *              obtained on the relation.  (Generally, NoLock should only be
1026  *              used if the caller knows it has some appropriate lock on the
1027  *              relation already.)
1028  *
1029  *              An error is raised if the relation does not exist.
1030  *
1031  *              NB: a "relation" is anything with a pg_class entry.  The caller is
1032  *              expected to check whether the relkind is something it can handle.
1033  * ----------------
1034  */
1035 Relation
1036 relation_open(Oid relationId, LOCKMODE lockmode)
1037 {
1038         Relation        r;
1039
1040         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1041
1042         /* Get the lock before trying to open the relcache entry */
1043         if (lockmode != NoLock)
1044                 LockRelationOid(relationId, lockmode);
1045
1046         /* The relcache does all the real work... */
1047         r = RelationIdGetRelation(relationId);
1048
1049         if (!RelationIsValid(r))
1050                 elog(ERROR, "could not open relation with OID %u", relationId);
1051
1052         /* Make note that we've accessed a temporary relation */
1053         if (RelationUsesLocalBuffers(r))
1054                 MyXactAccessedTempRel = true;
1055
1056         pgstat_initstats(r);
1057
1058         return r;
1059 }
1060
1061 /* ----------------
1062  *              try_relation_open - open any relation by relation OID
1063  *
1064  *              Same as relation_open, except return NULL instead of failing
1065  *              if the relation does not exist.
1066  * ----------------
1067  */
1068 Relation
1069 try_relation_open(Oid relationId, LOCKMODE lockmode)
1070 {
1071         Relation        r;
1072
1073         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1074
1075         /* Get the lock first */
1076         if (lockmode != NoLock)
1077                 LockRelationOid(relationId, lockmode);
1078
1079         /*
1080          * Now that we have the lock, probe to see if the relation really exists
1081          * or not.
1082          */
1083         if (!SearchSysCacheExists1(RELOID, ObjectIdGetDatum(relationId)))
1084         {
1085                 /* Release useless lock */
1086                 if (lockmode != NoLock)
1087                         UnlockRelationOid(relationId, lockmode);
1088
1089                 return NULL;
1090         }
1091
1092         /* Should be safe to do a relcache load */
1093         r = RelationIdGetRelation(relationId);
1094
1095         if (!RelationIsValid(r))
1096                 elog(ERROR, "could not open relation with OID %u", relationId);
1097
1098         /* Make note that we've accessed a temporary relation */
1099         if (RelationUsesLocalBuffers(r))
1100                 MyXactAccessedTempRel = true;
1101
1102         pgstat_initstats(r);
1103
1104         return r;
1105 }
1106
1107 /* ----------------
1108  *              relation_openrv - open any relation specified by a RangeVar
1109  *
1110  *              Same as relation_open, but the relation is specified by a RangeVar.
1111  * ----------------
1112  */
1113 Relation
1114 relation_openrv(const RangeVar *relation, LOCKMODE lockmode)
1115 {
1116         Oid                     relOid;
1117
1118         /*
1119          * Check for shared-cache-inval messages before trying to open the
1120          * relation.  This is needed even if we already hold a lock on the
1121          * relation, because GRANT/REVOKE are executed without taking any lock on
1122          * the target relation, and we want to be sure we see current ACL
1123          * information.  We can skip this if asked for NoLock, on the assumption
1124          * that such a call is not the first one in the current command, and so we
1125          * should be reasonably up-to-date already.  (XXX this all could stand to
1126          * be redesigned, but for the moment we'll keep doing this like it's been
1127          * done historically.)
1128          */
1129         if (lockmode != NoLock)
1130                 AcceptInvalidationMessages();
1131
1132         /* Look up and lock the appropriate relation using namespace search */
1133         relOid = RangeVarGetRelid(relation, lockmode, false);
1134
1135         /* Let relation_open do the rest */
1136         return relation_open(relOid, NoLock);
1137 }
1138
1139 /* ----------------
1140  *              relation_openrv_extended - open any relation specified by a RangeVar
1141  *
1142  *              Same as relation_openrv, but with an additional missing_ok argument
1143  *              allowing a NULL return rather than an error if the relation is not
1144  *              found.  (Note that some other causes, such as permissions problems,
1145  *              will still result in an ereport.)
1146  * ----------------
1147  */
1148 Relation
1149 relation_openrv_extended(const RangeVar *relation, LOCKMODE lockmode,
1150                                                  bool missing_ok)
1151 {
1152         Oid                     relOid;
1153
1154         /*
1155          * Check for shared-cache-inval messages before trying to open the
1156          * relation.  See comments in relation_openrv().
1157          */
1158         if (lockmode != NoLock)
1159                 AcceptInvalidationMessages();
1160
1161         /* Look up and lock the appropriate relation using namespace search */
1162         relOid = RangeVarGetRelid(relation, lockmode, missing_ok);
1163
1164         /* Return NULL on not-found */
1165         if (!OidIsValid(relOid))
1166                 return NULL;
1167
1168         /* Let relation_open do the rest */
1169         return relation_open(relOid, NoLock);
1170 }
1171
1172 /* ----------------
1173  *              relation_close - close any relation
1174  *
1175  *              If lockmode is not "NoLock", we then release the specified lock.
1176  *
1177  *              Note that it is often sensible to hold a lock beyond relation_close;
1178  *              in that case, the lock is released automatically at xact end.
1179  * ----------------
1180  */
1181 void
1182 relation_close(Relation relation, LOCKMODE lockmode)
1183 {
1184         LockRelId       relid = relation->rd_lockInfo.lockRelId;
1185
1186         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1187
1188         /* The relcache does the real work... */
1189         RelationClose(relation);
1190
1191         if (lockmode != NoLock)
1192                 UnlockRelationId(&relid, lockmode);
1193 }
1194
1195
1196 /* ----------------
1197  *              heap_open - open a heap relation by relation OID
1198  *
1199  *              This is essentially relation_open plus check that the relation
1200  *              is not an index nor a composite type.  (The caller should also
1201  *              check that it's not a view or foreign table before assuming it has
1202  *              storage.)
1203  * ----------------
1204  */
1205 Relation
1206 heap_open(Oid relationId, LOCKMODE lockmode)
1207 {
1208         Relation        r;
1209
1210         r = relation_open(relationId, lockmode);
1211
1212         if (r->rd_rel->relkind == RELKIND_INDEX)
1213                 ereport(ERROR,
1214                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1215                                  errmsg("\"%s\" is an index",
1216                                                 RelationGetRelationName(r))));
1217         else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1218                 ereport(ERROR,
1219                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1220                                  errmsg("\"%s\" is a composite type",
1221                                                 RelationGetRelationName(r))));
1222
1223         return r;
1224 }
1225
1226 /* ----------------
1227  *              heap_openrv - open a heap relation specified
1228  *              by a RangeVar node
1229  *
1230  *              As above, but relation is specified by a RangeVar.
1231  * ----------------
1232  */
1233 Relation
1234 heap_openrv(const RangeVar *relation, LOCKMODE lockmode)
1235 {
1236         Relation        r;
1237
1238         r = relation_openrv(relation, lockmode);
1239
1240         if (r->rd_rel->relkind == RELKIND_INDEX)
1241                 ereport(ERROR,
1242                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1243                                  errmsg("\"%s\" is an index",
1244                                                 RelationGetRelationName(r))));
1245         else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1246                 ereport(ERROR,
1247                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1248                                  errmsg("\"%s\" is a composite type",
1249                                                 RelationGetRelationName(r))));
1250
1251         return r;
1252 }
1253
1254 /* ----------------
1255  *              heap_openrv_extended - open a heap relation specified
1256  *              by a RangeVar node
1257  *
1258  *              As above, but optionally return NULL instead of failing for
1259  *              relation-not-found.
1260  * ----------------
1261  */
1262 Relation
1263 heap_openrv_extended(const RangeVar *relation, LOCKMODE lockmode,
1264                                          bool missing_ok)
1265 {
1266         Relation        r;
1267
1268         r = relation_openrv_extended(relation, lockmode, missing_ok);
1269
1270         if (r)
1271         {
1272                 if (r->rd_rel->relkind == RELKIND_INDEX)
1273                         ereport(ERROR,
1274                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1275                                          errmsg("\"%s\" is an index",
1276                                                         RelationGetRelationName(r))));
1277                 else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1278                         ereport(ERROR,
1279                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1280                                          errmsg("\"%s\" is a composite type",
1281                                                         RelationGetRelationName(r))));
1282         }
1283
1284         return r;
1285 }
1286
1287
1288 /* ----------------
1289  *              heap_beginscan  - begin relation scan
1290  *
1291  * heap_beginscan_strat offers an extended API that lets the caller control
1292  * whether a nondefault buffer access strategy can be used, and whether
1293  * syncscan can be chosen (possibly resulting in the scan not starting from
1294  * block zero).  Both of these default to TRUE with plain heap_beginscan.
1295  *
1296  * heap_beginscan_bm is an alternative entry point for setting up a
1297  * HeapScanDesc for a bitmap heap scan.  Although that scan technology is
1298  * really quite unlike a standard seqscan, there is just enough commonality
1299  * to make it worth using the same data structure.
1300  * ----------------
1301  */
1302 HeapScanDesc
1303 heap_beginscan(Relation relation, Snapshot snapshot,
1304                            int nkeys, ScanKey key)
1305 {
1306         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1307                                                                    true, true, false, false);
1308 }
1309
1310 HeapScanDesc
1311 heap_beginscan_catalog(Relation relation, int nkeys, ScanKey key)
1312 {
1313         Oid                     relid = RelationGetRelid(relation);
1314         Snapshot        snapshot = RegisterSnapshot(GetCatalogSnapshot(relid));
1315
1316         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1317                                                                    true, true, false, true);
1318 }
1319
1320 HeapScanDesc
1321 heap_beginscan_strat(Relation relation, Snapshot snapshot,
1322                                          int nkeys, ScanKey key,
1323                                          bool allow_strat, bool allow_sync)
1324 {
1325         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1326                                                                    allow_strat, allow_sync, false, false);
1327 }
1328
1329 HeapScanDesc
1330 heap_beginscan_bm(Relation relation, Snapshot snapshot,
1331                                   int nkeys, ScanKey key)
1332 {
1333         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1334                                                                    false, false, true, false);
1335 }
1336
1337 static HeapScanDesc
1338 heap_beginscan_internal(Relation relation, Snapshot snapshot,
1339                                                 int nkeys, ScanKey key,
1340                                                 bool allow_strat, bool allow_sync,
1341                                                 bool is_bitmapscan, bool temp_snap)
1342 {
1343         HeapScanDesc scan;
1344
1345         /*
1346          * increment relation ref count while scanning relation
1347          *
1348          * This is just to make really sure the relcache entry won't go away while
1349          * the scan has a pointer to it.  Caller should be holding the rel open
1350          * anyway, so this is redundant in all normal scenarios...
1351          */
1352         RelationIncrementReferenceCount(relation);
1353
1354         /*
1355          * allocate and initialize scan descriptor
1356          */
1357         scan = (HeapScanDesc) palloc(sizeof(HeapScanDescData));
1358
1359         scan->rs_rd = relation;
1360         scan->rs_snapshot = snapshot;
1361         scan->rs_nkeys = nkeys;
1362         scan->rs_bitmapscan = is_bitmapscan;
1363         scan->rs_strategy = NULL;       /* set in initscan */
1364         scan->rs_allow_strat = allow_strat;
1365         scan->rs_allow_sync = allow_sync;
1366         scan->rs_temp_snap = temp_snap;
1367
1368         /*
1369          * we can use page-at-a-time mode if it's an MVCC-safe snapshot
1370          */
1371         scan->rs_pageatatime = IsMVCCSnapshot(snapshot);
1372
1373         /*
1374          * For a seqscan in a serializable transaction, acquire a predicate lock
1375          * on the entire relation. This is required not only to lock all the
1376          * matching tuples, but also to conflict with new insertions into the
1377          * table. In an indexscan, we take page locks on the index pages covering
1378          * the range specified in the scan qual, but in a heap scan there is
1379          * nothing more fine-grained to lock. A bitmap scan is a different story,
1380          * there we have already scanned the index and locked the index pages
1381          * covering the predicate. But in that case we still have to lock any
1382          * matching heap tuples.
1383          */
1384         if (!is_bitmapscan)
1385                 PredicateLockRelation(relation, snapshot);
1386
1387         /* we only need to set this up once */
1388         scan->rs_ctup.t_tableOid = RelationGetRelid(relation);
1389
1390         /*
1391          * we do this here instead of in initscan() because heap_rescan also calls
1392          * initscan() and we don't want to allocate memory again
1393          */
1394         if (nkeys > 0)
1395                 scan->rs_key = (ScanKey) palloc(sizeof(ScanKeyData) * nkeys);
1396         else
1397                 scan->rs_key = NULL;
1398
1399         initscan(scan, key, false);
1400
1401         return scan;
1402 }
1403
1404 /* ----------------
1405  *              heap_rescan             - restart a relation scan
1406  * ----------------
1407  */
1408 void
1409 heap_rescan(HeapScanDesc scan,
1410                         ScanKey key)
1411 {
1412         /*
1413          * unpin scan buffers
1414          */
1415         if (BufferIsValid(scan->rs_cbuf))
1416                 ReleaseBuffer(scan->rs_cbuf);
1417
1418         /*
1419          * reinitialize scan descriptor
1420          */
1421         initscan(scan, key, true);
1422 }
1423
1424 /* ----------------
1425  *              heap_endscan    - end relation scan
1426  *
1427  *              See how to integrate with index scans.
1428  *              Check handling if reldesc caching.
1429  * ----------------
1430  */
1431 void
1432 heap_endscan(HeapScanDesc scan)
1433 {
1434         /* Note: no locking manipulations needed */
1435
1436         /*
1437          * unpin scan buffers
1438          */
1439         if (BufferIsValid(scan->rs_cbuf))
1440                 ReleaseBuffer(scan->rs_cbuf);
1441
1442         /*
1443          * decrement relation reference count and free scan descriptor storage
1444          */
1445         RelationDecrementReferenceCount(scan->rs_rd);
1446
1447         if (scan->rs_key)
1448                 pfree(scan->rs_key);
1449
1450         if (scan->rs_strategy != NULL)
1451                 FreeAccessStrategy(scan->rs_strategy);
1452
1453         if (scan->rs_temp_snap)
1454                 UnregisterSnapshot(scan->rs_snapshot);
1455
1456         pfree(scan);
1457 }
1458
1459 /* ----------------
1460  *              heap_getnext    - retrieve next tuple in scan
1461  *
1462  *              Fix to work with index relations.
1463  *              We don't return the buffer anymore, but you can get it from the
1464  *              returned HeapTuple.
1465  * ----------------
1466  */
1467
1468 #ifdef HEAPDEBUGALL
1469 #define HEAPDEBUG_1 \
1470         elog(DEBUG2, "heap_getnext([%s,nkeys=%d],dir=%d) called", \
1471                  RelationGetRelationName(scan->rs_rd), scan->rs_nkeys, (int) direction)
1472 #define HEAPDEBUG_2 \
1473         elog(DEBUG2, "heap_getnext returning EOS")
1474 #define HEAPDEBUG_3 \
1475         elog(DEBUG2, "heap_getnext returning tuple")
1476 #else
1477 #define HEAPDEBUG_1
1478 #define HEAPDEBUG_2
1479 #define HEAPDEBUG_3
1480 #endif   /* !defined(HEAPDEBUGALL) */
1481
1482
1483 HeapTuple
1484 heap_getnext(HeapScanDesc scan, ScanDirection direction)
1485 {
1486         /* Note: no locking manipulations needed */
1487
1488         HEAPDEBUG_1;                            /* heap_getnext( info ) */
1489
1490         if (scan->rs_pageatatime)
1491                 heapgettup_pagemode(scan, direction,
1492                                                         scan->rs_nkeys, scan->rs_key);
1493         else
1494                 heapgettup(scan, direction, scan->rs_nkeys, scan->rs_key);
1495
1496         if (scan->rs_ctup.t_data == NULL)
1497         {
1498                 HEAPDEBUG_2;                    /* heap_getnext returning EOS */
1499                 return NULL;
1500         }
1501
1502         /*
1503          * if we get here it means we have a new current scan tuple, so point to
1504          * the proper return buffer and return the tuple.
1505          */
1506         HEAPDEBUG_3;                            /* heap_getnext returning tuple */
1507
1508         pgstat_count_heap_getnext(scan->rs_rd);
1509
1510         return &(scan->rs_ctup);
1511 }
1512
1513 /*
1514  *      heap_fetch              - retrieve tuple with given tid
1515  *
1516  * On entry, tuple->t_self is the TID to fetch.  We pin the buffer holding
1517  * the tuple, fill in the remaining fields of *tuple, and check the tuple
1518  * against the specified snapshot.
1519  *
1520  * If successful (tuple found and passes snapshot time qual), then *userbuf
1521  * is set to the buffer holding the tuple and TRUE is returned.  The caller
1522  * must unpin the buffer when done with the tuple.
1523  *
1524  * If the tuple is not found (ie, item number references a deleted slot),
1525  * then tuple->t_data is set to NULL and FALSE is returned.
1526  *
1527  * If the tuple is found but fails the time qual check, then FALSE is returned
1528  * but tuple->t_data is left pointing to the tuple.
1529  *
1530  * keep_buf determines what is done with the buffer in the FALSE-result cases.
1531  * When the caller specifies keep_buf = true, we retain the pin on the buffer
1532  * and return it in *userbuf (so the caller must eventually unpin it); when
1533  * keep_buf = false, the pin is released and *userbuf is set to InvalidBuffer.
1534  *
1535  * stats_relation is the relation to charge the heap_fetch operation against
1536  * for statistical purposes.  (This could be the heap rel itself, an
1537  * associated index, or NULL to not count the fetch at all.)
1538  *
1539  * heap_fetch does not follow HOT chains: only the exact TID requested will
1540  * be fetched.
1541  *
1542  * It is somewhat inconsistent that we ereport() on invalid block number but
1543  * return false on invalid item number.  There are a couple of reasons though.
1544  * One is that the caller can relatively easily check the block number for
1545  * validity, but cannot check the item number without reading the page
1546  * himself.  Another is that when we are following a t_ctid link, we can be
1547  * reasonably confident that the page number is valid (since VACUUM shouldn't
1548  * truncate off the destination page without having killed the referencing
1549  * tuple first), but the item number might well not be good.
1550  */
1551 bool
1552 heap_fetch(Relation relation,
1553                    Snapshot snapshot,
1554                    HeapTuple tuple,
1555                    Buffer *userbuf,
1556                    bool keep_buf,
1557                    Relation stats_relation)
1558 {
1559         ItemPointer tid = &(tuple->t_self);
1560         ItemId          lp;
1561         Buffer          buffer;
1562         Page            page;
1563         OffsetNumber offnum;
1564         bool            valid;
1565
1566         /*
1567          * Fetch and pin the appropriate page of the relation.
1568          */
1569         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1570
1571         /*
1572          * Need share lock on buffer to examine tuple commit status.
1573          */
1574         LockBuffer(buffer, BUFFER_LOCK_SHARE);
1575         page = BufferGetPage(buffer);
1576
1577         /*
1578          * We'd better check for out-of-range offnum in case of VACUUM since the
1579          * TID was obtained.
1580          */
1581         offnum = ItemPointerGetOffsetNumber(tid);
1582         if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1583         {
1584                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1585                 if (keep_buf)
1586                         *userbuf = buffer;
1587                 else
1588                 {
1589                         ReleaseBuffer(buffer);
1590                         *userbuf = InvalidBuffer;
1591                 }
1592                 tuple->t_data = NULL;
1593                 return false;
1594         }
1595
1596         /*
1597          * get the item line pointer corresponding to the requested tid
1598          */
1599         lp = PageGetItemId(page, offnum);
1600
1601         /*
1602          * Must check for deleted tuple.
1603          */
1604         if (!ItemIdIsNormal(lp))
1605         {
1606                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1607                 if (keep_buf)
1608                         *userbuf = buffer;
1609                 else
1610                 {
1611                         ReleaseBuffer(buffer);
1612                         *userbuf = InvalidBuffer;
1613                 }
1614                 tuple->t_data = NULL;
1615                 return false;
1616         }
1617
1618         /*
1619          * fill in *tuple fields
1620          */
1621         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1622         tuple->t_len = ItemIdGetLength(lp);
1623         tuple->t_tableOid = RelationGetRelid(relation);
1624
1625         /*
1626          * check time qualification of tuple, then release lock
1627          */
1628         valid = HeapTupleSatisfiesVisibility(tuple, snapshot, buffer);
1629
1630         if (valid)
1631                 PredicateLockTuple(relation, tuple, snapshot);
1632
1633         CheckForSerializableConflictOut(valid, relation, tuple, buffer, snapshot);
1634
1635         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1636
1637         if (valid)
1638         {
1639                 /*
1640                  * All checks passed, so return the tuple as valid. Caller is now
1641                  * responsible for releasing the buffer.
1642                  */
1643                 *userbuf = buffer;
1644
1645                 /* Count the successful fetch against appropriate rel, if any */
1646                 if (stats_relation != NULL)
1647                         pgstat_count_heap_fetch(stats_relation);
1648
1649                 return true;
1650         }
1651
1652         /* Tuple failed time qual, but maybe caller wants to see it anyway. */
1653         if (keep_buf)
1654                 *userbuf = buffer;
1655         else
1656         {
1657                 ReleaseBuffer(buffer);
1658                 *userbuf = InvalidBuffer;
1659         }
1660
1661         return false;
1662 }
1663
1664 /*
1665  *      heap_hot_search_buffer  - search HOT chain for tuple satisfying snapshot
1666  *
1667  * On entry, *tid is the TID of a tuple (either a simple tuple, or the root
1668  * of a HOT chain), and buffer is the buffer holding this tuple.  We search
1669  * for the first chain member satisfying the given snapshot.  If one is
1670  * found, we update *tid to reference that tuple's offset number, and
1671  * return TRUE.  If no match, return FALSE without modifying *tid.
1672  *
1673  * heapTuple is a caller-supplied buffer.  When a match is found, we return
1674  * the tuple here, in addition to updating *tid.  If no match is found, the
1675  * contents of this buffer on return are undefined.
1676  *
1677  * If all_dead is not NULL, we check non-visible tuples to see if they are
1678  * globally dead; *all_dead is set TRUE if all members of the HOT chain
1679  * are vacuumable, FALSE if not.
1680  *
1681  * Unlike heap_fetch, the caller must already have pin and (at least) share
1682  * lock on the buffer; it is still pinned/locked at exit.  Also unlike
1683  * heap_fetch, we do not report any pgstats count; caller may do so if wanted.
1684  */
1685 bool
1686 heap_hot_search_buffer(ItemPointer tid, Relation relation, Buffer buffer,
1687                                            Snapshot snapshot, HeapTuple heapTuple,
1688                                            bool *all_dead, bool first_call)
1689 {
1690         Page            dp = (Page) BufferGetPage(buffer);
1691         TransactionId prev_xmax = InvalidTransactionId;
1692         OffsetNumber offnum;
1693         bool            at_chain_start;
1694         bool            valid;
1695         bool            skip;
1696
1697         /* If this is not the first call, previous call returned a (live!) tuple */
1698         if (all_dead)
1699                 *all_dead = first_call;
1700
1701         Assert(TransactionIdIsValid(RecentGlobalXmin));
1702
1703         Assert(ItemPointerGetBlockNumber(tid) == BufferGetBlockNumber(buffer));
1704         offnum = ItemPointerGetOffsetNumber(tid);
1705         at_chain_start = first_call;
1706         skip = !first_call;
1707
1708         heapTuple->t_self = *tid;
1709
1710         /* Scan through possible multiple members of HOT-chain */
1711         for (;;)
1712         {
1713                 ItemId          lp;
1714
1715                 /* check for bogus TID */
1716                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(dp))
1717                         break;
1718
1719                 lp = PageGetItemId(dp, offnum);
1720
1721                 /* check for unused, dead, or redirected items */
1722                 if (!ItemIdIsNormal(lp))
1723                 {
1724                         /* We should only see a redirect at start of chain */
1725                         if (ItemIdIsRedirected(lp) && at_chain_start)
1726                         {
1727                                 /* Follow the redirect */
1728                                 offnum = ItemIdGetRedirect(lp);
1729                                 at_chain_start = false;
1730                                 continue;
1731                         }
1732                         /* else must be end of chain */
1733                         break;
1734                 }
1735
1736                 heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
1737                 heapTuple->t_len = ItemIdGetLength(lp);
1738                 heapTuple->t_tableOid = RelationGetRelid(relation);
1739                 ItemPointerSetOffsetNumber(&heapTuple->t_self, offnum);
1740
1741                 /*
1742                  * Shouldn't see a HEAP_ONLY tuple at chain start.
1743                  */
1744                 if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
1745                         break;
1746
1747                 /*
1748                  * The xmin should match the previous xmax value, else chain is
1749                  * broken.
1750                  */
1751                 if (TransactionIdIsValid(prev_xmax) &&
1752                         !TransactionIdEquals(prev_xmax,
1753                                                                  HeapTupleHeaderGetXmin(heapTuple->t_data)))
1754                         break;
1755
1756                 /*
1757                  * When first_call is true (and thus, skip is initially false) we'll
1758                  * return the first tuple we find.  But on later passes, heapTuple
1759                  * will initially be pointing to the tuple we returned last time.
1760                  * Returning it again would be incorrect (and would loop forever), so
1761                  * we skip it and return the next match we find.
1762                  */
1763                 if (!skip)
1764                 {
1765                         /*
1766                          * For the benefit of logical decoding, have t_self point at the
1767                          * element of the HOT chain we're currently investigating instead
1768                          * of the root tuple of the HOT chain. This is important because
1769                          * the *Satisfies routine for historical mvcc snapshots needs the
1770                          * correct tid to decide about the visibility in some cases.
1771                          */
1772                         ItemPointerSet(&(heapTuple->t_self), BufferGetBlockNumber(buffer), offnum);
1773
1774                         /* If it's visible per the snapshot, we must return it */
1775                         valid = HeapTupleSatisfiesVisibility(heapTuple, snapshot, buffer);
1776                         CheckForSerializableConflictOut(valid, relation, heapTuple,
1777                                                                                         buffer, snapshot);
1778                         /* reset to original, non-redirected, tid */
1779                         heapTuple->t_self = *tid;
1780
1781                         if (valid)
1782                         {
1783                                 ItemPointerSetOffsetNumber(tid, offnum);
1784                                 PredicateLockTuple(relation, heapTuple, snapshot);
1785                                 if (all_dead)
1786                                         *all_dead = false;
1787                                 return true;
1788                         }
1789                 }
1790                 skip = false;
1791
1792                 /*
1793                  * If we can't see it, maybe no one else can either.  At caller
1794                  * request, check whether all chain members are dead to all
1795                  * transactions.
1796                  */
1797                 if (all_dead && *all_dead &&
1798                         !HeapTupleIsSurelyDead(heapTuple, RecentGlobalXmin))
1799                         *all_dead = false;
1800
1801                 /*
1802                  * Check to see if HOT chain continues past this tuple; if so fetch
1803                  * the next offnum and loop around.
1804                  */
1805                 if (HeapTupleIsHotUpdated(heapTuple))
1806                 {
1807                         Assert(ItemPointerGetBlockNumber(&heapTuple->t_data->t_ctid) ==
1808                                    ItemPointerGetBlockNumber(tid));
1809                         offnum = ItemPointerGetOffsetNumber(&heapTuple->t_data->t_ctid);
1810                         at_chain_start = false;
1811                         prev_xmax = HeapTupleHeaderGetUpdateXid(heapTuple->t_data);
1812                 }
1813                 else
1814                         break;                          /* end of chain */
1815         }
1816
1817         return false;
1818 }
1819
1820 /*
1821  *      heap_hot_search         - search HOT chain for tuple satisfying snapshot
1822  *
1823  * This has the same API as heap_hot_search_buffer, except that the caller
1824  * does not provide the buffer containing the page, rather we access it
1825  * locally.
1826  */
1827 bool
1828 heap_hot_search(ItemPointer tid, Relation relation, Snapshot snapshot,
1829                                 bool *all_dead)
1830 {
1831         bool            result;
1832         Buffer          buffer;
1833         HeapTupleData heapTuple;
1834
1835         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1836         LockBuffer(buffer, BUFFER_LOCK_SHARE);
1837         result = heap_hot_search_buffer(tid, relation, buffer, snapshot,
1838                                                                         &heapTuple, all_dead, true);
1839         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1840         ReleaseBuffer(buffer);
1841         return result;
1842 }
1843
1844 /*
1845  *      heap_get_latest_tid -  get the latest tid of a specified tuple
1846  *
1847  * Actually, this gets the latest version that is visible according to
1848  * the passed snapshot.  You can pass SnapshotDirty to get the very latest,
1849  * possibly uncommitted version.
1850  *
1851  * *tid is both an input and an output parameter: it is updated to
1852  * show the latest version of the row.  Note that it will not be changed
1853  * if no version of the row passes the snapshot test.
1854  */
1855 void
1856 heap_get_latest_tid(Relation relation,
1857                                         Snapshot snapshot,
1858                                         ItemPointer tid)
1859 {
1860         BlockNumber blk;
1861         ItemPointerData ctid;
1862         TransactionId priorXmax;
1863
1864         /* this is to avoid Assert failures on bad input */
1865         if (!ItemPointerIsValid(tid))
1866                 return;
1867
1868         /*
1869          * Since this can be called with user-supplied TID, don't trust the input
1870          * too much.  (RelationGetNumberOfBlocks is an expensive check, so we
1871          * don't check t_ctid links again this way.  Note that it would not do to
1872          * call it just once and save the result, either.)
1873          */
1874         blk = ItemPointerGetBlockNumber(tid);
1875         if (blk >= RelationGetNumberOfBlocks(relation))
1876                 elog(ERROR, "block number %u is out of range for relation \"%s\"",
1877                          blk, RelationGetRelationName(relation));
1878
1879         /*
1880          * Loop to chase down t_ctid links.  At top of loop, ctid is the tuple we
1881          * need to examine, and *tid is the TID we will return if ctid turns out
1882          * to be bogus.
1883          *
1884          * Note that we will loop until we reach the end of the t_ctid chain.
1885          * Depending on the snapshot passed, there might be at most one visible
1886          * version of the row, but we don't try to optimize for that.
1887          */
1888         ctid = *tid;
1889         priorXmax = InvalidTransactionId;       /* cannot check first XMIN */
1890         for (;;)
1891         {
1892                 Buffer          buffer;
1893                 Page            page;
1894                 OffsetNumber offnum;
1895                 ItemId          lp;
1896                 HeapTupleData tp;
1897                 bool            valid;
1898
1899                 /*
1900                  * Read, pin, and lock the page.
1901                  */
1902                 buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&ctid));
1903                 LockBuffer(buffer, BUFFER_LOCK_SHARE);
1904                 page = BufferGetPage(buffer);
1905
1906                 /*
1907                  * Check for bogus item number.  This is not treated as an error
1908                  * condition because it can happen while following a t_ctid link. We
1909                  * just assume that the prior tid is OK and return it unchanged.
1910                  */
1911                 offnum = ItemPointerGetOffsetNumber(&ctid);
1912                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1913                 {
1914                         UnlockReleaseBuffer(buffer);
1915                         break;
1916                 }
1917                 lp = PageGetItemId(page, offnum);
1918                 if (!ItemIdIsNormal(lp))
1919                 {
1920                         UnlockReleaseBuffer(buffer);
1921                         break;
1922                 }
1923
1924                 /* OK to access the tuple */
1925                 tp.t_self = ctid;
1926                 tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
1927                 tp.t_len = ItemIdGetLength(lp);
1928                 tp.t_tableOid = RelationGetRelid(relation);
1929
1930                 /*
1931                  * After following a t_ctid link, we might arrive at an unrelated
1932                  * tuple.  Check for XMIN match.
1933                  */
1934                 if (TransactionIdIsValid(priorXmax) &&
1935                   !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(tp.t_data)))
1936                 {
1937                         UnlockReleaseBuffer(buffer);
1938                         break;
1939                 }
1940
1941                 /*
1942                  * Check time qualification of tuple; if visible, set it as the new
1943                  * result candidate.
1944                  */
1945                 valid = HeapTupleSatisfiesVisibility(&tp, snapshot, buffer);
1946                 CheckForSerializableConflictOut(valid, relation, &tp, buffer, snapshot);
1947                 if (valid)
1948                         *tid = ctid;
1949
1950                 /*
1951                  * If there's a valid t_ctid link, follow it, else we're done.
1952                  */
1953                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
1954                         HeapTupleHeaderIsOnlyLocked(tp.t_data) ||
1955                         ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
1956                 {
1957                         UnlockReleaseBuffer(buffer);
1958                         break;
1959                 }
1960
1961                 ctid = tp.t_data->t_ctid;
1962                 priorXmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
1963                 UnlockReleaseBuffer(buffer);
1964         }                                                       /* end of loop */
1965 }
1966
1967
1968 /*
1969  * UpdateXmaxHintBits - update tuple hint bits after xmax transaction ends
1970  *
1971  * This is called after we have waited for the XMAX transaction to terminate.
1972  * If the transaction aborted, we guarantee the XMAX_INVALID hint bit will
1973  * be set on exit.  If the transaction committed, we set the XMAX_COMMITTED
1974  * hint bit if possible --- but beware that that may not yet be possible,
1975  * if the transaction committed asynchronously.
1976  *
1977  * Note that if the transaction was a locker only, we set HEAP_XMAX_INVALID
1978  * even if it commits.
1979  *
1980  * Hence callers should look only at XMAX_INVALID.
1981  *
1982  * Note this is not allowed for tuples whose xmax is a multixact.
1983  */
1984 static void
1985 UpdateXmaxHintBits(HeapTupleHeader tuple, Buffer buffer, TransactionId xid)
1986 {
1987         Assert(TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple), xid));
1988         Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));
1989
1990         if (!(tuple->t_infomask & (HEAP_XMAX_COMMITTED | HEAP_XMAX_INVALID)))
1991         {
1992                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
1993                         TransactionIdDidCommit(xid))
1994                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_COMMITTED,
1995                                                                  xid);
1996                 else
1997                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_INVALID,
1998                                                                  InvalidTransactionId);
1999         }
2000 }
2001
2002
2003 /*
2004  * GetBulkInsertState - prepare status object for a bulk insert
2005  */
2006 BulkInsertState
2007 GetBulkInsertState(void)
2008 {
2009         BulkInsertState bistate;
2010
2011         bistate = (BulkInsertState) palloc(sizeof(BulkInsertStateData));
2012         bistate->strategy = GetAccessStrategy(BAS_BULKWRITE);
2013         bistate->current_buf = InvalidBuffer;
2014         return bistate;
2015 }
2016
2017 /*
2018  * FreeBulkInsertState - clean up after finishing a bulk insert
2019  */
2020 void
2021 FreeBulkInsertState(BulkInsertState bistate)
2022 {
2023         if (bistate->current_buf != InvalidBuffer)
2024                 ReleaseBuffer(bistate->current_buf);
2025         FreeAccessStrategy(bistate->strategy);
2026         pfree(bistate);
2027 }
2028
2029
2030 /*
2031  *      heap_insert             - insert tuple into a heap
2032  *
2033  * The new tuple is stamped with current transaction ID and the specified
2034  * command ID.
2035  *
2036  * If the HEAP_INSERT_SKIP_WAL option is specified, the new tuple is not
2037  * logged in WAL, even for a non-temp relation.  Safe usage of this behavior
2038  * requires that we arrange that all new tuples go into new pages not
2039  * containing any tuples from other transactions, and that the relation gets
2040  * fsync'd before commit.  (See also heap_sync() comments)
2041  *
2042  * The HEAP_INSERT_SKIP_FSM option is passed directly to
2043  * RelationGetBufferForTuple, which see for more info.
2044  *
2045  * HEAP_INSERT_FROZEN should only be specified for inserts into
2046  * relfilenodes created during the current subtransaction and when
2047  * there are no prior snapshots or pre-existing portals open.
2048  * This causes rows to be frozen, which is an MVCC violation and
2049  * requires explicit options chosen by user.
2050  *
2051  * Note that these options will be applied when inserting into the heap's
2052  * TOAST table, too, if the tuple requires any out-of-line data.
2053  *
2054  * The BulkInsertState object (if any; bistate can be NULL for default
2055  * behavior) is also just passed through to RelationGetBufferForTuple.
2056  *
2057  * The return value is the OID assigned to the tuple (either here or by the
2058  * caller), or InvalidOid if no OID.  The header fields of *tup are updated
2059  * to match the stored tuple; in particular tup->t_self receives the actual
2060  * TID where the tuple was stored.  But note that any toasting of fields
2061  * within the tuple data is NOT reflected into *tup.
2062  */
2063 Oid
2064 heap_insert(Relation relation, HeapTuple tup, CommandId cid,
2065                         int options, BulkInsertState bistate)
2066 {
2067         TransactionId xid = GetCurrentTransactionId();
2068         HeapTuple       heaptup;
2069         Buffer          buffer;
2070         Buffer          vmbuffer = InvalidBuffer;
2071         bool            all_visible_cleared = false;
2072
2073         /*
2074          * Fill in tuple header fields, assign an OID, and toast the tuple if
2075          * necessary.
2076          *
2077          * Note: below this point, heaptup is the data we actually intend to store
2078          * into the relation; tup is the caller's original untoasted data.
2079          */
2080         heaptup = heap_prepare_insert(relation, tup, xid, cid, options);
2081
2082         /*
2083          * We're about to do the actual insert -- but check for conflict first, to
2084          * avoid possibly having to roll back work we've just done.
2085          *
2086          * For a heap insert, we only need to check for table-level SSI locks. Our
2087          * new tuple can't possibly conflict with existing tuple locks, and heap
2088          * page locks are only consolidated versions of tuple locks; they do not
2089          * lock "gaps" as index page locks do.  So we don't need to identify a
2090          * buffer before making the call.
2091          */
2092         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2093
2094         /*
2095          * Find buffer to insert this tuple into.  If the page is all visible,
2096          * this will also pin the requisite visibility map page.
2097          */
2098         buffer = RelationGetBufferForTuple(relation, heaptup->t_len,
2099                                                                            InvalidBuffer, options, bistate,
2100                                                                            &vmbuffer, NULL);
2101
2102         /* NO EREPORT(ERROR) from here till changes are logged */
2103         START_CRIT_SECTION();
2104
2105         RelationPutHeapTuple(relation, buffer, heaptup);
2106
2107         if (PageIsAllVisible(BufferGetPage(buffer)))
2108         {
2109                 all_visible_cleared = true;
2110                 PageClearAllVisible(BufferGetPage(buffer));
2111                 visibilitymap_clear(relation,
2112                                                         ItemPointerGetBlockNumber(&(heaptup->t_self)),
2113                                                         vmbuffer);
2114         }
2115
2116         /*
2117          * XXX Should we set PageSetPrunable on this page ?
2118          *
2119          * The inserting transaction may eventually abort thus making this tuple
2120          * DEAD and hence available for pruning. Though we don't want to optimize
2121          * for aborts, if no other tuple in this page is UPDATEd/DELETEd, the
2122          * aborted tuple will never be pruned until next vacuum is triggered.
2123          *
2124          * If you do add PageSetPrunable here, add it in heap_xlog_insert too.
2125          */
2126
2127         MarkBufferDirty(buffer);
2128
2129         /* XLOG stuff */
2130         if (!(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation))
2131         {
2132                 xl_heap_insert xlrec;
2133                 xl_heap_header xlhdr;
2134                 XLogRecPtr      recptr;
2135                 Page            page = BufferGetPage(buffer);
2136                 uint8           info = XLOG_HEAP_INSERT;
2137                 int                     bufflags = 0;
2138
2139                 /*
2140                  * If this is a catalog, we need to transmit combocids to properly
2141                  * decode, so log that as well.
2142                  */
2143                 if (RelationIsAccessibleInLogicalDecoding(relation))
2144                         log_heap_new_cid(relation, heaptup);
2145
2146                 /*
2147                  * If this is the single and first tuple on page, we can reinit the
2148                  * page instead of restoring the whole thing.  Set flag, and hide
2149                  * buffer references from XLogInsert.
2150                  */
2151                 if (ItemPointerGetOffsetNumber(&(heaptup->t_self)) == FirstOffsetNumber &&
2152                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
2153                 {
2154                         info |= XLOG_HEAP_INIT_PAGE;
2155                         bufflags |= REGBUF_WILL_INIT;
2156                 }
2157
2158                 xlrec.offnum = ItemPointerGetOffsetNumber(&heaptup->t_self);
2159                 xlrec.flags = all_visible_cleared ? XLOG_HEAP_ALL_VISIBLE_CLEARED : 0;
2160                 Assert(ItemPointerGetBlockNumber(&heaptup->t_self) == BufferGetBlockNumber(buffer));
2161
2162                 /*
2163                  * For logical decoding, we need the tuple even if we're doing a full
2164                  * page write, so make sure it's included even if we take a full-page
2165                  * image. (XXX We could alternatively store a pointer into the FPW).
2166                  */
2167                 if (RelationIsLogicallyLogged(relation))
2168                 {
2169                         xlrec.flags |= XLOG_HEAP_CONTAINS_NEW_TUPLE;
2170                         bufflags |= REGBUF_KEEP_DATA;
2171                 }
2172
2173                 XLogBeginInsert();
2174                 XLogRegisterData((char *) &xlrec, SizeOfHeapInsert);
2175
2176                 xlhdr.t_infomask2 = heaptup->t_data->t_infomask2;
2177                 xlhdr.t_infomask = heaptup->t_data->t_infomask;
2178                 xlhdr.t_hoff = heaptup->t_data->t_hoff;
2179
2180                 /*
2181                  * note we mark xlhdr as belonging to buffer; if XLogInsert decides to
2182                  * write the whole page to the xlog, we don't need to store
2183                  * xl_heap_header in the xlog.
2184                  */
2185                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2186                 XLogRegisterBufData(0, (char *) &xlhdr, SizeOfHeapHeader);
2187                 /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
2188                 XLogRegisterBufData(0,
2189                         (char *) heaptup->t_data + offsetof(HeapTupleHeaderData, t_bits),
2190                                          heaptup->t_len - offsetof(HeapTupleHeaderData, t_bits));
2191
2192                 recptr = XLogInsert(RM_HEAP_ID, info);
2193
2194                 PageSetLSN(page, recptr);
2195         }
2196
2197         END_CRIT_SECTION();
2198
2199         UnlockReleaseBuffer(buffer);
2200         if (vmbuffer != InvalidBuffer)
2201                 ReleaseBuffer(vmbuffer);
2202
2203         /*
2204          * If tuple is cachable, mark it for invalidation from the caches in case
2205          * we abort.  Note it is OK to do this after releasing the buffer, because
2206          * the heaptup data structure is all in local memory, not in the shared
2207          * buffer.
2208          */
2209         CacheInvalidateHeapTuple(relation, heaptup, NULL);
2210
2211         pgstat_count_heap_insert(relation, 1);
2212
2213         /*
2214          * If heaptup is a private copy, release it.  Don't forget to copy t_self
2215          * back to the caller's image, too.
2216          */
2217         if (heaptup != tup)
2218         {
2219                 tup->t_self = heaptup->t_self;
2220                 heap_freetuple(heaptup);
2221         }
2222
2223         return HeapTupleGetOid(tup);
2224 }
2225
2226 /*
2227  * Subroutine for heap_insert(). Prepares a tuple for insertion. This sets the
2228  * tuple header fields, assigns an OID, and toasts the tuple if necessary.
2229  * Returns a toasted version of the tuple if it was toasted, or the original
2230  * tuple if not. Note that in any case, the header fields are also set in
2231  * the original tuple.
2232  */
2233 static HeapTuple
2234 heap_prepare_insert(Relation relation, HeapTuple tup, TransactionId xid,
2235                                         CommandId cid, int options)
2236 {
2237         if (relation->rd_rel->relhasoids)
2238         {
2239 #ifdef NOT_USED
2240                 /* this is redundant with an Assert in HeapTupleSetOid */
2241                 Assert(tup->t_data->t_infomask & HEAP_HASOID);
2242 #endif
2243
2244                 /*
2245                  * If the object id of this tuple has already been assigned, trust the
2246                  * caller.  There are a couple of ways this can happen.  At initial db
2247                  * creation, the backend program sets oids for tuples. When we define
2248                  * an index, we set the oid.  Finally, in the future, we may allow
2249                  * users to set their own object ids in order to support a persistent
2250                  * object store (objects need to contain pointers to one another).
2251                  */
2252                 if (!OidIsValid(HeapTupleGetOid(tup)))
2253                         HeapTupleSetOid(tup, GetNewOid(relation));
2254         }
2255         else
2256         {
2257                 /* check there is not space for an OID */
2258                 Assert(!(tup->t_data->t_infomask & HEAP_HASOID));
2259         }
2260
2261         tup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
2262         tup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
2263         tup->t_data->t_infomask |= HEAP_XMAX_INVALID;
2264         HeapTupleHeaderSetXmin(tup->t_data, xid);
2265         if (options & HEAP_INSERT_FROZEN)
2266                 HeapTupleHeaderSetXminFrozen(tup->t_data);
2267
2268         HeapTupleHeaderSetCmin(tup->t_data, cid);
2269         HeapTupleHeaderSetXmax(tup->t_data, 0);         /* for cleanliness */
2270         tup->t_tableOid = RelationGetRelid(relation);
2271
2272         /*
2273          * If the new tuple is too big for storage or contains already toasted
2274          * out-of-line attributes from some other relation, invoke the toaster.
2275          */
2276         if (relation->rd_rel->relkind != RELKIND_RELATION &&
2277                 relation->rd_rel->relkind != RELKIND_MATVIEW)
2278         {
2279                 /* toast table entries should never be recursively toasted */
2280                 Assert(!HeapTupleHasExternal(tup));
2281                 return tup;
2282         }
2283         else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
2284                 return toast_insert_or_update(relation, tup, NULL, options);
2285         else
2286                 return tup;
2287 }
2288
2289 /*
2290  *      heap_multi_insert       - insert multiple tuple into a heap
2291  *
2292  * This is like heap_insert(), but inserts multiple tuples in one operation.
2293  * That's faster than calling heap_insert() in a loop, because when multiple
2294  * tuples can be inserted on a single page, we can write just a single WAL
2295  * record covering all of them, and only need to lock/unlock the page once.
2296  *
2297  * Note: this leaks memory into the current memory context. You can create a
2298  * temporary context before calling this, if that's a problem.
2299  */
2300 void
2301 heap_multi_insert(Relation relation, HeapTuple *tuples, int ntuples,
2302                                   CommandId cid, int options, BulkInsertState bistate)
2303 {
2304         TransactionId xid = GetCurrentTransactionId();
2305         HeapTuple  *heaptuples;
2306         int                     i;
2307         int                     ndone;
2308         char       *scratch = NULL;
2309         Page            page;
2310         bool            needwal;
2311         Size            saveFreeSpace;
2312         bool            need_tuple_data = RelationIsLogicallyLogged(relation);
2313         bool            need_cids = RelationIsAccessibleInLogicalDecoding(relation);
2314
2315         needwal = !(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation);
2316         saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
2317                                                                                                    HEAP_DEFAULT_FILLFACTOR);
2318
2319         /* Toast and set header data in all the tuples */
2320         heaptuples = palloc(ntuples * sizeof(HeapTuple));
2321         for (i = 0; i < ntuples; i++)
2322                 heaptuples[i] = heap_prepare_insert(relation, tuples[i],
2323                                                                                         xid, cid, options);
2324
2325         /*
2326          * Allocate some memory to use for constructing the WAL record. Using
2327          * palloc() within a critical section is not safe, so we allocate this
2328          * beforehand.
2329          */
2330         if (needwal)
2331                 scratch = palloc(BLCKSZ);
2332
2333         /*
2334          * We're about to do the actual inserts -- but check for conflict first,
2335          * to avoid possibly having to roll back work we've just done.
2336          *
2337          * For a heap insert, we only need to check for table-level SSI locks. Our
2338          * new tuple can't possibly conflict with existing tuple locks, and heap
2339          * page locks are only consolidated versions of tuple locks; they do not
2340          * lock "gaps" as index page locks do.  So we don't need to identify a
2341          * buffer before making the call.
2342          */
2343         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2344
2345         ndone = 0;
2346         while (ndone < ntuples)
2347         {
2348                 Buffer          buffer;
2349                 Buffer          vmbuffer = InvalidBuffer;
2350                 bool            all_visible_cleared = false;
2351                 int                     nthispage;
2352
2353                 CHECK_FOR_INTERRUPTS();
2354
2355                 /*
2356                  * Find buffer where at least the next tuple will fit.  If the page is
2357                  * all-visible, this will also pin the requisite visibility map page.
2358                  */
2359                 buffer = RelationGetBufferForTuple(relation, heaptuples[ndone]->t_len,
2360                                                                                    InvalidBuffer, options, bistate,
2361                                                                                    &vmbuffer, NULL);
2362                 page = BufferGetPage(buffer);
2363
2364                 /* NO EREPORT(ERROR) from here till changes are logged */
2365                 START_CRIT_SECTION();
2366
2367                 /*
2368                  * RelationGetBufferForTuple has ensured that the first tuple fits.
2369                  * Put that on the page, and then as many other tuples as fit.
2370                  */
2371                 RelationPutHeapTuple(relation, buffer, heaptuples[ndone]);
2372                 for (nthispage = 1; ndone + nthispage < ntuples; nthispage++)
2373                 {
2374                         HeapTuple       heaptup = heaptuples[ndone + nthispage];
2375
2376                         if (PageGetHeapFreeSpace(page) < MAXALIGN(heaptup->t_len) + saveFreeSpace)
2377                                 break;
2378
2379                         RelationPutHeapTuple(relation, buffer, heaptup);
2380
2381                         /*
2382                          * We don't use heap_multi_insert for catalog tuples yet, but
2383                          * better be prepared...
2384                          */
2385                         if (needwal && need_cids)
2386                                 log_heap_new_cid(relation, heaptup);
2387                 }
2388
2389                 if (PageIsAllVisible(page))
2390                 {
2391                         all_visible_cleared = true;
2392                         PageClearAllVisible(page);
2393                         visibilitymap_clear(relation,
2394                                                                 BufferGetBlockNumber(buffer),
2395                                                                 vmbuffer);
2396                 }
2397
2398                 /*
2399                  * XXX Should we set PageSetPrunable on this page ? See heap_insert()
2400                  */
2401
2402                 MarkBufferDirty(buffer);
2403
2404                 /* XLOG stuff */
2405                 if (needwal)
2406                 {
2407                         XLogRecPtr      recptr;
2408                         xl_heap_multi_insert *xlrec;
2409                         uint8           info = XLOG_HEAP2_MULTI_INSERT;
2410                         char       *tupledata;
2411                         int                     totaldatalen;
2412                         char       *scratchptr = scratch;
2413                         bool            init;
2414                         int                     bufflags = 0;
2415
2416                         /*
2417                          * If the page was previously empty, we can reinit the page
2418                          * instead of restoring the whole thing.
2419                          */
2420                         init = (ItemPointerGetOffsetNumber(&(heaptuples[ndone]->t_self)) == FirstOffsetNumber &&
2421                                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber + nthispage - 1);
2422
2423                         /* allocate xl_heap_multi_insert struct from the scratch area */
2424                         xlrec = (xl_heap_multi_insert *) scratchptr;
2425                         scratchptr += SizeOfHeapMultiInsert;
2426
2427                         /*
2428                          * Allocate offsets array. Unless we're reinitializing the page,
2429                          * in that case the tuples are stored in order starting at
2430                          * FirstOffsetNumber and we don't need to store the offsets
2431                          * explicitly.
2432                          */
2433                         if (!init)
2434                                 scratchptr += nthispage * sizeof(OffsetNumber);
2435
2436                         /* the rest of the scratch space is used for tuple data */
2437                         tupledata = scratchptr;
2438
2439                         xlrec->flags = all_visible_cleared ? XLOG_HEAP_ALL_VISIBLE_CLEARED : 0;
2440                         xlrec->ntuples = nthispage;
2441
2442                         /*
2443                          * Write out an xl_multi_insert_tuple and the tuple data itself
2444                          * for each tuple.
2445                          */
2446                         for (i = 0; i < nthispage; i++)
2447                         {
2448                                 HeapTuple       heaptup = heaptuples[ndone + i];
2449                                 xl_multi_insert_tuple *tuphdr;
2450                                 int                     datalen;
2451
2452                                 if (!init)
2453                                         xlrec->offsets[i] = ItemPointerGetOffsetNumber(&heaptup->t_self);
2454                                 /* xl_multi_insert_tuple needs two-byte alignment. */
2455                                 tuphdr = (xl_multi_insert_tuple *) SHORTALIGN(scratchptr);
2456                                 scratchptr = ((char *) tuphdr) + SizeOfMultiInsertTuple;
2457
2458                                 tuphdr->t_infomask2 = heaptup->t_data->t_infomask2;
2459                                 tuphdr->t_infomask = heaptup->t_data->t_infomask;
2460                                 tuphdr->t_hoff = heaptup->t_data->t_hoff;
2461
2462                                 /* write bitmap [+ padding] [+ oid] + data */
2463                                 datalen = heaptup->t_len - offsetof(HeapTupleHeaderData, t_bits);
2464                                 memcpy(scratchptr,
2465                                            (char *) heaptup->t_data + offsetof(HeapTupleHeaderData, t_bits),
2466                                            datalen);
2467                                 tuphdr->datalen = datalen;
2468                                 scratchptr += datalen;
2469                         }
2470                         totaldatalen = scratchptr - tupledata;
2471                         Assert((scratchptr - scratch) < BLCKSZ);
2472
2473                         if (need_tuple_data)
2474                                 xlrec->flags |= XLOG_HEAP_CONTAINS_NEW_TUPLE;
2475
2476                         /*
2477                          * Signal that this is the last xl_heap_multi_insert record
2478                          * emitted by this call to heap_multi_insert(). Needed for logical
2479                          * decoding so it knows when to cleanup temporary data.
2480                          */
2481                         if (ndone + nthispage == ntuples)
2482                                 xlrec->flags |= XLOG_HEAP_LAST_MULTI_INSERT;
2483
2484                         if (init)
2485                         {
2486                                 info |= XLOG_HEAP_INIT_PAGE;
2487                                 bufflags |= REGBUF_WILL_INIT;
2488                         }
2489
2490                         /*
2491                          * If we're doing logical decoding, include the new tuple data
2492                          * even if we take a full-page image of the page.
2493                          */
2494                         if (need_tuple_data)
2495                                 bufflags |= REGBUF_KEEP_DATA;
2496
2497                         XLogBeginInsert();
2498                         XLogRegisterData((char *) xlrec, tupledata - scratch);
2499                         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2500
2501                         XLogRegisterBufData(0, tupledata, totaldatalen);
2502                         recptr = XLogInsert(RM_HEAP2_ID, info);
2503
2504                         PageSetLSN(page, recptr);
2505                 }
2506
2507                 END_CRIT_SECTION();
2508
2509                 UnlockReleaseBuffer(buffer);
2510                 if (vmbuffer != InvalidBuffer)
2511                         ReleaseBuffer(vmbuffer);
2512
2513                 ndone += nthispage;
2514         }
2515
2516         /*
2517          * If tuples are cachable, mark them for invalidation from the caches in
2518          * case we abort.  Note it is OK to do this after releasing the buffer,
2519          * because the heaptuples data structure is all in local memory, not in
2520          * the shared buffer.
2521          */
2522         if (IsCatalogRelation(relation))
2523         {
2524                 for (i = 0; i < ntuples; i++)
2525                         CacheInvalidateHeapTuple(relation, heaptuples[i], NULL);
2526         }
2527
2528         /*
2529          * Copy t_self fields back to the caller's original tuples. This does
2530          * nothing for untoasted tuples (tuples[i] == heaptuples[i)], but it's
2531          * probably faster to always copy than check.
2532          */
2533         for (i = 0; i < ntuples; i++)
2534                 tuples[i]->t_self = heaptuples[i]->t_self;
2535
2536         pgstat_count_heap_insert(relation, ntuples);
2537 }
2538
2539 /*
2540  *      simple_heap_insert - insert a tuple
2541  *
2542  * Currently, this routine differs from heap_insert only in supplying
2543  * a default command ID and not allowing access to the speedup options.
2544  *
2545  * This should be used rather than using heap_insert directly in most places
2546  * where we are modifying system catalogs.
2547  */
2548 Oid
2549 simple_heap_insert(Relation relation, HeapTuple tup)
2550 {
2551         return heap_insert(relation, tup, GetCurrentCommandId(true), 0, NULL);
2552 }
2553
2554 /*
2555  * Given infomask/infomask2, compute the bits that must be saved in the
2556  * "infobits" field of xl_heap_delete, xl_heap_update, xl_heap_lock,
2557  * xl_heap_lock_updated WAL records.
2558  *
2559  * See fix_infomask_from_infobits.
2560  */
2561 static uint8
2562 compute_infobits(uint16 infomask, uint16 infomask2)
2563 {
2564         return
2565                 ((infomask & HEAP_XMAX_IS_MULTI) != 0 ? XLHL_XMAX_IS_MULTI : 0) |
2566                 ((infomask & HEAP_XMAX_LOCK_ONLY) != 0 ? XLHL_XMAX_LOCK_ONLY : 0) |
2567                 ((infomask & HEAP_XMAX_EXCL_LOCK) != 0 ? XLHL_XMAX_EXCL_LOCK : 0) |
2568         /* note we ignore HEAP_XMAX_SHR_LOCK here */
2569                 ((infomask & HEAP_XMAX_KEYSHR_LOCK) != 0 ? XLHL_XMAX_KEYSHR_LOCK : 0) |
2570                 ((infomask2 & HEAP_KEYS_UPDATED) != 0 ?
2571                  XLHL_KEYS_UPDATED : 0);
2572 }
2573
2574 /*
2575  * Given two versions of the same t_infomask for a tuple, compare them and
2576  * return whether the relevant status for a tuple Xmax has changed.  This is
2577  * used after a buffer lock has been released and reacquired: we want to ensure
2578  * that the tuple state continues to be the same it was when we previously
2579  * examined it.
2580  *
2581  * Note the Xmax field itself must be compared separately.
2582  */
2583 static inline bool
2584 xmax_infomask_changed(uint16 new_infomask, uint16 old_infomask)
2585 {
2586         const uint16 interesting =
2587         HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY | HEAP_LOCK_MASK;
2588
2589         if ((new_infomask & interesting) != (old_infomask & interesting))
2590                 return true;
2591
2592         return false;
2593 }
2594
2595 /*
2596  *      heap_delete - delete a tuple
2597  *
2598  * NB: do not call this directly unless you are prepared to deal with
2599  * concurrent-update conditions.  Use simple_heap_delete instead.
2600  *
2601  *      relation - table to be modified (caller must hold suitable lock)
2602  *      tid - TID of tuple to be deleted
2603  *      cid - delete command ID (used for visibility test, and stored into
2604  *              cmax if successful)
2605  *      crosscheck - if not InvalidSnapshot, also check tuple against this
2606  *      wait - true if should wait for any conflicting update to commit/abort
2607  *      hufd - output parameter, filled in failure cases (see below)
2608  *
2609  * Normal, successful return value is HeapTupleMayBeUpdated, which
2610  * actually means we did delete it.  Failure return codes are
2611  * HeapTupleSelfUpdated, HeapTupleUpdated, or HeapTupleBeingUpdated
2612  * (the last only possible if wait == false).
2613  *
2614  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
2615  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
2616  * (the last only for HeapTupleSelfUpdated, since we
2617  * cannot obtain cmax from a combocid generated by another transaction).
2618  * See comments for struct HeapUpdateFailureData for additional info.
2619  */
2620 HTSU_Result
2621 heap_delete(Relation relation, ItemPointer tid,
2622                         CommandId cid, Snapshot crosscheck, bool wait,
2623                         HeapUpdateFailureData *hufd)
2624 {
2625         HTSU_Result result;
2626         TransactionId xid = GetCurrentTransactionId();
2627         ItemId          lp;
2628         HeapTupleData tp;
2629         Page            page;
2630         BlockNumber block;
2631         Buffer          buffer;
2632         Buffer          vmbuffer = InvalidBuffer;
2633         TransactionId new_xmax;
2634         uint16          new_infomask,
2635                                 new_infomask2;
2636         bool            have_tuple_lock = false;
2637         bool            iscombo;
2638         bool            all_visible_cleared = false;
2639         HeapTuple       old_key_tuple = NULL;   /* replica identity of the tuple */
2640         bool            old_key_copied = false;
2641
2642         Assert(ItemPointerIsValid(tid));
2643
2644         block = ItemPointerGetBlockNumber(tid);
2645         buffer = ReadBuffer(relation, block);
2646         page = BufferGetPage(buffer);
2647
2648         /*
2649          * Before locking the buffer, pin the visibility map page if it appears to
2650          * be necessary.  Since we haven't got the lock yet, someone else might be
2651          * in the middle of changing this, so we'll need to recheck after we have
2652          * the lock.
2653          */
2654         if (PageIsAllVisible(page))
2655                 visibilitymap_pin(relation, block, &vmbuffer);
2656
2657         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2658
2659         /*
2660          * If we didn't pin the visibility map page and the page has become all
2661          * visible while we were busy locking the buffer, we'll have to unlock and
2662          * re-lock, to avoid holding the buffer lock across an I/O.  That's a bit
2663          * unfortunate, but hopefully shouldn't happen often.
2664          */
2665         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
2666         {
2667                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2668                 visibilitymap_pin(relation, block, &vmbuffer);
2669                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2670         }
2671
2672         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
2673         Assert(ItemIdIsNormal(lp));
2674
2675         tp.t_tableOid = RelationGetRelid(relation);
2676         tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
2677         tp.t_len = ItemIdGetLength(lp);
2678         tp.t_self = *tid;
2679
2680 l1:
2681         result = HeapTupleSatisfiesUpdate(&tp, cid, buffer);
2682
2683         if (result == HeapTupleInvisible)
2684         {
2685                 UnlockReleaseBuffer(buffer);
2686                 elog(ERROR, "attempted to delete invisible tuple");
2687         }
2688         else if (result == HeapTupleBeingUpdated && wait)
2689         {
2690                 TransactionId xwait;
2691                 uint16          infomask;
2692
2693                 /* must copy state data before unlocking buffer */
2694                 xwait = HeapTupleHeaderGetRawXmax(tp.t_data);
2695                 infomask = tp.t_data->t_infomask;
2696
2697                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2698
2699                 /*
2700                  * Acquire tuple lock to establish our priority for the tuple (see
2701                  * heap_lock_tuple).  LockTuple will release us when we are
2702                  * next-in-line for the tuple.
2703                  *
2704                  * If we are forced to "start over" below, we keep the tuple lock;
2705                  * this arranges that we stay at the head of the line while rechecking
2706                  * tuple state.
2707                  */
2708                 if (!have_tuple_lock)
2709                 {
2710                         LockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2711                         have_tuple_lock = true;
2712                 }
2713
2714                 /*
2715                  * Sleep until concurrent transaction ends.  Note that we don't care
2716                  * which lock mode the locker has, because we need the strongest one.
2717                  */
2718
2719                 if (infomask & HEAP_XMAX_IS_MULTI)
2720                 {
2721                         /* wait for multixact */
2722                         MultiXactIdWait((MultiXactId) xwait, MultiXactStatusUpdate, infomask,
2723                                                         relation, &tp.t_data->t_ctid, XLTW_Delete,
2724                                                         NULL);
2725                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2726
2727                         /*
2728                          * If xwait had just locked the tuple then some other xact could
2729                          * update this tuple before we get to this point.  Check for xmax
2730                          * change, and start over if so.
2731                          */
2732                         if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2733                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2734                                                                          xwait))
2735                                 goto l1;
2736
2737                         /*
2738                          * You might think the multixact is necessarily done here, but not
2739                          * so: it could have surviving members, namely our own xact or
2740                          * other subxacts of this backend.  It is legal for us to delete
2741                          * the tuple in either case, however (the latter case is
2742                          * essentially a situation of upgrading our former shared lock to
2743                          * exclusive).  We don't bother changing the on-disk hint bits
2744                          * since we are about to overwrite the xmax altogether.
2745                          */
2746                 }
2747                 else
2748                 {
2749                         /* wait for regular transaction to end */
2750                         XactLockTableWait(xwait, relation, &tp.t_data->t_ctid, XLTW_Delete);
2751                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2752
2753                         /*
2754                          * xwait is done, but if xwait had just locked the tuple then some
2755                          * other xact could update this tuple before we get to this point.
2756                          * Check for xmax change, and start over if so.
2757                          */
2758                         if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2759                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2760                                                                          xwait))
2761                                 goto l1;
2762
2763                         /* Otherwise check if it committed or aborted */
2764                         UpdateXmaxHintBits(tp.t_data, buffer, xwait);
2765                 }
2766
2767                 /*
2768                  * We may overwrite if previous xmax aborted, or if it committed but
2769                  * only locked the tuple without updating it.
2770                  */
2771                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
2772                         HEAP_XMAX_IS_LOCKED_ONLY(tp.t_data->t_infomask) ||
2773                         HeapTupleHeaderIsOnlyLocked(tp.t_data))
2774                         result = HeapTupleMayBeUpdated;
2775                 else
2776                         result = HeapTupleUpdated;
2777         }
2778
2779         if (crosscheck != InvalidSnapshot && result == HeapTupleMayBeUpdated)
2780         {
2781                 /* Perform additional check for transaction-snapshot mode RI updates */
2782                 if (!HeapTupleSatisfiesVisibility(&tp, crosscheck, buffer))
2783                         result = HeapTupleUpdated;
2784         }
2785
2786         if (result != HeapTupleMayBeUpdated)
2787         {
2788                 Assert(result == HeapTupleSelfUpdated ||
2789                            result == HeapTupleUpdated ||
2790                            result == HeapTupleBeingUpdated);
2791                 Assert(!(tp.t_data->t_infomask & HEAP_XMAX_INVALID));
2792                 hufd->ctid = tp.t_data->t_ctid;
2793                 hufd->xmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
2794                 if (result == HeapTupleSelfUpdated)
2795                         hufd->cmax = HeapTupleHeaderGetCmax(tp.t_data);
2796                 else
2797                         hufd->cmax = InvalidCommandId;
2798                 UnlockReleaseBuffer(buffer);
2799                 if (have_tuple_lock)
2800                         UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2801                 if (vmbuffer != InvalidBuffer)
2802                         ReleaseBuffer(vmbuffer);
2803                 return result;
2804         }
2805
2806         /*
2807          * We're about to do the actual delete -- check for conflict first, to
2808          * avoid possibly having to roll back work we've just done.
2809          */
2810         CheckForSerializableConflictIn(relation, &tp, buffer);
2811
2812         /* replace cid with a combo cid if necessary */
2813         HeapTupleHeaderAdjustCmax(tp.t_data, &cid, &iscombo);
2814
2815         /*
2816          * Compute replica identity tuple before entering the critical section so
2817          * we don't PANIC upon a memory allocation failure.
2818          */
2819         old_key_tuple = ExtractReplicaIdentity(relation, &tp, true, &old_key_copied);
2820
2821         /*
2822          * If this is the first possibly-multixact-able operation in the current
2823          * transaction, set my per-backend OldestMemberMXactId setting. We can be
2824          * certain that the transaction will never become a member of any older
2825          * MultiXactIds than that.  (We have to do this even if we end up just
2826          * using our own TransactionId below, since some other backend could
2827          * incorporate our XID into a MultiXact immediately afterwards.)
2828          */
2829         MultiXactIdSetOldestMember();
2830
2831         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(tp.t_data),
2832                                                           tp.t_data->t_infomask, tp.t_data->t_infomask2,
2833                                                           xid, LockTupleExclusive, true,
2834                                                           &new_xmax, &new_infomask, &new_infomask2);
2835
2836         START_CRIT_SECTION();
2837
2838         /*
2839          * If this transaction commits, the tuple will become DEAD sooner or
2840          * later.  Set flag that this page is a candidate for pruning once our xid
2841          * falls below the OldestXmin horizon.  If the transaction finally aborts,
2842          * the subsequent page pruning will be a no-op and the hint will be
2843          * cleared.
2844          */
2845         PageSetPrunable(page, xid);
2846
2847         if (PageIsAllVisible(page))
2848         {
2849                 all_visible_cleared = true;
2850                 PageClearAllVisible(page);
2851                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
2852                                                         vmbuffer);
2853         }
2854
2855         /* store transaction information of xact deleting the tuple */
2856         tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
2857         tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
2858         tp.t_data->t_infomask |= new_infomask;
2859         tp.t_data->t_infomask2 |= new_infomask2;
2860         HeapTupleHeaderClearHotUpdated(tp.t_data);
2861         HeapTupleHeaderSetXmax(tp.t_data, new_xmax);
2862         HeapTupleHeaderSetCmax(tp.t_data, cid, iscombo);
2863         /* Make sure there is no forward chain link in t_ctid */
2864         tp.t_data->t_ctid = tp.t_self;
2865
2866         MarkBufferDirty(buffer);
2867
2868         /* XLOG stuff */
2869         if (RelationNeedsWAL(relation))
2870         {
2871                 xl_heap_delete xlrec;
2872                 XLogRecPtr      recptr;
2873
2874                 /* For logical decode we need combocids to properly decode the catalog */
2875                 if (RelationIsAccessibleInLogicalDecoding(relation))
2876                         log_heap_new_cid(relation, &tp);
2877
2878                 xlrec.flags = all_visible_cleared ? XLOG_HEAP_ALL_VISIBLE_CLEARED : 0;
2879                 xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
2880                                                                                           tp.t_data->t_infomask2);
2881                 xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
2882                 xlrec.xmax = new_xmax;
2883
2884                 if (old_key_tuple != NULL)
2885                 {
2886                         if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
2887                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_TUPLE;
2888                         else
2889                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_KEY;
2890                 }
2891
2892                 XLogBeginInsert();
2893                 XLogRegisterData((char *) &xlrec, SizeOfHeapDelete);
2894
2895                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
2896
2897                 /*
2898                  * Log replica identity of the deleted tuple if there is one
2899                  */
2900                 if (old_key_tuple != NULL)
2901                 {
2902                         xl_heap_header xlhdr;
2903
2904                         xlhdr.t_infomask2 = old_key_tuple->t_data->t_infomask2;
2905                         xlhdr.t_infomask = old_key_tuple->t_data->t_infomask;
2906                         xlhdr.t_hoff = old_key_tuple->t_data->t_hoff;
2907
2908                         XLogRegisterData((char *) &xlhdr, SizeOfHeapHeader);
2909                         XLogRegisterData((char *) old_key_tuple->t_data
2910                                                          + offsetof(HeapTupleHeaderData, t_bits),
2911                                                          old_key_tuple->t_len
2912                                                          - offsetof(HeapTupleHeaderData, t_bits));
2913                 }
2914
2915                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
2916
2917                 PageSetLSN(page, recptr);
2918         }
2919
2920         END_CRIT_SECTION();
2921
2922         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2923
2924         if (vmbuffer != InvalidBuffer)
2925                 ReleaseBuffer(vmbuffer);
2926
2927         /*
2928          * If the tuple has toasted out-of-line attributes, we need to delete
2929          * those items too.  We have to do this before releasing the buffer
2930          * because we need to look at the contents of the tuple, but it's OK to
2931          * release the content lock on the buffer first.
2932          */
2933         if (relation->rd_rel->relkind != RELKIND_RELATION &&
2934                 relation->rd_rel->relkind != RELKIND_MATVIEW)
2935         {
2936                 /* toast table entries should never be recursively toasted */
2937                 Assert(!HeapTupleHasExternal(&tp));
2938         }
2939         else if (HeapTupleHasExternal(&tp))
2940                 toast_delete(relation, &tp);
2941
2942         /*
2943          * Mark tuple for invalidation from system caches at next command
2944          * boundary. We have to do this before releasing the buffer because we
2945          * need to look at the contents of the tuple.
2946          */
2947         CacheInvalidateHeapTuple(relation, &tp, NULL);
2948
2949         /* Now we can release the buffer */
2950         ReleaseBuffer(buffer);
2951
2952         /*
2953          * Release the lmgr tuple lock, if we had it.
2954          */
2955         if (have_tuple_lock)
2956                 UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2957
2958         pgstat_count_heap_delete(relation);
2959
2960         if (old_key_tuple != NULL && old_key_copied)
2961                 heap_freetuple(old_key_tuple);
2962
2963         return HeapTupleMayBeUpdated;
2964 }
2965
2966 /*
2967  *      simple_heap_delete - delete a tuple
2968  *
2969  * This routine may be used to delete a tuple when concurrent updates of
2970  * the target tuple are not expected (for example, because we have a lock
2971  * on the relation associated with the tuple).  Any failure is reported
2972  * via ereport().
2973  */
2974 void
2975 simple_heap_delete(Relation relation, ItemPointer tid)
2976 {
2977         HTSU_Result result;
2978         HeapUpdateFailureData hufd;
2979
2980         result = heap_delete(relation, tid,
2981                                                  GetCurrentCommandId(true), InvalidSnapshot,
2982                                                  true /* wait for commit */ ,
2983                                                  &hufd);
2984         switch (result)
2985         {
2986                 case HeapTupleSelfUpdated:
2987                         /* Tuple was already updated in current command? */
2988                         elog(ERROR, "tuple already updated by self");
2989                         break;
2990
2991                 case HeapTupleMayBeUpdated:
2992                         /* done successfully */
2993                         break;
2994
2995                 case HeapTupleUpdated:
2996                         elog(ERROR, "tuple concurrently updated");
2997                         break;
2998
2999                 default:
3000                         elog(ERROR, "unrecognized heap_delete status: %u", result);
3001                         break;
3002         }
3003 }
3004
3005 /*
3006  *      heap_update - replace a tuple
3007  *
3008  * NB: do not call this directly unless you are prepared to deal with
3009  * concurrent-update conditions.  Use simple_heap_update instead.
3010  *
3011  *      relation - table to be modified (caller must hold suitable lock)
3012  *      otid - TID of old tuple to be replaced
3013  *      newtup - newly constructed tuple data to store
3014  *      cid - update command ID (used for visibility test, and stored into
3015  *              cmax/cmin if successful)
3016  *      crosscheck - if not InvalidSnapshot, also check old tuple against this
3017  *      wait - true if should wait for any conflicting update to commit/abort
3018  *      hufd - output parameter, filled in failure cases (see below)
3019  *      lockmode - output parameter, filled with lock mode acquired on tuple
3020  *
3021  * Normal, successful return value is HeapTupleMayBeUpdated, which
3022  * actually means we *did* update it.  Failure return codes are
3023  * HeapTupleSelfUpdated, HeapTupleUpdated, or HeapTupleBeingUpdated
3024  * (the last only possible if wait == false).
3025  *
3026  * On success, the header fields of *newtup are updated to match the new
3027  * stored tuple; in particular, newtup->t_self is set to the TID where the
3028  * new tuple was inserted, and its HEAP_ONLY_TUPLE flag is set iff a HOT
3029  * update was done.  However, any TOAST changes in the new tuple's
3030  * data are not reflected into *newtup.
3031  *
3032  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
3033  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
3034  * (the last only for HeapTupleSelfUpdated, since we
3035  * cannot obtain cmax from a combocid generated by another transaction).
3036  * See comments for struct HeapUpdateFailureData for additional info.
3037  */
3038 HTSU_Result
3039 heap_update(Relation relation, ItemPointer otid, HeapTuple newtup,
3040                         CommandId cid, Snapshot crosscheck, bool wait,
3041                         HeapUpdateFailureData *hufd, LockTupleMode *lockmode)
3042 {
3043         HTSU_Result result;
3044         TransactionId xid = GetCurrentTransactionId();
3045         Bitmapset  *hot_attrs;
3046         Bitmapset  *key_attrs;
3047         Bitmapset  *id_attrs;
3048         ItemId          lp;
3049         HeapTupleData oldtup;
3050         HeapTuple       heaptup;
3051         HeapTuple       old_key_tuple = NULL;
3052         bool            old_key_copied = false;
3053         Page            page;
3054         BlockNumber block;
3055         MultiXactStatus mxact_status;
3056         Buffer          buffer,
3057                                 newbuf,
3058                                 vmbuffer = InvalidBuffer,
3059                                 vmbuffer_new = InvalidBuffer;
3060         bool            need_toast,
3061                                 already_marked;
3062         Size            newtupsize,
3063                                 pagefree;
3064         bool            have_tuple_lock = false;
3065         bool            iscombo;
3066         bool            satisfies_hot;
3067         bool            satisfies_key;
3068         bool            satisfies_id;
3069         bool            use_hot_update = false;
3070         bool            key_intact;
3071         bool            all_visible_cleared = false;
3072         bool            all_visible_cleared_new = false;
3073         bool            checked_lockers;
3074         bool            locker_remains;
3075         TransactionId xmax_new_tuple,
3076                                 xmax_old_tuple;
3077         uint16          infomask_old_tuple,
3078                                 infomask2_old_tuple,
3079                                 infomask_new_tuple,
3080                                 infomask2_new_tuple;
3081
3082         Assert(ItemPointerIsValid(otid));
3083
3084         /*
3085          * Fetch the list of attributes to be checked for HOT update.  This is
3086          * wasted effort if we fail to update or have to put the new tuple on a
3087          * different page.  But we must compute the list before obtaining buffer
3088          * lock --- in the worst case, if we are doing an update on one of the
3089          * relevant system catalogs, we could deadlock if we try to fetch the list
3090          * later.  In any case, the relcache caches the data so this is usually
3091          * pretty cheap.
3092          *
3093          * Note that we get a copy here, so we need not worry about relcache flush
3094          * happening midway through.
3095          */
3096         hot_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_ALL);
3097         key_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_KEY);
3098         id_attrs = RelationGetIndexAttrBitmap(relation,
3099                                                                                   INDEX_ATTR_BITMAP_IDENTITY_KEY);
3100
3101         block = ItemPointerGetBlockNumber(otid);
3102         buffer = ReadBuffer(relation, block);
3103         page = BufferGetPage(buffer);
3104
3105         /*
3106          * Before locking the buffer, pin the visibility map page if it appears to
3107          * be necessary.  Since we haven't got the lock yet, someone else might be
3108          * in the middle of changing this, so we'll need to recheck after we have
3109          * the lock.
3110          */
3111         if (PageIsAllVisible(page))
3112                 visibilitymap_pin(relation, block, &vmbuffer);
3113
3114         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3115
3116         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(otid));
3117         Assert(ItemIdIsNormal(lp));
3118
3119         /*
3120          * Fill in enough data in oldtup for HeapSatisfiesHOTandKeyUpdate to work
3121          * properly.
3122          */
3123         oldtup.t_tableOid = RelationGetRelid(relation);
3124         oldtup.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3125         oldtup.t_len = ItemIdGetLength(lp);
3126         oldtup.t_self = *otid;
3127
3128         /* the new tuple is ready, except for this: */
3129         newtup->t_tableOid = RelationGetRelid(relation);
3130
3131         /* Fill in OID for newtup */
3132         if (relation->rd_rel->relhasoids)
3133         {
3134 #ifdef NOT_USED
3135                 /* this is redundant with an Assert in HeapTupleSetOid */
3136                 Assert(newtup->t_data->t_infomask & HEAP_HASOID);
3137 #endif
3138                 HeapTupleSetOid(newtup, HeapTupleGetOid(&oldtup));
3139         }
3140         else
3141         {
3142                 /* check there is not space for an OID */
3143                 Assert(!(newtup->t_data->t_infomask & HEAP_HASOID));
3144         }
3145
3146         /*
3147          * If we're not updating any "key" column, we can grab a weaker lock type.
3148          * This allows for more concurrency when we are running simultaneously
3149          * with foreign key checks.
3150          *
3151          * Note that if a column gets detoasted while executing the update, but
3152          * the value ends up being the same, this test will fail and we will use
3153          * the stronger lock.  This is acceptable; the important case to optimize
3154          * is updates that don't manipulate key columns, not those that
3155          * serendipitiously arrive at the same key values.
3156          */
3157         HeapSatisfiesHOTandKeyUpdate(relation, hot_attrs, key_attrs, id_attrs,
3158                                                                  &satisfies_hot, &satisfies_key,
3159                                                                  &satisfies_id, &oldtup, newtup);
3160         if (satisfies_key)
3161         {
3162                 *lockmode = LockTupleNoKeyExclusive;
3163                 mxact_status = MultiXactStatusNoKeyUpdate;
3164                 key_intact = true;
3165
3166                 /*
3167                  * If this is the first possibly-multixact-able operation in the
3168                  * current transaction, set my per-backend OldestMemberMXactId
3169                  * setting. We can be certain that the transaction will never become a
3170                  * member of any older MultiXactIds than that.  (We have to do this
3171                  * even if we end up just using our own TransactionId below, since
3172                  * some other backend could incorporate our XID into a MultiXact
3173                  * immediately afterwards.)
3174                  */
3175                 MultiXactIdSetOldestMember();
3176         }
3177         else
3178         {
3179                 *lockmode = LockTupleExclusive;
3180                 mxact_status = MultiXactStatusUpdate;
3181                 key_intact = false;
3182         }
3183
3184         /*
3185          * Note: beyond this point, use oldtup not otid to refer to old tuple.
3186          * otid may very well point at newtup->t_self, which we will overwrite
3187          * with the new tuple's location, so there's great risk of confusion if we
3188          * use otid anymore.
3189          */
3190
3191 l2:
3192         checked_lockers = false;
3193         locker_remains = false;
3194         result = HeapTupleSatisfiesUpdate(&oldtup, cid, buffer);
3195
3196         /* see below about the "no wait" case */
3197         Assert(result != HeapTupleBeingUpdated || wait);
3198
3199         if (result == HeapTupleInvisible)
3200         {
3201                 UnlockReleaseBuffer(buffer);
3202                 elog(ERROR, "attempted to update invisible tuple");
3203         }
3204         else if (result == HeapTupleBeingUpdated && wait)
3205         {
3206                 TransactionId xwait;
3207                 uint16          infomask;
3208                 bool            can_continue = false;
3209
3210                 checked_lockers = true;
3211
3212                 /*
3213                  * XXX note that we don't consider the "no wait" case here.  This
3214                  * isn't a problem currently because no caller uses that case, but it
3215                  * should be fixed if such a caller is introduced.  It wasn't a
3216                  * problem previously because this code would always wait, but now
3217                  * that some tuple locks do not conflict with one of the lock modes we
3218                  * use, it is possible that this case is interesting to handle
3219                  * specially.
3220                  *
3221                  * This may cause failures with third-party code that calls
3222                  * heap_update directly.
3223                  */
3224
3225                 /* must copy state data before unlocking buffer */
3226                 xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3227                 infomask = oldtup.t_data->t_infomask;
3228
3229                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3230
3231                 /*
3232                  * Acquire tuple lock to establish our priority for the tuple (see
3233                  * heap_lock_tuple).  LockTuple will release us when we are
3234                  * next-in-line for the tuple.
3235                  *
3236                  * If we are forced to "start over" below, we keep the tuple lock;
3237                  * this arranges that we stay at the head of the line while rechecking
3238                  * tuple state.
3239                  */
3240                 if (!have_tuple_lock)
3241                 {
3242                         LockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3243                         have_tuple_lock = true;
3244                 }
3245
3246                 /*
3247                  * Now we have to do something about the existing locker.  If it's a
3248                  * multi, sleep on it; we might be awakened before it is completely
3249                  * gone (or even not sleep at all in some cases); we need to preserve
3250                  * it as locker, unless it is gone completely.
3251                  *
3252                  * If it's not a multi, we need to check for sleeping conditions
3253                  * before actually going to sleep.  If the update doesn't conflict
3254                  * with the locks, we just continue without sleeping (but making sure
3255                  * it is preserved).
3256                  */
3257                 if (infomask & HEAP_XMAX_IS_MULTI)
3258                 {
3259                         TransactionId update_xact;
3260                         int                     remain;
3261
3262                         /* wait for multixact */
3263                         MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
3264                                                         relation, &oldtup.t_data->t_ctid, XLTW_Update,
3265                                                         &remain);
3266                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3267
3268                         /*
3269                          * If xwait had just locked the tuple then some other xact could
3270                          * update this tuple before we get to this point.  Check for xmax
3271                          * change, and start over if so.
3272                          */
3273                         if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3274                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3275                                                                          xwait))
3276                                 goto l2;
3277
3278                         /*
3279                          * Note that the multixact may not be done by now.  It could have
3280                          * surviving members; our own xact or other subxacts of this
3281                          * backend, and also any other concurrent transaction that locked
3282                          * the tuple with KeyShare if we only got TupleLockUpdate.  If
3283                          * this is the case, we have to be careful to mark the updated
3284                          * tuple with the surviving members in Xmax.
3285                          *
3286                          * Note that there could have been another update in the
3287                          * MultiXact. In that case, we need to check whether it committed
3288                          * or aborted. If it aborted we are safe to update it again;
3289                          * otherwise there is an update conflict, and we have to return
3290                          * HeapTupleUpdated below.
3291                          *
3292                          * In the LockTupleExclusive case, we still need to preserve the
3293                          * surviving members: those would include the tuple locks we had
3294                          * before this one, which are important to keep in case this
3295                          * subxact aborts.
3296                          */
3297                         update_xact = InvalidTransactionId;
3298                         if (!HEAP_XMAX_IS_LOCKED_ONLY(oldtup.t_data->t_infomask))
3299                                 update_xact = HeapTupleGetUpdateXid(oldtup.t_data);
3300
3301                         /*
3302                          * There was no UPDATE in the MultiXact; or it aborted. No
3303                          * TransactionIdIsInProgress() call needed here, since we called
3304                          * MultiXactIdWait() above.
3305                          */
3306                         if (!TransactionIdIsValid(update_xact) ||
3307                                 TransactionIdDidAbort(update_xact))
3308                                 can_continue = true;
3309
3310                         locker_remains = remain != 0;
3311                 }
3312                 else
3313                 {
3314                         /*
3315                          * If it's just a key-share locker, and we're not changing the key
3316                          * columns, we don't need to wait for it to end; but we need to
3317                          * preserve it as locker.
3318                          */
3319                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) && key_intact)
3320                         {
3321                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3322
3323                                 /*
3324                                  * recheck the locker; if someone else changed the tuple while
3325                                  * we weren't looking, start over.
3326                                  */
3327                                 if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3328                                         !TransactionIdEquals(
3329                                                                         HeapTupleHeaderGetRawXmax(oldtup.t_data),
3330                                                                                  xwait))
3331                                         goto l2;
3332
3333                                 can_continue = true;
3334                                 locker_remains = true;
3335                         }
3336                         else
3337                         {
3338                                 /* wait for regular transaction to end */
3339                                 XactLockTableWait(xwait, relation, &oldtup.t_data->t_ctid,
3340                                                                   XLTW_Update);
3341                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3342
3343                                 /*
3344                                  * xwait is done, but if xwait had just locked the tuple then
3345                                  * some other xact could update this tuple before we get to
3346                                  * this point. Check for xmax change, and start over if so.
3347                                  */
3348                                 if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3349                                         !TransactionIdEquals(
3350                                                                         HeapTupleHeaderGetRawXmax(oldtup.t_data),
3351                                                                                  xwait))
3352                                         goto l2;
3353
3354                                 /* Otherwise check if it committed or aborted */
3355                                 UpdateXmaxHintBits(oldtup.t_data, buffer, xwait);
3356                                 if (oldtup.t_data->t_infomask & HEAP_XMAX_INVALID)
3357                                         can_continue = true;
3358                         }
3359                 }
3360
3361                 result = can_continue ? HeapTupleMayBeUpdated : HeapTupleUpdated;
3362         }
3363
3364         if (crosscheck != InvalidSnapshot && result == HeapTupleMayBeUpdated)
3365         {
3366                 /* Perform additional check for transaction-snapshot mode RI updates */
3367                 if (!HeapTupleSatisfiesVisibility(&oldtup, crosscheck, buffer))
3368                         result = HeapTupleUpdated;
3369         }
3370
3371         if (result != HeapTupleMayBeUpdated)
3372         {
3373                 Assert(result == HeapTupleSelfUpdated ||
3374                            result == HeapTupleUpdated ||
3375                            result == HeapTupleBeingUpdated);
3376                 Assert(!(oldtup.t_data->t_infomask & HEAP_XMAX_INVALID));
3377                 hufd->ctid = oldtup.t_data->t_ctid;
3378                 hufd->xmax = HeapTupleHeaderGetUpdateXid(oldtup.t_data);
3379                 if (result == HeapTupleSelfUpdated)
3380                         hufd->cmax = HeapTupleHeaderGetCmax(oldtup.t_data);
3381                 else
3382                         hufd->cmax = InvalidCommandId;
3383                 UnlockReleaseBuffer(buffer);
3384                 if (have_tuple_lock)
3385                         UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3386                 if (vmbuffer != InvalidBuffer)
3387                         ReleaseBuffer(vmbuffer);
3388                 bms_free(hot_attrs);
3389                 bms_free(key_attrs);
3390                 return result;
3391         }
3392
3393         /*
3394          * If we didn't pin the visibility map page and the page has become all
3395          * visible while we were busy locking the buffer, or during some
3396          * subsequent window during which we had it unlocked, we'll have to unlock
3397          * and re-lock, to avoid holding the buffer lock across an I/O.  That's a
3398          * bit unfortunate, especially since we'll now have to recheck whether the
3399          * tuple has been locked or updated under us, but hopefully it won't
3400          * happen very often.
3401          */
3402         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3403         {
3404                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3405                 visibilitymap_pin(relation, block, &vmbuffer);
3406                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3407                 goto l2;
3408         }
3409
3410         /*
3411          * We're about to do the actual update -- check for conflict first, to
3412          * avoid possibly having to roll back work we've just done.
3413          */
3414         CheckForSerializableConflictIn(relation, &oldtup, buffer);
3415
3416         /* Fill in transaction status data */
3417
3418         /*
3419          * If the tuple we're updating is locked, we need to preserve the locking
3420          * info in the old tuple's Xmax.  Prepare a new Xmax value for this.
3421          */
3422         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3423                                                           oldtup.t_data->t_infomask,
3424                                                           oldtup.t_data->t_infomask2,
3425                                                           xid, *lockmode, true,
3426                                                           &xmax_old_tuple, &infomask_old_tuple,
3427                                                           &infomask2_old_tuple);
3428
3429         /*
3430          * And also prepare an Xmax value for the new copy of the tuple.  If there
3431          * was no xmax previously, or there was one but all lockers are now gone,
3432          * then use InvalidXid; otherwise, get the xmax from the old tuple.  (In
3433          * rare cases that might also be InvalidXid and yet not have the
3434          * HEAP_XMAX_INVALID bit set; that's fine.)
3435          */
3436         if ((oldtup.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3437                 (checked_lockers && !locker_remains))
3438                 xmax_new_tuple = InvalidTransactionId;
3439         else
3440                 xmax_new_tuple = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3441
3442         if (!TransactionIdIsValid(xmax_new_tuple))
3443         {
3444                 infomask_new_tuple = HEAP_XMAX_INVALID;
3445                 infomask2_new_tuple = 0;
3446         }
3447         else
3448         {
3449                 /*
3450                  * If we found a valid Xmax for the new tuple, then the infomask bits
3451                  * to use on the new tuple depend on what was there on the old one.
3452                  * Note that since we're doing an update, the only possibility is that
3453                  * the lockers had FOR KEY SHARE lock.
3454                  */
3455                 if (oldtup.t_data->t_infomask & HEAP_XMAX_IS_MULTI)
3456                 {
3457                         GetMultiXactIdHintBits(xmax_new_tuple, &infomask_new_tuple,
3458                                                                    &infomask2_new_tuple);
3459                 }
3460                 else
3461                 {
3462                         infomask_new_tuple = HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_LOCK_ONLY;
3463                         infomask2_new_tuple = 0;
3464                 }
3465         }
3466
3467         /*
3468          * Prepare the new tuple with the appropriate initial values of Xmin and
3469          * Xmax, as well as initial infomask bits as computed above.
3470          */
3471         newtup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
3472         newtup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
3473         HeapTupleHeaderSetXmin(newtup->t_data, xid);
3474         HeapTupleHeaderSetCmin(newtup->t_data, cid);
3475         newtup->t_data->t_infomask |= HEAP_UPDATED | infomask_new_tuple;
3476         newtup->t_data->t_infomask2 |= infomask2_new_tuple;
3477         HeapTupleHeaderSetXmax(newtup->t_data, xmax_new_tuple);
3478
3479         /*
3480          * Replace cid with a combo cid if necessary.  Note that we already put
3481          * the plain cid into the new tuple.
3482          */
3483         HeapTupleHeaderAdjustCmax(oldtup.t_data, &cid, &iscombo);
3484
3485         /*
3486          * If the toaster needs to be activated, OR if the new tuple will not fit
3487          * on the same page as the old, then we need to release the content lock
3488          * (but not the pin!) on the old tuple's buffer while we are off doing
3489          * TOAST and/or table-file-extension work.  We must mark the old tuple to
3490          * show that it's already being updated, else other processes may try to
3491          * update it themselves.
3492          *
3493          * We need to invoke the toaster if there are already any out-of-line
3494          * toasted values present, or if the new tuple is over-threshold.
3495          */
3496         if (relation->rd_rel->relkind != RELKIND_RELATION &&
3497                 relation->rd_rel->relkind != RELKIND_MATVIEW)
3498         {
3499                 /* toast table entries should never be recursively toasted */
3500                 Assert(!HeapTupleHasExternal(&oldtup));
3501                 Assert(!HeapTupleHasExternal(newtup));
3502                 need_toast = false;
3503         }
3504         else
3505                 need_toast = (HeapTupleHasExternal(&oldtup) ||
3506                                           HeapTupleHasExternal(newtup) ||
3507                                           newtup->t_len > TOAST_TUPLE_THRESHOLD);
3508
3509         pagefree = PageGetHeapFreeSpace(page);
3510
3511         newtupsize = MAXALIGN(newtup->t_len);
3512
3513         if (need_toast || newtupsize > pagefree)
3514         {
3515                 /* Clear obsolete visibility flags ... */
3516                 oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3517                 oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3518                 HeapTupleClearHotUpdated(&oldtup);
3519                 /* ... and store info about transaction updating this tuple */
3520                 Assert(TransactionIdIsValid(xmax_old_tuple));
3521                 HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
3522                 oldtup.t_data->t_infomask |= infomask_old_tuple;
3523                 oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
3524                 HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3525                 /* temporarily make it look not-updated */
3526                 oldtup.t_data->t_ctid = oldtup.t_self;
3527                 already_marked = true;
3528                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3529
3530                 /*
3531                  * Let the toaster do its thing, if needed.
3532                  *
3533                  * Note: below this point, heaptup is the data we actually intend to
3534                  * store into the relation; newtup is the caller's original untoasted
3535                  * data.
3536                  */
3537                 if (need_toast)
3538                 {
3539                         /* Note we always use WAL and FSM during updates */
3540                         heaptup = toast_insert_or_update(relation, newtup, &oldtup, 0);
3541                         newtupsize = MAXALIGN(heaptup->t_len);
3542                 }
3543                 else
3544                         heaptup = newtup;
3545
3546                 /*
3547                  * Now, do we need a new page for the tuple, or not?  This is a bit
3548                  * tricky since someone else could have added tuples to the page while
3549                  * we weren't looking.  We have to recheck the available space after
3550                  * reacquiring the buffer lock.  But don't bother to do that if the
3551                  * former amount of free space is still not enough; it's unlikely
3552                  * there's more free now than before.
3553                  *
3554                  * What's more, if we need to get a new page, we will need to acquire
3555                  * buffer locks on both old and new pages.  To avoid deadlock against
3556                  * some other backend trying to get the same two locks in the other
3557                  * order, we must be consistent about the order we get the locks in.
3558                  * We use the rule "lock the lower-numbered page of the relation
3559                  * first".  To implement this, we must do RelationGetBufferForTuple
3560                  * while not holding the lock on the old page, and we must rely on it
3561                  * to get the locks on both pages in the correct order.
3562                  */
3563                 if (newtupsize > pagefree)
3564                 {
3565                         /* Assume there's no chance to put heaptup on same page. */
3566                         newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3567                                                                                            buffer, 0, NULL,
3568                                                                                            &vmbuffer_new, &vmbuffer);
3569                 }
3570                 else
3571                 {
3572                         /* Re-acquire the lock on the old tuple's page. */
3573                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3574                         /* Re-check using the up-to-date free space */
3575                         pagefree = PageGetHeapFreeSpace(page);
3576                         if (newtupsize > pagefree)
3577                         {
3578                                 /*
3579                                  * Rats, it doesn't fit anymore.  We must now unlock and
3580                                  * relock to avoid deadlock.  Fortunately, this path should
3581                                  * seldom be taken.
3582                                  */
3583                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3584                                 newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3585                                                                                                    buffer, 0, NULL,
3586                                                                                                    &vmbuffer_new, &vmbuffer);
3587                         }
3588                         else
3589                         {
3590                                 /* OK, it fits here, so we're done. */
3591                                 newbuf = buffer;
3592                         }
3593                 }
3594         }
3595         else
3596         {
3597                 /* No TOAST work needed, and it'll fit on same page */
3598                 already_marked = false;
3599                 newbuf = buffer;
3600                 heaptup = newtup;
3601         }
3602
3603         /*
3604          * We're about to create the new tuple -- check for conflict first, to
3605          * avoid possibly having to roll back work we've just done.
3606          *
3607          * NOTE: For a tuple insert, we only need to check for table locks, since
3608          * predicate locking at the index level will cover ranges for anything
3609          * except a table scan.  Therefore, only provide the relation.
3610          */
3611         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
3612
3613         /*
3614          * At this point newbuf and buffer are both pinned and locked, and newbuf
3615          * has enough space for the new tuple.  If they are the same buffer, only
3616          * one pin is held.
3617          */
3618
3619         if (newbuf == buffer)
3620         {
3621                 /*
3622                  * Since the new tuple is going into the same page, we might be able
3623                  * to do a HOT update.  Check if any of the index columns have been
3624                  * changed.  If not, then HOT update is possible.
3625                  */
3626                 if (satisfies_hot)
3627                         use_hot_update = true;
3628         }
3629         else
3630         {
3631                 /* Set a hint that the old page could use prune/defrag */
3632                 PageSetFull(page);
3633         }
3634
3635         /*
3636          * Compute replica identity tuple before entering the critical section so
3637          * we don't PANIC upon a memory allocation failure.
3638          * ExtractReplicaIdentity() will return NULL if nothing needs to be
3639          * logged.
3640          */
3641         old_key_tuple = ExtractReplicaIdentity(relation, &oldtup, !satisfies_id, &old_key_copied);
3642
3643         /* NO EREPORT(ERROR) from here till changes are logged */
3644         START_CRIT_SECTION();
3645
3646         /*
3647          * If this transaction commits, the old tuple will become DEAD sooner or
3648          * later.  Set flag that this page is a candidate for pruning once our xid
3649          * falls below the OldestXmin horizon.  If the transaction finally aborts,
3650          * the subsequent page pruning will be a no-op and the hint will be
3651          * cleared.
3652          *
3653          * XXX Should we set hint on newbuf as well?  If the transaction aborts,
3654          * there would be a prunable tuple in the newbuf; but for now we choose
3655          * not to optimize for aborts.  Note that heap_xlog_update must be kept in
3656          * sync if this decision changes.
3657          */
3658         PageSetPrunable(page, xid);
3659
3660         if (use_hot_update)
3661         {
3662                 /* Mark the old tuple as HOT-updated */
3663                 HeapTupleSetHotUpdated(&oldtup);
3664                 /* And mark the new tuple as heap-only */
3665                 HeapTupleSetHeapOnly(heaptup);
3666                 /* Mark the caller's copy too, in case different from heaptup */
3667                 HeapTupleSetHeapOnly(newtup);
3668         }
3669         else
3670         {
3671                 /* Make sure tuples are correctly marked as not-HOT */
3672                 HeapTupleClearHotUpdated(&oldtup);
3673                 HeapTupleClearHeapOnly(heaptup);
3674                 HeapTupleClearHeapOnly(newtup);
3675         }
3676
3677         RelationPutHeapTuple(relation, newbuf, heaptup);        /* insert new tuple */
3678
3679         if (!already_marked)
3680         {
3681                 /* Clear obsolete visibility flags ... */
3682                 oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3683                 oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3684                 /* ... and store info about transaction updating this tuple */
3685                 Assert(TransactionIdIsValid(xmax_old_tuple));
3686                 HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
3687                 oldtup.t_data->t_infomask |= infomask_old_tuple;
3688                 oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
3689                 HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3690         }
3691
3692         /* record address of new tuple in t_ctid of old one */
3693         oldtup.t_data->t_ctid = heaptup->t_self;
3694
3695         /* clear PD_ALL_VISIBLE flags */
3696         if (PageIsAllVisible(BufferGetPage(buffer)))
3697         {
3698                 all_visible_cleared = true;
3699                 PageClearAllVisible(BufferGetPage(buffer));
3700                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
3701                                                         vmbuffer);
3702         }
3703         if (newbuf != buffer && PageIsAllVisible(BufferGetPage(newbuf)))
3704         {
3705                 all_visible_cleared_new = true;
3706                 PageClearAllVisible(BufferGetPage(newbuf));
3707                 visibilitymap_clear(relation, BufferGetBlockNumber(newbuf),
3708                                                         vmbuffer_new);
3709         }
3710
3711         if (newbuf != buffer)
3712                 MarkBufferDirty(newbuf);
3713         MarkBufferDirty(buffer);
3714
3715         /* XLOG stuff */
3716         if (RelationNeedsWAL(relation))
3717         {
3718                 XLogRecPtr      recptr;
3719
3720                 /*
3721                  * For logical decoding we need combocids to properly decode the
3722                  * catalog.
3723                  */
3724                 if (RelationIsAccessibleInLogicalDecoding(relation))
3725                 {
3726                         log_heap_new_cid(relation, &oldtup);
3727                         log_heap_new_cid(relation, heaptup);
3728                 }
3729
3730                 recptr = log_heap_update(relation, buffer,
3731                                                                  newbuf, &oldtup, heaptup,
3732                                                                  old_key_tuple,
3733                                                                  all_visible_cleared,
3734                                                                  all_visible_cleared_new);
3735                 if (newbuf != buffer)
3736                 {
3737                         PageSetLSN(BufferGetPage(newbuf), recptr);
3738                 }
3739                 PageSetLSN(BufferGetPage(buffer), recptr);
3740         }
3741
3742         END_CRIT_SECTION();
3743
3744         if (newbuf != buffer)
3745                 LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
3746         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3747
3748         /*
3749          * Mark old tuple for invalidation from system caches at next command
3750          * boundary, and mark the new tuple for invalidation in case we abort. We
3751          * have to do this before releasing the buffer because oldtup is in the
3752          * buffer.  (heaptup is all in local memory, but it's necessary to process
3753          * both tuple versions in one call to inval.c so we can avoid redundant
3754          * sinval messages.)
3755          */
3756         CacheInvalidateHeapTuple(relation, &oldtup, heaptup);
3757
3758         /* Now we can release the buffer(s) */
3759         if (newbuf != buffer)
3760                 ReleaseBuffer(newbuf);
3761         ReleaseBuffer(buffer);
3762         if (BufferIsValid(vmbuffer_new))
3763                 ReleaseBuffer(vmbuffer_new);
3764         if (BufferIsValid(vmbuffer))
3765                 ReleaseBuffer(vmbuffer);
3766
3767         /*
3768          * Release the lmgr tuple lock, if we had it.
3769          */
3770         if (have_tuple_lock)
3771                 UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3772
3773         pgstat_count_heap_update(relation, use_hot_update);
3774
3775         /*
3776          * If heaptup is a private copy, release it.  Don't forget to copy t_self
3777          * back to the caller's image, too.
3778          */
3779         if (heaptup != newtup)
3780         {
3781                 newtup->t_self = heaptup->t_self;
3782                 heap_freetuple(heaptup);
3783         }
3784
3785         if (old_key_tuple != NULL && old_key_copied)
3786                 heap_freetuple(old_key_tuple);
3787
3788         bms_free(hot_attrs);
3789         bms_free(key_attrs);
3790
3791         return HeapTupleMayBeUpdated;
3792 }
3793
3794 /*
3795  * Check if the specified attribute's value is same in both given tuples.
3796  * Subroutine for HeapSatisfiesHOTandKeyUpdate.
3797  */
3798 static bool
3799 heap_tuple_attr_equals(TupleDesc tupdesc, int attrnum,
3800                                            HeapTuple tup1, HeapTuple tup2)
3801 {
3802         Datum           value1,
3803                                 value2;
3804         bool            isnull1,
3805                                 isnull2;
3806         Form_pg_attribute att;
3807
3808         /*
3809          * If it's a whole-tuple reference, say "not equal".  It's not really
3810          * worth supporting this case, since it could only succeed after a no-op
3811          * update, which is hardly a case worth optimizing for.
3812          */
3813         if (attrnum == 0)
3814                 return false;
3815
3816         /*
3817          * Likewise, automatically say "not equal" for any system attribute other
3818          * than OID and tableOID; we cannot expect these to be consistent in a HOT
3819          * chain, or even to be set correctly yet in the new tuple.
3820          */
3821         if (attrnum < 0)
3822         {
3823                 if (attrnum != ObjectIdAttributeNumber &&
3824                         attrnum != TableOidAttributeNumber)
3825                         return false;
3826         }
3827
3828         /*
3829          * Extract the corresponding values.  XXX this is pretty inefficient if
3830          * there are many indexed columns.  Should HeapSatisfiesHOTandKeyUpdate do
3831          * a single heap_deform_tuple call on each tuple, instead?      But that
3832          * doesn't work for system columns ...
3833          */
3834         value1 = heap_getattr(tup1, attrnum, tupdesc, &isnull1);
3835         value2 = heap_getattr(tup2, attrnum, tupdesc, &isnull2);
3836
3837         /*
3838          * If one value is NULL and other is not, then they are certainly not
3839          * equal
3840          */
3841         if (isnull1 != isnull2)
3842                 return false;
3843
3844         /*
3845          * If both are NULL, they can be considered equal.
3846          */
3847         if (isnull1)
3848                 return true;
3849
3850         /*
3851          * We do simple binary comparison of the two datums.  This may be overly
3852          * strict because there can be multiple binary representations for the
3853          * same logical value.  But we should be OK as long as there are no false
3854          * positives.  Using a type-specific equality operator is messy because
3855          * there could be multiple notions of equality in different operator
3856          * classes; furthermore, we cannot safely invoke user-defined functions
3857          * while holding exclusive buffer lock.
3858          */
3859         if (attrnum <= 0)
3860         {
3861                 /* The only allowed system columns are OIDs, so do this */
3862                 return (DatumGetObjectId(value1) == DatumGetObjectId(value2));
3863         }
3864         else
3865         {
3866                 Assert(attrnum <= tupdesc->natts);
3867                 att = tupdesc->attrs[attrnum - 1];
3868                 return datumIsEqual(value1, value2, att->attbyval, att->attlen);
3869         }
3870 }
3871
3872 /*
3873  * Check which columns are being updated.
3874  *
3875  * This simultaneously checks conditions for HOT updates, for FOR KEY
3876  * SHARE updates, and REPLICA IDENTITY concerns.  Since much of the time they
3877  * will be checking very similar sets of columns, and doing the same tests on
3878  * them, it makes sense to optimize and do them together.
3879  *
3880  * We receive three bitmapsets comprising the three sets of columns we're
3881  * interested in.  Note these are destructively modified; that is OK since
3882  * this is invoked at most once in heap_update.
3883  *
3884  * hot_result is set to TRUE if it's okay to do a HOT update (i.e. it does not
3885  * modified indexed columns); key_result is set to TRUE if the update does not
3886  * modify columns used in the key; id_result is set to TRUE if the update does
3887  * not modify columns in any index marked as the REPLICA IDENTITY.
3888  */
3889 static void
3890 HeapSatisfiesHOTandKeyUpdate(Relation relation, Bitmapset *hot_attrs,
3891                                                          Bitmapset *key_attrs, Bitmapset *id_attrs,
3892                                                          bool *satisfies_hot, bool *satisfies_key,
3893                                                          bool *satisfies_id,
3894                                                          HeapTuple oldtup, HeapTuple newtup)
3895 {
3896         int                     next_hot_attnum;
3897         int                     next_key_attnum;
3898         int                     next_id_attnum;
3899         bool            hot_result = true;
3900         bool            key_result = true;
3901         bool            id_result = true;
3902
3903         /* If REPLICA IDENTITY is set to FULL, id_attrs will be empty. */
3904         Assert(bms_is_subset(id_attrs, key_attrs));
3905         Assert(bms_is_subset(key_attrs, hot_attrs));
3906
3907         /*
3908          * If one of these sets contains no remaining bits, bms_first_member will
3909          * return -1, and after adding FirstLowInvalidHeapAttributeNumber (which
3910          * is negative!)  we'll get an attribute number that can't possibly be
3911          * real, and thus won't match any actual attribute number.
3912          */
3913         next_hot_attnum = bms_first_member(hot_attrs);
3914         next_hot_attnum += FirstLowInvalidHeapAttributeNumber;
3915         next_key_attnum = bms_first_member(key_attrs);
3916         next_key_attnum += FirstLowInvalidHeapAttributeNumber;
3917         next_id_attnum = bms_first_member(id_attrs);
3918         next_id_attnum += FirstLowInvalidHeapAttributeNumber;
3919
3920         for (;;)
3921         {
3922                 bool            changed;
3923                 int                     check_now;
3924
3925                 /*
3926                  * Since the HOT attributes are a superset of the key attributes and
3927                  * the key attributes are a superset of the id attributes, this logic
3928                  * is guaranteed to identify the next column that needs to be checked.
3929                  */
3930                 if (hot_result && next_hot_attnum > FirstLowInvalidHeapAttributeNumber)
3931                         check_now = next_hot_attnum;
3932                 else if (key_result && next_key_attnum > FirstLowInvalidHeapAttributeNumber)
3933                         check_now = next_key_attnum;
3934                 else if (id_result && next_id_attnum > FirstLowInvalidHeapAttributeNumber)
3935                         check_now = next_id_attnum;
3936                 else
3937                         break;
3938
3939                 /* See whether it changed. */
3940                 changed = !heap_tuple_attr_equals(RelationGetDescr(relation),
3941                                                                                   check_now, oldtup, newtup);
3942                 if (changed)
3943                 {
3944                         if (check_now == next_hot_attnum)
3945                                 hot_result = false;
3946                         if (check_now == next_key_attnum)
3947                                 key_result = false;
3948                         if (check_now == next_id_attnum)
3949                                 id_result = false;
3950
3951                         /* if all are false now, we can stop checking */
3952                         if (!hot_result && !key_result && !id_result)
3953                                 break;
3954                 }
3955
3956                 /*
3957                  * Advance the next attribute numbers for the sets that contain the
3958                  * attribute we just checked.  As we work our way through the columns,
3959                  * the next_attnum values will rise; but when each set becomes empty,
3960                  * bms_first_member() will return -1 and the attribute number will end
3961                  * up with a value less than FirstLowInvalidHeapAttributeNumber.
3962                  */
3963                 if (hot_result && check_now == next_hot_attnum)
3964                 {
3965                         next_hot_attnum = bms_first_member(hot_attrs);
3966                         next_hot_attnum += FirstLowInvalidHeapAttributeNumber;
3967                 }
3968                 if (key_result && check_now == next_key_attnum)
3969                 {
3970                         next_key_attnum = bms_first_member(key_attrs);
3971                         next_key_attnum += FirstLowInvalidHeapAttributeNumber;
3972                 }
3973                 if (id_result && check_now == next_id_attnum)
3974                 {
3975                         next_id_attnum = bms_first_member(id_attrs);
3976                         next_id_attnum += FirstLowInvalidHeapAttributeNumber;
3977                 }
3978         }
3979
3980         *satisfies_hot = hot_result;
3981         *satisfies_key = key_result;
3982         *satisfies_id = id_result;
3983 }
3984
3985 /*
3986  *      simple_heap_update - replace a tuple
3987  *
3988  * This routine may be used to update a tuple when concurrent updates of
3989  * the target tuple are not expected (for example, because we have a lock
3990  * on the relation associated with the tuple).  Any failure is reported
3991  * via ereport().
3992  */
3993 void
3994 simple_heap_update(Relation relation, ItemPointer otid, HeapTuple tup)
3995 {
3996         HTSU_Result result;
3997         HeapUpdateFailureData hufd;
3998         LockTupleMode lockmode;
3999
4000         result = heap_update(relation, otid, tup,
4001                                                  GetCurrentCommandId(true), InvalidSnapshot,
4002                                                  true /* wait for commit */ ,
4003                                                  &hufd, &lockmode);
4004         switch (result)
4005         {
4006                 case HeapTupleSelfUpdated:
4007                         /* Tuple was already updated in current command? */
4008                         elog(ERROR, "tuple already updated by self");
4009                         break;
4010
4011                 case HeapTupleMayBeUpdated:
4012                         /* done successfully */
4013                         break;
4014
4015                 case HeapTupleUpdated:
4016                         elog(ERROR, "tuple concurrently updated");
4017                         break;
4018
4019                 default:
4020                         elog(ERROR, "unrecognized heap_update status: %u", result);
4021                         break;
4022         }
4023 }
4024
4025
4026 /*
4027  * Return the MultiXactStatus corresponding to the given tuple lock mode.
4028  */
4029 static MultiXactStatus
4030 get_mxact_status_for_lock(LockTupleMode mode, bool is_update)
4031 {
4032         int                     retval;
4033
4034         if (is_update)
4035                 retval = tupleLockExtraInfo[mode].updstatus;
4036         else
4037                 retval = tupleLockExtraInfo[mode].lockstatus;
4038
4039         if (retval == -1)
4040                 elog(ERROR, "invalid lock tuple mode %d/%s", mode,
4041                          is_update ? "true" : "false");
4042
4043         return (MultiXactStatus) retval;
4044 }
4045
4046
4047 /*
4048  *      heap_lock_tuple - lock a tuple in shared or exclusive mode
4049  *
4050  * Note that this acquires a buffer pin, which the caller must release.
4051  *
4052  * Input parameters:
4053  *      relation: relation containing tuple (caller must hold suitable lock)
4054  *      tuple->t_self: TID of tuple to lock (rest of struct need not be valid)
4055  *      cid: current command ID (used for visibility test, and stored into
4056  *              tuple's cmax if lock is successful)
4057  *      mode: indicates if shared or exclusive tuple lock is desired
4058  *      wait_policy: what to do if tuple lock is not available
4059  *      follow_updates: if true, follow the update chain to also lock descendant
4060  *              tuples.
4061  *
4062  * Output parameters:
4063  *      *tuple: all fields filled in
4064  *      *buffer: set to buffer holding tuple (pinned but not locked at exit)
4065  *      *hufd: filled in failure cases (see below)
4066  *
4067  * Function result may be:
4068  *      HeapTupleMayBeUpdated: lock was successfully acquired
4069  *      HeapTupleSelfUpdated: lock failed because tuple updated by self
4070  *      HeapTupleUpdated: lock failed because tuple updated by other xact
4071  *      HeapTupleWouldBlock: lock couldn't be acquired and wait_policy is skip
4072  *
4073  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
4074  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
4075  * (the last only for HeapTupleSelfUpdated, since we
4076  * cannot obtain cmax from a combocid generated by another transaction).
4077  * See comments for struct HeapUpdateFailureData for additional info.
4078  *
4079  * See README.tuplock for a thorough explanation of this mechanism.
4080  */
4081 HTSU_Result
4082 heap_lock_tuple(Relation relation, HeapTuple tuple,
4083                                 CommandId cid, LockTupleMode mode, LockWaitPolicy wait_policy,
4084                                 bool follow_updates,
4085                                 Buffer *buffer, HeapUpdateFailureData *hufd)
4086 {
4087         HTSU_Result result;
4088         ItemPointer tid = &(tuple->t_self);
4089         ItemId          lp;
4090         Page            page;
4091         TransactionId xid,
4092                                 xmax;
4093         uint16          old_infomask,
4094                                 new_infomask,
4095                                 new_infomask2;
4096         bool            have_tuple_lock = false;
4097
4098         *buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
4099         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4100
4101         page = BufferGetPage(*buffer);
4102         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
4103         Assert(ItemIdIsNormal(lp));
4104
4105         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
4106         tuple->t_len = ItemIdGetLength(lp);
4107         tuple->t_tableOid = RelationGetRelid(relation);
4108
4109 l3:
4110         result = HeapTupleSatisfiesUpdate(tuple, cid, *buffer);
4111
4112         if (result == HeapTupleInvisible)
4113         {
4114                 UnlockReleaseBuffer(*buffer);
4115                 elog(ERROR, "attempted to lock invisible tuple");
4116         }
4117         else if (result == HeapTupleBeingUpdated)
4118         {
4119                 TransactionId xwait;
4120                 uint16          infomask;
4121                 uint16          infomask2;
4122                 bool            require_sleep;
4123                 ItemPointerData t_ctid;
4124
4125                 /* must copy state data before unlocking buffer */
4126                 xwait = HeapTupleHeaderGetRawXmax(tuple->t_data);
4127                 infomask = tuple->t_data->t_infomask;
4128                 infomask2 = tuple->t_data->t_infomask2;
4129                 ItemPointerCopy(&tuple->t_data->t_ctid, &t_ctid);
4130
4131                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4132
4133                 /*
4134                  * If any subtransaction of the current top transaction already holds
4135                  * a lock as strong or stronger than what we're requesting, we
4136                  * effectively hold the desired lock already.  We *must* succeed
4137                  * without trying to take the tuple lock, else we will deadlock
4138                  * against anyone wanting to acquire a stronger lock.
4139                  */
4140                 if (infomask & HEAP_XMAX_IS_MULTI)
4141                 {
4142                         int                     i;
4143                         int                     nmembers;
4144                         MultiXactMember *members;
4145
4146                         /*
4147                          * We don't need to allow old multixacts here; if that had been
4148                          * the case, HeapTupleSatisfiesUpdate would have returned
4149                          * MayBeUpdated and we wouldn't be here.
4150                          */
4151                         nmembers =
4152                                 GetMultiXactIdMembers(xwait, &members, false,
4153                                                                           HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4154
4155                         for (i = 0; i < nmembers; i++)
4156                         {
4157                                 if (TransactionIdIsCurrentTransactionId(members[i].xid))
4158                                 {
4159                                         LockTupleMode membermode;
4160
4161                                         membermode = TUPLOCK_from_mxstatus(members[i].status);
4162
4163                                         if (membermode >= mode)
4164                                         {
4165                                                 if (have_tuple_lock)
4166                                                         UnlockTupleTuplock(relation, tid, mode);
4167
4168                                                 pfree(members);
4169                                                 return HeapTupleMayBeUpdated;
4170                                         }
4171                                 }
4172                         }
4173
4174                         if (members)
4175                                 pfree(members);
4176                 }
4177
4178                 /*
4179                  * Acquire tuple lock to establish our priority for the tuple.
4180                  * LockTuple will release us when we are next-in-line for the tuple.
4181                  * We must do this even if we are share-locking.
4182                  *
4183                  * If we are forced to "start over" below, we keep the tuple lock;
4184                  * this arranges that we stay at the head of the line while rechecking
4185                  * tuple state.
4186                  */
4187                 if (!have_tuple_lock)
4188                 {
4189                         switch (wait_policy)
4190                         {
4191                                 case LockWaitBlock:
4192                                         LockTupleTuplock(relation, tid, mode);
4193                                         break;
4194                                 case LockWaitSkip:
4195                                         if (!ConditionalLockTupleTuplock(relation, tid, mode))
4196                                         {
4197                                                 result = HeapTupleWouldBlock;
4198                                                 /* recovery code expects to have buffer lock held */
4199                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4200                                                 goto failed;
4201                                         }
4202                                         break;
4203                                 case LockWaitError:
4204                                         if (!ConditionalLockTupleTuplock(relation, tid, mode))
4205                                                 ereport(ERROR,
4206                                                                 (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4207                                                                  errmsg("could not obtain lock on row in relation \"%s\"",
4208                                                                                 RelationGetRelationName(relation))));
4209                                         break;
4210                         }
4211                         have_tuple_lock = true;
4212                 }
4213
4214                 /*
4215                  * Initially assume that we will have to wait for the locking
4216                  * transaction(s) to finish.  We check various cases below in which
4217                  * this can be turned off.
4218                  */
4219                 require_sleep = true;
4220                 if (mode == LockTupleKeyShare)
4221                 {
4222                         /*
4223                          * If we're requesting KeyShare, and there's no update present, we
4224                          * don't need to wait.  Even if there is an update, we can still
4225                          * continue if the key hasn't been modified.
4226                          *
4227                          * However, if there are updates, we need to walk the update chain
4228                          * to mark future versions of the row as locked, too.  That way,
4229                          * if somebody deletes that future version, we're protected
4230                          * against the key going away.  This locking of future versions
4231                          * could block momentarily, if a concurrent transaction is
4232                          * deleting a key; or it could return a value to the effect that
4233                          * the transaction deleting the key has already committed.  So we
4234                          * do this before re-locking the buffer; otherwise this would be
4235                          * prone to deadlocks.
4236                          *
4237                          * Note that the TID we're locking was grabbed before we unlocked
4238                          * the buffer.  For it to change while we're not looking, the
4239                          * other properties we're testing for below after re-locking the
4240                          * buffer would also change, in which case we would restart this
4241                          * loop above.
4242                          */
4243                         if (!(infomask2 & HEAP_KEYS_UPDATED))
4244                         {
4245                                 bool            updated;
4246
4247                                 updated = !HEAP_XMAX_IS_LOCKED_ONLY(infomask);
4248
4249                                 /*
4250                                  * If there are updates, follow the update chain; bail out if
4251                                  * that cannot be done.
4252                                  */
4253                                 if (follow_updates && updated)
4254                                 {
4255                                         HTSU_Result res;
4256
4257                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4258                                                                                                   GetCurrentTransactionId(),
4259                                                                                                   mode);
4260                                         if (res != HeapTupleMayBeUpdated)
4261                                         {
4262                                                 result = res;
4263                                                 /* recovery code expects to have buffer lock held */
4264                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4265                                                 goto failed;
4266                                         }
4267                                 }
4268
4269                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4270
4271                                 /*
4272                                  * Make sure it's still an appropriate lock, else start over.
4273                                  * Also, if it wasn't updated before we released the lock, but
4274                                  * is updated now, we start over too; the reason is that we
4275                                  * now need to follow the update chain to lock the new
4276                                  * versions.
4277                                  */
4278                                 if (!HeapTupleHeaderIsOnlyLocked(tuple->t_data) &&
4279                                         ((tuple->t_data->t_infomask2 & HEAP_KEYS_UPDATED) ||
4280                                          !updated))
4281                                         goto l3;
4282
4283                                 /* Things look okay, so we can skip sleeping */
4284                                 require_sleep = false;
4285
4286                                 /*
4287                                  * Note we allow Xmax to change here; other updaters/lockers
4288                                  * could have modified it before we grabbed the buffer lock.
4289                                  * However, this is not a problem, because with the recheck we
4290                                  * just did we ensure that they still don't conflict with the
4291                                  * lock we want.
4292                                  */
4293                         }
4294                 }
4295                 else if (mode == LockTupleShare)
4296                 {
4297                         /*
4298                          * If we're requesting Share, we can similarly avoid sleeping if
4299                          * there's no update and no exclusive lock present.
4300                          */
4301                         if (HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
4302                                 !HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4303                         {
4304                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4305
4306                                 /*
4307                                  * Make sure it's still an appropriate lock, else start over.
4308                                  * See above about allowing xmax to change.
4309                                  */
4310                                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4311                                         HEAP_XMAX_IS_EXCL_LOCKED(tuple->t_data->t_infomask))
4312                                         goto l3;
4313                                 require_sleep = false;
4314                         }
4315                 }
4316                 else if (mode == LockTupleNoKeyExclusive)
4317                 {
4318                         /*
4319                          * If we're requesting NoKeyExclusive, we might also be able to
4320                          * avoid sleeping; just ensure that there's no other lock type
4321                          * than KeyShare.  Note that this is a bit more involved than just
4322                          * checking hint bits -- we need to expand the multixact to figure
4323                          * out lock modes for each one (unless there was only one such
4324                          * locker).
4325                          */
4326                         if (infomask & HEAP_XMAX_IS_MULTI)
4327                         {
4328                                 int                     nmembers;
4329                                 MultiXactMember *members;
4330
4331                                 /*
4332                                  * We don't need to allow old multixacts here; if that had
4333                                  * been the case, HeapTupleSatisfiesUpdate would have returned
4334                                  * MayBeUpdated and we wouldn't be here.
4335                                  */
4336                                 nmembers =
4337                                         GetMultiXactIdMembers(xwait, &members, false,
4338                                                                                   HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4339
4340                                 if (nmembers <= 0)
4341                                 {
4342                                         /*
4343                                          * No need to keep the previous xmax here. This is
4344                                          * unlikely to happen.
4345                                          */
4346                                         require_sleep = false;
4347                                 }
4348                                 else
4349                                 {
4350                                         int                     i;
4351                                         bool            allowed = true;
4352
4353                                         for (i = 0; i < nmembers; i++)
4354                                         {
4355                                                 if (members[i].status != MultiXactStatusForKeyShare)
4356                                                 {
4357                                                         allowed = false;
4358                                                         break;
4359                                                 }
4360                                         }
4361                                         if (allowed)
4362                                         {
4363                                                 /*
4364                                                  * if the xmax changed under us in the meantime, start
4365                                                  * over.
4366                                                  */
4367                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4368                                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4369                                                         !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4370                                                                                                  xwait))
4371                                                 {
4372                                                         pfree(members);
4373                                                         goto l3;
4374                                                 }
4375                                                 /* otherwise, we're good */
4376                                                 require_sleep = false;
4377                                         }
4378
4379                                         pfree(members);
4380                                 }
4381                         }
4382                         else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
4383                         {
4384                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4385
4386                                 /* if the xmax changed in the meantime, start over */
4387                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4388                                         !TransactionIdEquals(
4389                                                                         HeapTupleHeaderGetRawXmax(tuple->t_data),
4390                                                                                  xwait))
4391                                         goto l3;
4392                                 /* otherwise, we're good */
4393                                 require_sleep = false;
4394                         }
4395                 }
4396
4397                 /*
4398                  * By here, we either have already acquired the buffer exclusive lock,
4399                  * or we must wait for the locking transaction or multixact; so below
4400                  * we ensure that we grab buffer lock after the sleep.
4401                  */
4402
4403                 if (require_sleep)
4404                 {
4405                         if (infomask & HEAP_XMAX_IS_MULTI)
4406                         {
4407                                 MultiXactStatus status = get_mxact_status_for_lock(mode, false);
4408
4409                                 /* We only ever lock tuples, never update them */
4410                                 if (status >= MultiXactStatusNoKeyUpdate)
4411                                         elog(ERROR, "invalid lock mode in heap_lock_tuple");
4412
4413                                 /* wait for multixact to end, or die trying  */
4414                                 switch (wait_policy)
4415                                 {
4416                                         case LockWaitBlock:
4417                                                 MultiXactIdWait((MultiXactId) xwait, status, infomask,
4418                                                                                 relation, &tuple->t_data->t_ctid, XLTW_Lock, NULL);
4419                                                 break;
4420                                         case LockWaitSkip:
4421                                                 if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4422                                                                                                                 status, infomask, relation,
4423                                                                                                                 NULL))
4424                                                 {
4425                                                         result = HeapTupleWouldBlock;
4426                                                         /* recovery code expects to have buffer lock held */
4427                                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4428                                                         goto failed;
4429                                                 }
4430                                                 break;
4431                                         case LockWaitError:
4432                                                 if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4433                                                                                                                 status, infomask, relation,
4434                                                                                                                 NULL))
4435                                                         ereport(ERROR,
4436                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4437                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
4438                                                                                         RelationGetRelationName(relation))));
4439
4440                                                 break;
4441                                 }
4442
4443                                 /* if there are updates, follow the update chain */
4444                                 if (follow_updates &&
4445                                         !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
4446                                 {
4447                                         HTSU_Result res;
4448
4449                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4450                                                                                                   GetCurrentTransactionId(),
4451                                                                                                   mode);
4452                                         if (res != HeapTupleMayBeUpdated)
4453                                         {
4454                                                 result = res;
4455                                                 /* recovery code expects to have buffer lock held */
4456                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4457                                                 goto failed;
4458                                         }
4459                                 }
4460
4461                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4462
4463                                 /*
4464                                  * If xwait had just locked the tuple then some other xact
4465                                  * could update this tuple before we get to this point. Check
4466                                  * for xmax change, and start over if so.
4467                                  */
4468                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4469                                         !TransactionIdEquals(
4470                                                                         HeapTupleHeaderGetRawXmax(tuple->t_data),
4471                                                                                  xwait))
4472                                         goto l3;
4473
4474                                 /*
4475                                  * Of course, the multixact might not be done here: if we're
4476                                  * requesting a light lock mode, other transactions with light
4477                                  * locks could still be alive, as well as locks owned by our
4478                                  * own xact or other subxacts of this backend.  We need to
4479                                  * preserve the surviving MultiXact members.  Note that it
4480                                  * isn't absolutely necessary in the latter case, but doing so
4481                                  * is simpler.
4482                                  */
4483                         }
4484                         else
4485                         {
4486                                 /* wait for regular transaction to end, or die trying */
4487                                 switch (wait_policy)
4488                                 {
4489                                         case LockWaitBlock:
4490                                                 XactLockTableWait(xwait, relation, &tuple->t_data->t_ctid,
4491                                                                                   XLTW_Lock);
4492                                                 break;
4493                                         case LockWaitSkip:
4494                                                 if (!ConditionalXactLockTableWait(xwait))
4495                                                 {
4496                                                         result = HeapTupleWouldBlock;
4497                                                         /* recovery code expects to have buffer lock held */
4498                                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4499                                                         goto failed;
4500                                                 }
4501                                                 break;
4502                                         case LockWaitError:
4503                                                 if (!ConditionalXactLockTableWait(xwait))
4504                                                         ereport(ERROR,
4505                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4506                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
4507                                                                                         RelationGetRelationName(relation))));
4508                                                 break;
4509                                 }
4510
4511                                 /* if there are updates, follow the update chain */
4512                                 if (follow_updates &&
4513                                         !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
4514                                 {
4515                                         HTSU_Result res;
4516
4517                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4518                                                                                                   GetCurrentTransactionId(),
4519                                                                                                   mode);
4520                                         if (res != HeapTupleMayBeUpdated)
4521                                         {
4522                                                 result = res;
4523                                                 /* recovery code expects to have buffer lock held */
4524                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4525                                                 goto failed;
4526                                         }
4527                                 }
4528
4529                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4530
4531                                 /*
4532                                  * xwait is done, but if xwait had just locked the tuple then
4533                                  * some other xact could update this tuple before we get to
4534                                  * this point.  Check for xmax change, and start over if so.
4535                                  */
4536                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4537                                         !TransactionIdEquals(
4538                                                                         HeapTupleHeaderGetRawXmax(tuple->t_data),
4539                                                                                  xwait))
4540                                         goto l3;
4541
4542                                 /*
4543                                  * Otherwise check if it committed or aborted.  Note we cannot
4544                                  * be here if the tuple was only locked by somebody who didn't
4545                                  * conflict with us; that should have been handled above.  So
4546                                  * that transaction must necessarily be gone by now.
4547                                  */
4548                                 UpdateXmaxHintBits(tuple->t_data, *buffer, xwait);
4549                         }
4550                 }
4551
4552                 /* By here, we're certain that we hold buffer exclusive lock again */
4553
4554                 /*
4555                  * We may lock if previous xmax aborted, or if it committed but only
4556                  * locked the tuple without updating it; or if we didn't have to wait
4557                  * at all for whatever reason.
4558                  */
4559                 if (!require_sleep ||
4560                         (tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
4561                         HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4562                         HeapTupleHeaderIsOnlyLocked(tuple->t_data))
4563                         result = HeapTupleMayBeUpdated;
4564                 else
4565                         result = HeapTupleUpdated;
4566         }
4567
4568 failed:
4569         if (result != HeapTupleMayBeUpdated)
4570         {
4571                 Assert(result == HeapTupleSelfUpdated || result == HeapTupleUpdated ||
4572                            result == HeapTupleWouldBlock);
4573                 Assert(!(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));
4574                 hufd->ctid = tuple->t_data->t_ctid;
4575                 hufd->xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
4576                 if (result == HeapTupleSelfUpdated)
4577                         hufd->cmax = HeapTupleHeaderGetCmax(tuple->t_data);
4578                 else
4579                         hufd->cmax = InvalidCommandId;
4580                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4581                 if (have_tuple_lock)
4582                         UnlockTupleTuplock(relation, tid, mode);
4583                 return result;
4584         }
4585
4586         xmax = HeapTupleHeaderGetRawXmax(tuple->t_data);
4587         old_infomask = tuple->t_data->t_infomask;
4588
4589         /*
4590          * We might already hold the desired lock (or stronger), possibly under a
4591          * different subtransaction of the current top transaction.  If so, there
4592          * is no need to change state or issue a WAL record.  We already handled
4593          * the case where this is true for xmax being a MultiXactId, so now check
4594          * for cases where it is a plain TransactionId.
4595          *
4596          * Note in particular that this covers the case where we already hold
4597          * exclusive lock on the tuple and the caller only wants key share or
4598          * share lock. It would certainly not do to give up the exclusive lock.
4599          */
4600         if (!(old_infomask & (HEAP_XMAX_INVALID |
4601                                                   HEAP_XMAX_COMMITTED |
4602                                                   HEAP_XMAX_IS_MULTI)) &&
4603                 (mode == LockTupleKeyShare ?
4604                  (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask) ||
4605                   HEAP_XMAX_IS_SHR_LOCKED(old_infomask) ||
4606                   HEAP_XMAX_IS_EXCL_LOCKED(old_infomask)) :
4607                  mode == LockTupleShare ?
4608                  (HEAP_XMAX_IS_SHR_LOCKED(old_infomask) ||
4609                   HEAP_XMAX_IS_EXCL_LOCKED(old_infomask)) :
4610                  (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))) &&
4611                 TransactionIdIsCurrentTransactionId(xmax))
4612         {
4613                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4614                 /* Probably can't hold tuple lock here, but may as well check */
4615                 if (have_tuple_lock)
4616                         UnlockTupleTuplock(relation, tid, mode);
4617                 return HeapTupleMayBeUpdated;
4618         }
4619
4620         /*
4621          * If this is the first possibly-multixact-able operation in the current
4622          * transaction, set my per-backend OldestMemberMXactId setting. We can be
4623          * certain that the transaction will never become a member of any older
4624          * MultiXactIds than that.  (We have to do this even if we end up just
4625          * using our own TransactionId below, since some other backend could
4626          * incorporate our XID into a MultiXact immediately afterwards.)
4627          */
4628         MultiXactIdSetOldestMember();
4629
4630         /*
4631          * Compute the new xmax and infomask to store into the tuple.  Note we do
4632          * not modify the tuple just yet, because that would leave it in the wrong
4633          * state if multixact.c elogs.
4634          */
4635         compute_new_xmax_infomask(xmax, old_infomask, tuple->t_data->t_infomask2,
4636                                                           GetCurrentTransactionId(), mode, false,
4637                                                           &xid, &new_infomask, &new_infomask2);
4638
4639         START_CRIT_SECTION();
4640
4641         /*
4642          * Store transaction information of xact locking the tuple.
4643          *
4644          * Note: Cmax is meaningless in this context, so don't set it; this avoids
4645          * possibly generating a useless combo CID.  Moreover, if we're locking a
4646          * previously updated tuple, it's important to preserve the Cmax.
4647          *
4648          * Also reset the HOT UPDATE bit, but only if there's no update; otherwise
4649          * we would break the HOT chain.
4650          */
4651         tuple->t_data->t_infomask &= ~HEAP_XMAX_BITS;
4652         tuple->t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
4653         tuple->t_data->t_infomask |= new_infomask;
4654         tuple->t_data->t_infomask2 |= new_infomask2;
4655         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
4656                 HeapTupleHeaderClearHotUpdated(tuple->t_data);
4657         HeapTupleHeaderSetXmax(tuple->t_data, xid);
4658
4659         /*
4660          * Make sure there is no forward chain link in t_ctid.  Note that in the
4661          * cases where the tuple has been updated, we must not overwrite t_ctid,
4662          * because it was set by the updater.  Moreover, if the tuple has been
4663          * updated, we need to follow the update chain to lock the new versions of
4664          * the tuple as well.
4665          */
4666         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
4667                 tuple->t_data->t_ctid = *tid;
4668
4669         MarkBufferDirty(*buffer);
4670
4671         /*
4672          * XLOG stuff.  You might think that we don't need an XLOG record because
4673          * there is no state change worth restoring after a crash.  You would be
4674          * wrong however: we have just written either a TransactionId or a
4675          * MultiXactId that may never have been seen on disk before, and we need
4676          * to make sure that there are XLOG entries covering those ID numbers.
4677          * Else the same IDs might be re-used after a crash, which would be
4678          * disastrous if this page made it to disk before the crash.  Essentially
4679          * we have to enforce the WAL log-before-data rule even in this case.
4680          * (Also, in a PITR log-shipping or 2PC environment, we have to have XLOG
4681          * entries for everything anyway.)
4682          */
4683         if (RelationNeedsWAL(relation))
4684         {
4685                 xl_heap_lock xlrec;
4686                 XLogRecPtr      recptr;
4687
4688                 XLogBeginInsert();
4689                 XLogRegisterBuffer(0, *buffer, REGBUF_STANDARD);
4690
4691                 xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
4692                 xlrec.locking_xid = xid;
4693                 xlrec.infobits_set = compute_infobits(new_infomask,
4694                                                                                           tuple->t_data->t_infomask2);
4695                 XLogRegisterData((char *) &xlrec, SizeOfHeapLock);
4696
4697                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
4698
4699                 PageSetLSN(page, recptr);
4700         }
4701
4702         END_CRIT_SECTION();
4703
4704         LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4705
4706         /*
4707          * Don't update the visibility map here. Locking a tuple doesn't change
4708          * visibility info.
4709          */
4710
4711         /*
4712          * Now that we have successfully marked the tuple as locked, we can
4713          * release the lmgr tuple lock, if we had it.
4714          */
4715         if (have_tuple_lock)
4716                 UnlockTupleTuplock(relation, tid, mode);
4717
4718         return HeapTupleMayBeUpdated;
4719 }
4720
4721
4722 /*
4723  * Given an original set of Xmax and infomask, and a transaction (identified by
4724  * add_to_xmax) acquiring a new lock of some mode, compute the new Xmax and
4725  * corresponding infomasks to use on the tuple.
4726  *
4727  * Note that this might have side effects such as creating a new MultiXactId.
4728  *
4729  * Most callers will have called HeapTupleSatisfiesUpdate before this function;
4730  * that will have set the HEAP_XMAX_INVALID bit if the xmax was a MultiXactId
4731  * but it was not running anymore. There is a race condition, which is that the
4732  * MultiXactId may have finished since then, but that uncommon case is handled
4733  * either here, or within MultiXactIdExpand.
4734  *
4735  * There is a similar race condition possible when the old xmax was a regular
4736  * TransactionId.  We test TransactionIdIsInProgress again just to narrow the
4737  * window, but it's still possible to end up creating an unnecessary
4738  * MultiXactId.  Fortunately this is harmless.
4739  */
4740 static void
4741 compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
4742                                                   uint16 old_infomask2, TransactionId add_to_xmax,
4743                                                   LockTupleMode mode, bool is_update,
4744                                                   TransactionId *result_xmax, uint16 *result_infomask,
4745                                                   uint16 *result_infomask2)
4746 {
4747         TransactionId new_xmax;
4748         uint16          new_infomask,
4749                                 new_infomask2;
4750
4751         Assert(TransactionIdIsCurrentTransactionId(add_to_xmax));
4752
4753 l5:
4754         new_infomask = 0;
4755         new_infomask2 = 0;
4756         if (old_infomask & HEAP_XMAX_INVALID)
4757         {
4758                 /*
4759                  * No previous locker; we just insert our own TransactionId.
4760                  *
4761                  * Note that it's critical that this case be the first one checked,
4762                  * because there are several blocks below that come back to this one
4763                  * to implement certain optimizations; old_infomask might contain
4764                  * other dirty bits in those cases, but we don't really care.
4765                  */
4766                 if (is_update)
4767                 {
4768                         new_xmax = add_to_xmax;
4769                         if (mode == LockTupleExclusive)
4770                                 new_infomask2 |= HEAP_KEYS_UPDATED;
4771                 }
4772                 else
4773                 {
4774                         new_infomask |= HEAP_XMAX_LOCK_ONLY;
4775                         switch (mode)
4776                         {
4777                                 case LockTupleKeyShare:
4778                                         new_xmax = add_to_xmax;
4779                                         new_infomask |= HEAP_XMAX_KEYSHR_LOCK;
4780                                         break;
4781                                 case LockTupleShare:
4782                                         new_xmax = add_to_xmax;
4783                                         new_infomask |= HEAP_XMAX_SHR_LOCK;
4784                                         break;
4785                                 case LockTupleNoKeyExclusive:
4786                                         new_xmax = add_to_xmax;
4787                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
4788                                         break;
4789                                 case LockTupleExclusive:
4790                                         new_xmax = add_to_xmax;
4791                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
4792                                         new_infomask2 |= HEAP_KEYS_UPDATED;
4793                                         break;
4794                                 default:
4795                                         new_xmax = InvalidTransactionId;        /* silence compiler */
4796                                         elog(ERROR, "invalid lock mode");
4797                         }
4798                 }
4799         }
4800         else if (old_infomask & HEAP_XMAX_IS_MULTI)
4801         {
4802                 MultiXactStatus new_status;
4803
4804                 /*
4805                  * Currently we don't allow XMAX_COMMITTED to be set for multis, so
4806                  * cross-check.
4807                  */
4808                 Assert(!(old_infomask & HEAP_XMAX_COMMITTED));
4809
4810                 /*
4811                  * A multixact together with LOCK_ONLY set but neither lock bit set
4812                  * (i.e. a pg_upgraded share locked tuple) cannot possibly be running
4813                  * anymore.  This check is critical for databases upgraded by
4814                  * pg_upgrade; both MultiXactIdIsRunning and MultiXactIdExpand assume
4815                  * that such multis are never passed.
4816                  */
4817                 if (!(old_infomask & HEAP_LOCK_MASK) &&
4818                         HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
4819                 {
4820                         old_infomask &= ~HEAP_XMAX_IS_MULTI;
4821                         old_infomask |= HEAP_XMAX_INVALID;
4822                         goto l5;
4823                 }
4824
4825                 /*
4826                  * If the XMAX is already a MultiXactId, then we need to expand it to
4827                  * include add_to_xmax; but if all the members were lockers and are
4828                  * all gone, we can do away with the IS_MULTI bit and just set
4829                  * add_to_xmax as the only locker/updater.  If all lockers are gone
4830                  * and we have an updater that aborted, we can also do without a
4831                  * multi.
4832                  *
4833                  * The cost of doing GetMultiXactIdMembers would be paid by
4834                  * MultiXactIdExpand if we weren't to do this, so this check is not
4835                  * incurring extra work anyhow.
4836                  */
4837                 if (!MultiXactIdIsRunning(xmax, HEAP_XMAX_IS_LOCKED_ONLY(old_infomask)))
4838                 {
4839                         if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) ||
4840                                 TransactionIdDidAbort(MultiXactIdGetUpdateXid(xmax,
4841                                                                                                                           old_infomask)))
4842                         {
4843                                 /*
4844                                  * Reset these bits and restart; otherwise fall through to
4845                                  * create a new multi below.
4846                                  */
4847                                 old_infomask &= ~HEAP_XMAX_IS_MULTI;
4848                                 old_infomask |= HEAP_XMAX_INVALID;
4849                                 goto l5;
4850                         }
4851                 }
4852
4853                 new_status = get_mxact_status_for_lock(mode, is_update);
4854
4855                 new_xmax = MultiXactIdExpand((MultiXactId) xmax, add_to_xmax,
4856                                                                          new_status);
4857                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4858         }
4859         else if (old_infomask & HEAP_XMAX_COMMITTED)
4860         {
4861                 /*
4862                  * It's a committed update, so we need to preserve him as updater of
4863                  * the tuple.
4864                  */
4865                 MultiXactStatus status;
4866                 MultiXactStatus new_status;
4867
4868                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4869                         status = MultiXactStatusUpdate;
4870                 else
4871                         status = MultiXactStatusNoKeyUpdate;
4872
4873                 new_status = get_mxact_status_for_lock(mode, is_update);
4874
4875                 /*
4876                  * since it's not running, it's obviously impossible for the old
4877                  * updater to be identical to the current one, so we need not check
4878                  * for that case as we do in the block above.
4879                  */
4880                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
4881                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4882         }
4883         else if (TransactionIdIsInProgress(xmax))
4884         {
4885                 /*
4886                  * If the XMAX is a valid, in-progress TransactionId, then we need to
4887                  * create a new MultiXactId that includes both the old locker or
4888                  * updater and our own TransactionId.
4889                  */
4890                 MultiXactStatus new_status;
4891                 MultiXactStatus old_status;
4892                 LockTupleMode old_mode;
4893
4894                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
4895                 {
4896                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
4897                                 old_status = MultiXactStatusForKeyShare;
4898                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
4899                                 old_status = MultiXactStatusForShare;
4900                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
4901                         {
4902                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4903                                         old_status = MultiXactStatusForUpdate;
4904                                 else
4905                                         old_status = MultiXactStatusForNoKeyUpdate;
4906                         }
4907                         else
4908                         {
4909                                 /*
4910                                  * LOCK_ONLY can be present alone only when a page has been
4911                                  * upgraded by pg_upgrade.  But in that case,
4912                                  * TransactionIdIsInProgress() should have returned false.  We
4913                                  * assume it's no longer locked in this case.
4914                                  */
4915                                 elog(WARNING, "LOCK_ONLY found for Xid in progress %u", xmax);
4916                                 old_infomask |= HEAP_XMAX_INVALID;
4917                                 old_infomask &= ~HEAP_XMAX_LOCK_ONLY;
4918                                 goto l5;
4919                         }
4920                 }
4921                 else
4922                 {
4923                         /* it's an update, but which kind? */
4924                         if (old_infomask2 & HEAP_KEYS_UPDATED)
4925                                 old_status = MultiXactStatusUpdate;
4926                         else
4927                                 old_status = MultiXactStatusNoKeyUpdate;
4928                 }
4929
4930                 old_mode = TUPLOCK_from_mxstatus(old_status);
4931
4932                 /*
4933                  * If the lock to be acquired is for the same TransactionId as the
4934                  * existing lock, there's an optimization possible: consider only the
4935                  * strongest of both locks as the only one present, and restart.
4936                  */
4937                 if (xmax == add_to_xmax)
4938                 {
4939                         /*
4940                          * Note that it's not possible for the original tuple to be
4941                          * updated: we wouldn't be here because the tuple would have been
4942                          * invisible and we wouldn't try to update it.  As a subtlety,
4943                          * this code can also run when traversing an update chain to lock
4944                          * future versions of a tuple.  But we wouldn't be here either,
4945                          * because the add_to_xmax would be different from the original
4946                          * updater.
4947                          */
4948                         Assert(HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
4949
4950                         /* acquire the strongest of both */
4951                         if (mode < old_mode)
4952                                 mode = old_mode;
4953                         /* mustn't touch is_update */
4954
4955                         old_infomask |= HEAP_XMAX_INVALID;
4956                         goto l5;
4957                 }
4958
4959                 /* otherwise, just fall back to creating a new multixact */
4960                 new_status = get_mxact_status_for_lock(mode, is_update);
4961                 new_xmax = MultiXactIdCreate(xmax, old_status,
4962                                                                          add_to_xmax, new_status);
4963                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4964         }
4965         else if (!HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) &&
4966                          TransactionIdDidCommit(xmax))
4967         {
4968                 /*
4969                  * It's a committed update, so we gotta preserve him as updater of the
4970                  * tuple.
4971                  */
4972                 MultiXactStatus status;
4973                 MultiXactStatus new_status;
4974
4975                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4976                         status = MultiXactStatusUpdate;
4977                 else
4978                         status = MultiXactStatusNoKeyUpdate;
4979
4980                 new_status = get_mxact_status_for_lock(mode, is_update);
4981
4982                 /*
4983                  * since it's not running, it's obviously impossible for the old
4984                  * updater to be identical to the current one, so we need not check
4985                  * for that case as we do in the block above.
4986                  */
4987                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
4988                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4989         }
4990         else
4991         {
4992                 /*
4993                  * Can get here iff the locking/updating transaction was running when
4994                  * the infomask was extracted from the tuple, but finished before
4995                  * TransactionIdIsInProgress got to run.  Deal with it as if there was
4996                  * no locker at all in the first place.
4997                  */
4998                 old_infomask |= HEAP_XMAX_INVALID;
4999                 goto l5;
5000         }
5001
5002         *result_infomask = new_infomask;
5003         *result_infomask2 = new_infomask2;
5004         *result_xmax = new_xmax;
5005 }
5006
5007 /*
5008  * Subroutine for heap_lock_updated_tuple_rec.
5009  *
5010  * Given an hypothetical multixact status held by the transaction identified
5011  * with the given xid, does the current transaction need to wait, fail, or can
5012  * it continue if it wanted to acquire a lock of the given mode?  "needwait"
5013  * is set to true if waiting is necessary; if it can continue, then
5014  * HeapTupleMayBeUpdated is returned.  In case of a conflict, a different
5015  * HeapTupleSatisfiesUpdate return code is returned.
5016  *
5017  * The held status is said to be hypothetical because it might correspond to a
5018  * lock held by a single Xid, i.e. not a real MultiXactId; we express it this
5019  * way for simplicity of API.
5020  */
5021 static HTSU_Result
5022 test_lockmode_for_conflict(MultiXactStatus status, TransactionId xid,
5023                                                    LockTupleMode mode, bool *needwait)
5024 {
5025         MultiXactStatus wantedstatus;
5026
5027         *needwait = false;
5028         wantedstatus = get_mxact_status_for_lock(mode, false);
5029
5030         /*
5031          * Note: we *must* check TransactionIdIsInProgress before
5032          * TransactionIdDidAbort/Commit; see comment at top of tqual.c for an
5033          * explanation.
5034          */
5035         if (TransactionIdIsCurrentTransactionId(xid))
5036         {
5037                 /*
5038                  * Updated by our own transaction? Just return failure.  This
5039                  * shouldn't normally happen.
5040                  */
5041                 return HeapTupleSelfUpdated;
5042         }
5043         else if (TransactionIdIsInProgress(xid))
5044         {
5045                 /*
5046                  * If the locking transaction is running, what we do depends on
5047                  * whether the lock modes conflict: if they do, then we must wait for
5048                  * it to finish; otherwise we can fall through to lock this tuple
5049                  * version without waiting.
5050                  */
5051                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5052                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5053                 {
5054                         *needwait = true;
5055                 }
5056
5057                 /*
5058                  * If we set needwait above, then this value doesn't matter;
5059                  * otherwise, this value signals to caller that it's okay to proceed.
5060                  */
5061                 return HeapTupleMayBeUpdated;
5062         }
5063         else if (TransactionIdDidAbort(xid))
5064                 return HeapTupleMayBeUpdated;
5065         else if (TransactionIdDidCommit(xid))
5066         {
5067                 /*
5068                  * The other transaction committed.  If it was only a locker, then the
5069                  * lock is completely gone now and we can return success; but if it
5070                  * was an update, then what we do depends on whether the two lock
5071                  * modes conflict.  If they conflict, then we must report error to
5072                  * caller. But if they don't, we can fall through to allow the current
5073                  * transaction to lock the tuple.
5074                  *
5075                  * Note: the reason we worry about ISUPDATE here is because as soon as
5076                  * a transaction ends, all its locks are gone and meaningless, and
5077                  * thus we can ignore them; whereas its updates persist.  In the
5078                  * TransactionIdIsInProgress case, above, we don't need to check
5079                  * because we know the lock is still "alive" and thus a conflict needs
5080                  * always be checked.
5081                  */
5082                 if (!ISUPDATE_from_mxstatus(status))
5083                         return HeapTupleMayBeUpdated;
5084
5085                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5086                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5087                         /* bummer */
5088                         return HeapTupleUpdated;
5089
5090                 return HeapTupleMayBeUpdated;
5091         }
5092
5093         /* Not in progress, not aborted, not committed -- must have crashed */
5094         return HeapTupleMayBeUpdated;
5095 }
5096
5097
5098 /*
5099  * Recursive part of heap_lock_updated_tuple
5100  *
5101  * Fetch the tuple pointed to by tid in rel, and mark it as locked by the given
5102  * xid with the given mode; if this tuple is updated, recurse to lock the new
5103  * version as well.
5104  */
5105 static HTSU_Result
5106 heap_lock_updated_tuple_rec(Relation rel, ItemPointer tid, TransactionId xid,
5107                                                         LockTupleMode mode)
5108 {
5109         ItemPointerData tupid;
5110         HeapTupleData mytup;
5111         Buffer          buf;
5112         uint16          new_infomask,
5113                                 new_infomask2,
5114                                 old_infomask,
5115                                 old_infomask2;
5116         TransactionId xmax,
5117                                 new_xmax;
5118         TransactionId priorXmax = InvalidTransactionId;
5119
5120         ItemPointerCopy(tid, &tupid);
5121
5122         for (;;)
5123         {
5124                 new_infomask = 0;
5125                 new_xmax = InvalidTransactionId;
5126                 ItemPointerCopy(&tupid, &(mytup.t_self));
5127
5128                 if (!heap_fetch(rel, SnapshotAny, &mytup, &buf, false, NULL))
5129                 {
5130                         /*
5131                          * if we fail to find the updated version of the tuple, it's
5132                          * because it was vacuumed/pruned away after its creator
5133                          * transaction aborted.  So behave as if we got to the end of the
5134                          * chain, and there's no further tuple to lock: return success to
5135                          * caller.
5136                          */
5137                         return HeapTupleMayBeUpdated;
5138                 }
5139
5140 l4:
5141                 CHECK_FOR_INTERRUPTS();
5142                 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5143
5144                 /*
5145                  * Check the tuple XMIN against prior XMAX, if any.  If we reached the
5146                  * end of the chain, we're done, so return success.
5147                  */
5148                 if (TransactionIdIsValid(priorXmax) &&
5149                         !TransactionIdEquals(HeapTupleHeaderGetXmin(mytup.t_data),
5150                                                                  priorXmax))
5151                 {
5152                         UnlockReleaseBuffer(buf);
5153                         return HeapTupleMayBeUpdated;
5154                 }
5155
5156                 old_infomask = mytup.t_data->t_infomask;
5157                 old_infomask2 = mytup.t_data->t_infomask2;
5158                 xmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5159
5160                 /*
5161                  * If this tuple version has been updated or locked by some concurrent
5162                  * transaction(s), what we do depends on whether our lock mode
5163                  * conflicts with what those other transactions hold, and also on the
5164                  * status of them.
5165                  */
5166                 if (!(old_infomask & HEAP_XMAX_INVALID))
5167                 {
5168                         TransactionId rawxmax;
5169                         bool            needwait;
5170
5171                         rawxmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5172                         if (old_infomask & HEAP_XMAX_IS_MULTI)
5173                         {
5174                                 int                     nmembers;
5175                                 int                     i;
5176                                 MultiXactMember *members;
5177
5178                                 nmembers = GetMultiXactIdMembers(rawxmax, &members, false,
5179                                                                          HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5180                                 for (i = 0; i < nmembers; i++)
5181                                 {
5182                                         HTSU_Result res;
5183
5184                                         res = test_lockmode_for_conflict(members[i].status,
5185                                                                                                          members[i].xid,
5186                                                                                                          mode, &needwait);
5187
5188                                         if (needwait)
5189                                         {
5190                                                 LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5191                                                 XactLockTableWait(members[i].xid, rel,
5192                                                                                   &mytup.t_data->t_ctid,
5193                                                                                   XLTW_LockUpdated);
5194                                                 pfree(members);
5195                                                 goto l4;
5196                                         }
5197                                         if (res != HeapTupleMayBeUpdated)
5198                                         {
5199                                                 UnlockReleaseBuffer(buf);
5200                                                 pfree(members);
5201                                                 return res;
5202                                         }
5203                                 }
5204                                 if (members)
5205                                         pfree(members);
5206                         }
5207                         else
5208                         {
5209                                 HTSU_Result res;
5210                                 MultiXactStatus status;
5211
5212                                 /*
5213                                  * For a non-multi Xmax, we first need to compute the
5214                                  * corresponding MultiXactStatus by using the infomask bits.
5215                                  */
5216                                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5217                                 {
5218                                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5219                                                 status = MultiXactStatusForKeyShare;
5220                                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5221                                                 status = MultiXactStatusForShare;
5222                                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5223                                         {
5224                                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
5225                                                         status = MultiXactStatusForUpdate;
5226                                                 else
5227                                                         status = MultiXactStatusForNoKeyUpdate;
5228                                         }
5229                                         else
5230                                         {
5231                                                 /*
5232                                                  * LOCK_ONLY present alone (a pg_upgraded tuple marked
5233                                                  * as share-locked in the old cluster) shouldn't be
5234                                                  * seen in the middle of an update chain.
5235                                                  */
5236                                                 elog(ERROR, "invalid lock status in tuple");
5237                                         }
5238                                 }
5239                                 else
5240                                 {
5241                                         /* it's an update, but which kind? */
5242                                         if (old_infomask2 & HEAP_KEYS_UPDATED)
5243                                                 status = MultiXactStatusUpdate;
5244                                         else
5245                                                 status = MultiXactStatusNoKeyUpdate;
5246                                 }
5247
5248                                 res = test_lockmode_for_conflict(status, rawxmax, mode,
5249                                                                                                  &needwait);
5250                                 if (needwait)
5251                                 {
5252                                         LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5253                                         XactLockTableWait(rawxmax, rel, &mytup.t_data->t_ctid,
5254                                                                           XLTW_LockUpdated);
5255                                         goto l4;
5256                                 }
5257                                 if (res != HeapTupleMayBeUpdated)
5258                                 {
5259                                         UnlockReleaseBuffer(buf);
5260                                         return res;
5261                                 }
5262                         }
5263                 }
5264
5265                 /* compute the new Xmax and infomask values for the tuple ... */
5266                 compute_new_xmax_infomask(xmax, old_infomask, mytup.t_data->t_infomask2,
5267                                                                   xid, mode, false,
5268                                                                   &new_xmax, &new_infomask, &new_infomask2);
5269
5270                 START_CRIT_SECTION();
5271
5272                 /* ... and set them */
5273                 HeapTupleHeaderSetXmax(mytup.t_data, new_xmax);
5274                 mytup.t_data->t_infomask &= ~HEAP_XMAX_BITS;
5275                 mytup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5276                 mytup.t_data->t_infomask |= new_infomask;
5277                 mytup.t_data->t_infomask2 |= new_infomask2;
5278
5279                 MarkBufferDirty(buf);
5280
5281                 /* XLOG stuff */
5282                 if (RelationNeedsWAL(rel))
5283                 {
5284                         xl_heap_lock_updated xlrec;
5285                         XLogRecPtr      recptr;
5286                         Page            page = BufferGetPage(buf);
5287
5288                         XLogBeginInsert();
5289                         XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
5290
5291                         xlrec.offnum = ItemPointerGetOffsetNumber(&mytup.t_self);
5292                         xlrec.xmax = new_xmax;
5293                         xlrec.infobits_set = compute_infobits(new_infomask, new_infomask2);
5294
5295                         XLogRegisterData((char *) &xlrec, SizeOfHeapLockUpdated);
5296
5297                         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_LOCK_UPDATED);
5298
5299                         PageSetLSN(page, recptr);
5300                 }
5301
5302                 END_CRIT_SECTION();
5303
5304                 /* if we find the end of update chain, we're done. */
5305                 if (mytup.t_data->t_infomask & HEAP_XMAX_INVALID ||
5306                         ItemPointerEquals(&mytup.t_self, &mytup.t_data->t_ctid) ||
5307                         HeapTupleHeaderIsOnlyLocked(mytup.t_data))
5308                 {
5309                         UnlockReleaseBuffer(buf);
5310                         return HeapTupleMayBeUpdated;
5311                 }
5312
5313                 /* tail recursion */
5314                 priorXmax = HeapTupleHeaderGetUpdateXid(mytup.t_data);
5315                 ItemPointerCopy(&(mytup.t_data->t_ctid), &tupid);
5316                 UnlockReleaseBuffer(buf);
5317         }
5318 }
5319
5320 /*
5321  * heap_lock_updated_tuple
5322  *              Follow update chain when locking an updated tuple, acquiring locks (row
5323  *              marks) on the updated versions.
5324  *
5325  * The initial tuple is assumed to be already locked.
5326  *
5327  * This function doesn't check visibility, it just inconditionally marks the
5328  * tuple(s) as locked.  If any tuple in the updated chain is being deleted
5329  * concurrently (or updated with the key being modified), sleep until the
5330  * transaction doing it is finished.
5331  *
5332  * Note that we don't acquire heavyweight tuple locks on the tuples we walk
5333  * when we have to wait for other transactions to release them, as opposed to
5334  * what heap_lock_tuple does.  The reason is that having more than one
5335  * transaction walking the chain is probably uncommon enough that risk of
5336  * starvation is not likely: one of the preconditions for being here is that
5337  * the snapshot in use predates the update that created this tuple (because we
5338  * started at an earlier version of the tuple), but at the same time such a
5339  * transaction cannot be using repeatable read or serializable isolation
5340  * levels, because that would lead to a serializability failure.
5341  */
5342 static HTSU_Result
5343 heap_lock_updated_tuple(Relation rel, HeapTuple tuple, ItemPointer ctid,
5344                                                 TransactionId xid, LockTupleMode mode)
5345 {
5346         if (!ItemPointerEquals(&tuple->t_self, ctid))
5347         {
5348                 /*
5349                  * If this is the first possibly-multixact-able operation in the
5350                  * current transaction, set my per-backend OldestMemberMXactId
5351                  * setting. We can be certain that the transaction will never become a
5352                  * member of any older MultiXactIds than that.  (We have to do this
5353                  * even if we end up just using our own TransactionId below, since
5354                  * some other backend could incorporate our XID into a MultiXact
5355                  * immediately afterwards.)
5356                  */
5357                 MultiXactIdSetOldestMember();
5358
5359                 return heap_lock_updated_tuple_rec(rel, ctid, xid, mode);
5360         }
5361
5362         /* nothing to lock */
5363         return HeapTupleMayBeUpdated;
5364 }
5365
5366
5367 /*
5368  * heap_inplace_update - update a tuple "in place" (ie, overwrite it)
5369  *
5370  * Overwriting violates both MVCC and transactional safety, so the uses
5371  * of this function in Postgres are extremely limited.  Nonetheless we
5372  * find some places to use it.
5373  *
5374  * The tuple cannot change size, and therefore it's reasonable to assume
5375  * that its null bitmap (if any) doesn't change either.  So we just
5376  * overwrite the data portion of the tuple without touching the null
5377  * bitmap or any of the header fields.
5378  *
5379  * tuple is an in-memory tuple structure containing the data to be written
5380  * over the target tuple.  Also, tuple->t_self identifies the target tuple.
5381  */
5382 void
5383 heap_inplace_update(Relation relation, HeapTuple tuple)
5384 {
5385         Buffer          buffer;
5386         Page            page;
5387         OffsetNumber offnum;
5388         ItemId          lp = NULL;
5389         HeapTupleHeader htup;
5390         uint32          oldlen;
5391         uint32          newlen;
5392
5393         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&(tuple->t_self)));
5394         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5395         page = (Page) BufferGetPage(buffer);
5396
5397         offnum = ItemPointerGetOffsetNumber(&(tuple->t_self));
5398         if (PageGetMaxOffsetNumber(page) >= offnum)
5399                 lp = PageGetItemId(page, offnum);
5400
5401         if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
5402                 elog(ERROR, "heap_inplace_update: invalid lp");
5403
5404         htup = (HeapTupleHeader) PageGetItem(page, lp);
5405
5406         oldlen = ItemIdGetLength(lp) - htup->t_hoff;
5407         newlen = tuple->t_len - tuple->t_data->t_hoff;
5408         if (oldlen != newlen || htup->t_hoff != tuple->t_data->t_hoff)
5409                 elog(ERROR, "heap_inplace_update: wrong tuple length");
5410
5411         /* NO EREPORT(ERROR) from here till changes are logged */
5412         START_CRIT_SECTION();
5413
5414         memcpy((char *) htup + htup->t_hoff,
5415                    (char *) tuple->t_data + tuple->t_data->t_hoff,
5416                    newlen);
5417
5418         MarkBufferDirty(buffer);
5419
5420         /* XLOG stuff */
5421         if (RelationNeedsWAL(relation))
5422         {
5423                 xl_heap_inplace xlrec;
5424                 XLogRecPtr      recptr;
5425
5426                 xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
5427
5428                 XLogBeginInsert();
5429                 XLogRegisterData((char *) &xlrec, SizeOfHeapInplace);
5430
5431                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
5432                 XLogRegisterBufData(0, (char *) htup + htup->t_hoff, newlen);
5433
5434                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_INPLACE);
5435
5436                 PageSetLSN(page, recptr);
5437         }
5438
5439         END_CRIT_SECTION();
5440
5441         UnlockReleaseBuffer(buffer);
5442
5443         /*
5444          * Send out shared cache inval if necessary.  Note that because we only
5445          * pass the new version of the tuple, this mustn't be used for any
5446          * operations that could change catcache lookup keys.  But we aren't
5447          * bothering with index updates either, so that's true a fortiori.
5448          */
5449         if (!IsBootstrapProcessingMode())
5450                 CacheInvalidateHeapTuple(relation, tuple, NULL);
5451 }
5452
5453 #define         FRM_NOOP                                0x0001
5454 #define         FRM_INVALIDATE_XMAX             0x0002
5455 #define         FRM_RETURN_IS_XID               0x0004
5456 #define         FRM_RETURN_IS_MULTI             0x0008
5457 #define         FRM_MARK_COMMITTED              0x0010
5458
5459 /*
5460  * FreezeMultiXactId
5461  *              Determine what to do during freezing when a tuple is marked by a
5462  *              MultiXactId.
5463  *
5464  * NB -- this might have the side-effect of creating a new MultiXactId!
5465  *
5466  * "flags" is an output value; it's used to tell caller what to do on return.
5467  * Possible flags are:
5468  * FRM_NOOP
5469  *              don't do anything -- keep existing Xmax
5470  * FRM_INVALIDATE_XMAX
5471  *              mark Xmax as InvalidTransactionId and set XMAX_INVALID flag.
5472  * FRM_RETURN_IS_XID
5473  *              The Xid return value is a single update Xid to set as xmax.
5474  * FRM_MARK_COMMITTED
5475  *              Xmax can be marked as HEAP_XMAX_COMMITTED
5476  * FRM_RETURN_IS_MULTI
5477  *              The return value is a new MultiXactId to set as new Xmax.
5478  *              (caller must obtain proper infomask bits using GetMultiXactIdHintBits)
5479  */
5480 static TransactionId
5481 FreezeMultiXactId(MultiXactId multi, uint16 t_infomask,
5482                                   TransactionId cutoff_xid, MultiXactId cutoff_multi,
5483                                   uint16 *flags)
5484 {
5485         TransactionId xid = InvalidTransactionId;
5486         int                     i;
5487         MultiXactMember *members;
5488         int                     nmembers;
5489         bool            need_replace;
5490         int                     nnewmembers;
5491         MultiXactMember *newmembers;
5492         bool            has_lockers;
5493         TransactionId update_xid;
5494         bool            update_committed;
5495         bool            allow_old;
5496
5497         *flags = 0;
5498
5499         /* We should only be called in Multis */
5500         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
5501
5502         if (!MultiXactIdIsValid(multi))
5503         {
5504                 /* Ensure infomask bits are appropriately set/reset */
5505                 *flags |= FRM_INVALIDATE_XMAX;
5506                 return InvalidTransactionId;
5507         }
5508         else if (MultiXactIdPrecedes(multi, cutoff_multi))
5509         {
5510                 /*
5511                  * This old multi cannot possibly have members still running.  If it
5512                  * was a locker only, it can be removed without any further
5513                  * consideration; but if it contained an update, we might need to
5514                  * preserve it.
5515                  *
5516                  * Don't assert MultiXactIdIsRunning if the multi came from a
5517                  * pg_upgrade'd share-locked tuple, though, as doing that causes an
5518                  * error to be raised unnecessarily.
5519                  */
5520                 Assert((!(t_infomask & HEAP_LOCK_MASK) &&
5521                                 HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)) ||
5522                            !MultiXactIdIsRunning(multi,
5523                                                                          HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)));
5524                 if (HEAP_XMAX_IS_LOCKED_ONLY(t_infomask))
5525                 {
5526                         *flags |= FRM_INVALIDATE_XMAX;
5527                         xid = InvalidTransactionId; /* not strictly necessary */
5528                 }
5529                 else
5530                 {
5531                         /* replace multi by update xid */
5532                         xid = MultiXactIdGetUpdateXid(multi, t_infomask);
5533
5534                         /* wasn't only a lock, xid needs to be valid */
5535                         Assert(TransactionIdIsValid(xid));
5536
5537                         /*
5538                          * If the xid is older than the cutoff, it has to have aborted,
5539                          * otherwise the tuple would have gotten pruned away.
5540                          */
5541                         if (TransactionIdPrecedes(xid, cutoff_xid))
5542                         {
5543                                 Assert(!TransactionIdDidCommit(xid));
5544                                 *flags |= FRM_INVALIDATE_XMAX;
5545                                 xid = InvalidTransactionId;             /* not strictly necessary */
5546                         }
5547                         else
5548                         {
5549                                 *flags |= FRM_RETURN_IS_XID;
5550                         }
5551                 }
5552
5553                 return xid;
5554         }
5555
5556         /*
5557          * This multixact might have or might not have members still running, but
5558          * we know it's valid and is newer than the cutoff point for multis.
5559          * However, some member(s) of it may be below the cutoff for Xids, so we
5560          * need to walk the whole members array to figure out what to do, if
5561          * anything.
5562          */
5563
5564         allow_old = !(t_infomask & HEAP_LOCK_MASK) &&
5565                 HEAP_XMAX_IS_LOCKED_ONLY(t_infomask);
5566         nmembers =
5567                 GetMultiXactIdMembers(multi, &members, allow_old,
5568                                                           HEAP_XMAX_IS_LOCKED_ONLY(t_infomask));
5569         if (nmembers <= 0)
5570         {
5571                 /* Nothing worth keeping */
5572                 *flags |= FRM_INVALIDATE_XMAX;
5573                 return InvalidTransactionId;
5574         }
5575
5576         /* is there anything older than the cutoff? */
5577         need_replace = false;
5578         for (i = 0; i < nmembers; i++)
5579         {
5580                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
5581                 {
5582                         need_replace = true;
5583                         break;
5584                 }
5585         }
5586
5587         /*
5588          * In the simplest case, there is no member older than the cutoff; we can
5589          * keep the existing MultiXactId as is.
5590          */
5591         if (!need_replace)
5592         {
5593                 *flags |= FRM_NOOP;
5594                 pfree(members);
5595                 return InvalidTransactionId;
5596         }
5597
5598         /*
5599          * If the multi needs to be updated, figure out which members do we need
5600          * to keep.
5601          */
5602         nnewmembers = 0;
5603         newmembers = palloc(sizeof(MultiXactMember) * nmembers);
5604         has_lockers = false;
5605         update_xid = InvalidTransactionId;
5606         update_committed = false;
5607
5608         for (i = 0; i < nmembers; i++)
5609         {
5610                 /*
5611                  * Determine whether to keep this member or ignore it.
5612                  */
5613                 if (ISUPDATE_from_mxstatus(members[i].status))
5614                 {
5615                         TransactionId xid = members[i].xid;
5616
5617                         /*
5618                          * It's an update; should we keep it?  If the transaction is known
5619                          * aborted or crashed then it's okay to ignore it, otherwise not.
5620                          * Note that an updater older than cutoff_xid cannot possibly be
5621                          * committed, because HeapTupleSatisfiesVacuum would have returned
5622                          * HEAPTUPLE_DEAD and we would not be trying to freeze the tuple.
5623                          *
5624                          * As with all tuple visibility routines, it's critical to test
5625                          * TransactionIdIsInProgress before TransactionIdDidCommit,
5626                          * because of race conditions explained in detail in tqual.c.
5627                          */
5628                         if (TransactionIdIsCurrentTransactionId(xid) ||
5629                                 TransactionIdIsInProgress(xid))
5630                         {
5631                                 Assert(!TransactionIdIsValid(update_xid));
5632                                 update_xid = xid;
5633                         }
5634                         else if (TransactionIdDidCommit(xid))
5635                         {
5636                                 /*
5637                                  * The transaction committed, so we can tell caller to set
5638                                  * HEAP_XMAX_COMMITTED.  (We can only do this because we know
5639                                  * the transaction is not running.)
5640                                  */
5641                                 Assert(!TransactionIdIsValid(update_xid));
5642                                 update_committed = true;
5643                                 update_xid = xid;
5644                         }
5645
5646                         /*
5647                          * Not in progress, not committed -- must be aborted or crashed;
5648                          * we can ignore it.
5649                          */
5650
5651                         /*
5652                          * Since the tuple wasn't marked HEAPTUPLE_DEAD by vacuum, the
5653                          * update Xid cannot possibly be older than the xid cutoff.
5654                          */
5655                         Assert(!TransactionIdIsValid(update_xid) ||
5656                                    !TransactionIdPrecedes(update_xid, cutoff_xid));
5657
5658                         /*
5659                          * If we determined that it's an Xid corresponding to an update
5660                          * that must be retained, additionally add it to the list of
5661                          * members of the new Multi, in case we end up using that.  (We
5662                          * might still decide to use only an update Xid and not a multi,
5663                          * but it's easier to maintain the list as we walk the old members
5664                          * list.)
5665                          */
5666                         if (TransactionIdIsValid(update_xid))
5667                                 newmembers[nnewmembers++] = members[i];
5668                 }
5669                 else
5670                 {
5671                         /* We only keep lockers if they are still running */
5672                         if (TransactionIdIsCurrentTransactionId(members[i].xid) ||
5673                                 TransactionIdIsInProgress(members[i].xid))
5674                         {
5675                                 /* running locker cannot possibly be older than the cutoff */
5676                                 Assert(!TransactionIdPrecedes(members[i].xid, cutoff_xid));
5677                                 newmembers[nnewmembers++] = members[i];
5678                                 has_lockers = true;
5679                         }
5680                 }
5681         }
5682
5683         pfree(members);
5684
5685         if (nnewmembers == 0)
5686         {
5687                 /* nothing worth keeping!? Tell caller to remove the whole thing */
5688                 *flags |= FRM_INVALIDATE_XMAX;
5689                 xid = InvalidTransactionId;
5690         }
5691         else if (TransactionIdIsValid(update_xid) && !has_lockers)
5692         {
5693                 /*
5694                  * If there's a single member and it's an update, pass it back alone
5695                  * without creating a new Multi.  (XXX we could do this when there's a
5696                  * single remaining locker, too, but that would complicate the API too
5697                  * much; moreover, the case with the single updater is more
5698                  * interesting, because those are longer-lived.)
5699                  */
5700                 Assert(nnewmembers == 1);
5701                 *flags |= FRM_RETURN_IS_XID;
5702                 if (update_committed)
5703                         *flags |= FRM_MARK_COMMITTED;
5704                 xid = update_xid;
5705         }
5706         else
5707         {
5708                 /*
5709                  * Create a new multixact with the surviving members of the previous
5710                  * one, to set as new Xmax in the tuple.
5711                  */
5712                 xid = MultiXactIdCreateFromMembers(nnewmembers, newmembers);
5713                 *flags |= FRM_RETURN_IS_MULTI;
5714         }
5715
5716         pfree(newmembers);
5717
5718         return xid;
5719 }
5720
5721 /*
5722  * heap_prepare_freeze_tuple
5723  *
5724  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
5725  * are older than the specified cutoff XID and cutoff MultiXactId.  If so,
5726  * setup enough state (in the *frz output argument) to later execute and
5727  * WAL-log what we would need to do, and return TRUE.  Return FALSE if nothing
5728  * is to be changed.
5729  *
5730  * Caller is responsible for setting the offset field, if appropriate.
5731  *
5732  * It is assumed that the caller has checked the tuple with
5733  * HeapTupleSatisfiesVacuum() and determined that it is not HEAPTUPLE_DEAD
5734  * (else we should be removing the tuple, not freezing it).
5735  *
5736  * NB: cutoff_xid *must* be <= the current global xmin, to ensure that any
5737  * XID older than it could neither be running nor seen as running by any
5738  * open transaction.  This ensures that the replacement will not change
5739  * anyone's idea of the tuple state.
5740  * Similarly, cutoff_multi must be less than or equal to the smallest
5741  * MultiXactId used by any transaction currently open.
5742  *
5743  * If the tuple is in a shared buffer, caller must hold an exclusive lock on
5744  * that buffer.
5745  *
5746  * NB: It is not enough to set hint bits to indicate something is
5747  * committed/invalid -- they might not be set on a standby, or after crash
5748  * recovery.  We really need to remove old xids.
5749  */
5750 bool
5751 heap_prepare_freeze_tuple(HeapTupleHeader tuple, TransactionId cutoff_xid,
5752                                                   TransactionId cutoff_multi,
5753                                                   xl_heap_freeze_tuple *frz)
5754
5755 {
5756         bool            changed = false;
5757         bool            freeze_xmax = false;
5758         TransactionId xid;
5759
5760         frz->frzflags = 0;
5761         frz->t_infomask2 = tuple->t_infomask2;
5762         frz->t_infomask = tuple->t_infomask;
5763         frz->xmax = HeapTupleHeaderGetRawXmax(tuple);
5764
5765         /* Process xmin */
5766         xid = HeapTupleHeaderGetXmin(tuple);
5767         if (TransactionIdIsNormal(xid) &&
5768                 TransactionIdPrecedes(xid, cutoff_xid))
5769         {
5770                 frz->t_infomask |= HEAP_XMIN_FROZEN;
5771                 changed = true;
5772         }
5773
5774         /*
5775          * Process xmax.  To thoroughly examine the current Xmax value we need to
5776          * resolve a MultiXactId to its member Xids, in case some of them are
5777          * below the given cutoff for Xids.  In that case, those values might need
5778          * freezing, too.  Also, if a multi needs freezing, we cannot simply take
5779          * it out --- if there's a live updater Xid, it needs to be kept.
5780          *
5781          * Make sure to keep heap_tuple_needs_freeze in sync with this.
5782          */
5783         xid = HeapTupleHeaderGetRawXmax(tuple);
5784
5785         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
5786         {
5787                 TransactionId newxmax;
5788                 uint16          flags;
5789
5790                 newxmax = FreezeMultiXactId(xid, tuple->t_infomask,
5791                                                                         cutoff_xid, cutoff_multi, &flags);
5792
5793                 if (flags & FRM_INVALIDATE_XMAX)
5794                         freeze_xmax = true;
5795                 else if (flags & FRM_RETURN_IS_XID)
5796                 {
5797                         /*
5798                          * NB -- some of these transformations are only valid because we
5799                          * know the return Xid is a tuple updater (i.e. not merely a
5800                          * locker.) Also note that the only reason we don't explicitely
5801                          * worry about HEAP_KEYS_UPDATED is because it lives in
5802                          * t_infomask2 rather than t_infomask.
5803                          */
5804                         frz->t_infomask &= ~HEAP_XMAX_BITS;
5805                         frz->xmax = newxmax;
5806                         if (flags & FRM_MARK_COMMITTED)
5807                                 frz->t_infomask &= HEAP_XMAX_COMMITTED;
5808                         changed = true;
5809                 }
5810                 else if (flags & FRM_RETURN_IS_MULTI)
5811                 {
5812                         uint16          newbits;
5813                         uint16          newbits2;
5814
5815                         /*
5816                          * We can't use GetMultiXactIdHintBits directly on the new multi
5817                          * here; that routine initializes the masks to all zeroes, which
5818                          * would lose other bits we need.  Doing it this way ensures all
5819                          * unrelated bits remain untouched.
5820                          */
5821                         frz->t_infomask &= ~HEAP_XMAX_BITS;
5822                         frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5823                         GetMultiXactIdHintBits(newxmax, &newbits, &newbits2);
5824                         frz->t_infomask |= newbits;
5825                         frz->t_infomask2 |= newbits2;
5826
5827                         frz->xmax = newxmax;
5828
5829                         changed = true;
5830                 }
5831                 else
5832                 {
5833                         Assert(flags & FRM_NOOP);
5834                 }
5835         }
5836         else if (TransactionIdIsNormal(xid) &&
5837                          TransactionIdPrecedes(xid, cutoff_xid))
5838         {
5839                 freeze_xmax = true;
5840         }
5841
5842         if (freeze_xmax)
5843         {
5844                 frz->xmax = InvalidTransactionId;
5845
5846                 /*
5847                  * The tuple might be marked either XMAX_INVALID or XMAX_COMMITTED +
5848                  * LOCKED.  Normalize to INVALID just to be sure no one gets confused.
5849                  * Also get rid of the HEAP_KEYS_UPDATED bit.
5850                  */
5851                 frz->t_infomask &= ~HEAP_XMAX_BITS;
5852                 frz->t_infomask |= HEAP_XMAX_INVALID;
5853                 frz->t_infomask2 &= ~HEAP_HOT_UPDATED;
5854                 frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5855                 changed = true;
5856         }
5857
5858         /*
5859          * Old-style VACUUM FULL is gone, but we have to keep this code as long as
5860          * we support having MOVED_OFF/MOVED_IN tuples in the database.
5861          */
5862         if (tuple->t_infomask & HEAP_MOVED)
5863         {
5864                 xid = HeapTupleHeaderGetXvac(tuple);
5865                 if (TransactionIdIsNormal(xid) &&
5866                         TransactionIdPrecedes(xid, cutoff_xid))
5867                 {
5868                         /*
5869                          * If a MOVED_OFF tuple is not dead, the xvac transaction must
5870                          * have failed; whereas a non-dead MOVED_IN tuple must mean the
5871                          * xvac transaction succeeded.
5872                          */
5873                         if (tuple->t_infomask & HEAP_MOVED_OFF)
5874                                 frz->frzflags |= XLH_INVALID_XVAC;
5875                         else
5876                                 frz->frzflags |= XLH_FREEZE_XVAC;
5877
5878                         /*
5879                          * Might as well fix the hint bits too; usually XMIN_COMMITTED
5880                          * will already be set here, but there's a small chance not.
5881                          */
5882                         Assert(!(tuple->t_infomask & HEAP_XMIN_INVALID));
5883                         frz->t_infomask |= HEAP_XMIN_COMMITTED;
5884                         changed = true;
5885                 }
5886         }
5887
5888         return changed;
5889 }
5890
5891 /*
5892  * heap_execute_freeze_tuple
5893  *              Execute the prepared freezing of a tuple.
5894  *
5895  * Caller is responsible for ensuring that no other backend can access the
5896  * storage underlying this tuple, either by holding an exclusive lock on the
5897  * buffer containing it (which is what lazy VACUUM does), or by having it by
5898  * in private storage (which is what CLUSTER and friends do).
5899  *
5900  * Note: it might seem we could make the changes without exclusive lock, since
5901  * TransactionId read/write is assumed atomic anyway.  However there is a race
5902  * condition: someone who just fetched an old XID that we overwrite here could
5903  * conceivably not finish checking the XID against pg_clog before we finish
5904  * the VACUUM and perhaps truncate off the part of pg_clog he needs.  Getting
5905  * exclusive lock ensures no other backend is in process of checking the
5906  * tuple status.  Also, getting exclusive lock makes it safe to adjust the
5907  * infomask bits.
5908  *
5909  * NB: All code in here must be safe to execute during crash recovery!
5910  */
5911 void
5912 heap_execute_freeze_tuple(HeapTupleHeader tuple, xl_heap_freeze_tuple *frz)
5913 {
5914         HeapTupleHeaderSetXmax(tuple, frz->xmax);
5915
5916         if (frz->frzflags & XLH_FREEZE_XVAC)
5917                 HeapTupleHeaderSetXvac(tuple, FrozenTransactionId);
5918
5919         if (frz->frzflags & XLH_INVALID_XVAC)
5920                 HeapTupleHeaderSetXvac(tuple, InvalidTransactionId);
5921
5922         tuple->t_infomask = frz->t_infomask;
5923         tuple->t_infomask2 = frz->t_infomask2;
5924 }
5925
5926 /*
5927  * heap_freeze_tuple
5928  *              Freeze tuple in place, without WAL logging.
5929  *
5930  * Useful for callers like CLUSTER that perform their own WAL logging.
5931  */
5932 bool
5933 heap_freeze_tuple(HeapTupleHeader tuple, TransactionId cutoff_xid,
5934                                   TransactionId cutoff_multi)
5935 {
5936         xl_heap_freeze_tuple frz;
5937         bool            do_freeze;
5938
5939         do_freeze = heap_prepare_freeze_tuple(tuple, cutoff_xid, cutoff_multi,
5940                                                                                   &frz);
5941
5942         /*
5943          * Note that because this is not a WAL-logged operation, we don't need to
5944          * fill in the offset in the freeze record.
5945          */
5946
5947         if (do_freeze)
5948                 heap_execute_freeze_tuple(tuple, &frz);
5949         return do_freeze;
5950 }
5951
5952 /*
5953  * For a given MultiXactId, return the hint bits that should be set in the
5954  * tuple's infomask.
5955  *
5956  * Normally this should be called for a multixact that was just created, and
5957  * so is on our local cache, so the GetMembers call is fast.
5958  */
5959 static void
5960 GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
5961                                            uint16 *new_infomask2)
5962 {
5963         int                     nmembers;
5964         MultiXactMember *members;
5965         int                     i;
5966         uint16          bits = HEAP_XMAX_IS_MULTI;
5967         uint16          bits2 = 0;
5968         bool            has_update = false;
5969         LockTupleMode strongest = LockTupleKeyShare;
5970
5971         /*
5972          * We only use this in multis we just created, so they cannot be values
5973          * pre-pg_upgrade.
5974          */
5975         nmembers = GetMultiXactIdMembers(multi, &members, false, false);
5976
5977         for (i = 0; i < nmembers; i++)
5978         {
5979                 LockTupleMode mode;
5980
5981                 /*
5982                  * Remember the strongest lock mode held by any member of the
5983                  * multixact.
5984                  */
5985                 mode = TUPLOCK_from_mxstatus(members[i].status);
5986                 if (mode > strongest)
5987                         strongest = mode;
5988
5989                 /* See what other bits we need */
5990                 switch (members[i].status)
5991                 {
5992                         case MultiXactStatusForKeyShare:
5993                         case MultiXactStatusForShare:
5994                         case MultiXactStatusForNoKeyUpdate:
5995                                 break;
5996
5997                         case MultiXactStatusForUpdate:
5998                                 bits2 |= HEAP_KEYS_UPDATED;
5999                                 break;
6000
6001                         case MultiXactStatusNoKeyUpdate:
6002                                 has_update = true;
6003                                 break;
6004
6005                         case MultiXactStatusUpdate:
6006                                 bits2 |= HEAP_KEYS_UPDATED;
6007                                 has_update = true;
6008                                 break;
6009                 }
6010         }
6011
6012         if (strongest == LockTupleExclusive ||
6013                 strongest == LockTupleNoKeyExclusive)
6014                 bits |= HEAP_XMAX_EXCL_LOCK;
6015         else if (strongest == LockTupleShare)
6016                 bits |= HEAP_XMAX_SHR_LOCK;
6017         else if (strongest == LockTupleKeyShare)
6018                 bits |= HEAP_XMAX_KEYSHR_LOCK;
6019
6020         if (!has_update)
6021                 bits |= HEAP_XMAX_LOCK_ONLY;
6022
6023         if (nmembers > 0)
6024                 pfree(members);
6025
6026         *new_infomask = bits;
6027         *new_infomask2 = bits2;
6028 }
6029
6030 /*
6031  * MultiXactIdGetUpdateXid
6032  *
6033  * Given a multixact Xmax and corresponding infomask, which does not have the
6034  * HEAP_XMAX_LOCK_ONLY bit set, obtain and return the Xid of the updating
6035  * transaction.
6036  *
6037  * Caller is expected to check the status of the updating transaction, if
6038  * necessary.
6039  */
6040 static TransactionId
6041 MultiXactIdGetUpdateXid(TransactionId xmax, uint16 t_infomask)
6042 {
6043         TransactionId update_xact = InvalidTransactionId;
6044         MultiXactMember *members;
6045         int                     nmembers;
6046
6047         Assert(!(t_infomask & HEAP_XMAX_LOCK_ONLY));
6048         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6049
6050         /*
6051          * Since we know the LOCK_ONLY bit is not set, this cannot be a multi from
6052          * pre-pg_upgrade.
6053          */
6054         nmembers = GetMultiXactIdMembers(xmax, &members, false, false);
6055
6056         if (nmembers > 0)
6057         {
6058                 int                     i;
6059
6060                 for (i = 0; i < nmembers; i++)
6061                 {
6062                         /* Ignore lockers */
6063                         if (!ISUPDATE_from_mxstatus(members[i].status))
6064                                 continue;
6065
6066                         /* there can be at most one updater */
6067                         Assert(update_xact == InvalidTransactionId);
6068                         update_xact = members[i].xid;
6069 #ifndef USE_ASSERT_CHECKING
6070
6071                         /*
6072                          * in an assert-enabled build, walk the whole array to ensure
6073                          * there's no other updater.
6074                          */
6075                         break;
6076 #endif
6077                 }
6078
6079                 pfree(members);
6080         }
6081
6082         return update_xact;
6083 }
6084
6085 /*
6086  * HeapTupleGetUpdateXid
6087  *              As above, but use a HeapTupleHeader
6088  *
6089  * See also HeapTupleHeaderGetUpdateXid, which can be used without previously
6090  * checking the hint bits.
6091  */
6092 TransactionId
6093 HeapTupleGetUpdateXid(HeapTupleHeader tuple)
6094 {
6095         return MultiXactIdGetUpdateXid(HeapTupleHeaderGetRawXmax(tuple),
6096                                                                    tuple->t_infomask);
6097 }
6098
6099 /*
6100  * Do_MultiXactIdWait
6101  *              Actual implementation for the two functions below.
6102  *
6103  * 'multi', 'status' and 'infomask' indicate what to sleep on (the status is
6104  * needed to ensure we only sleep on conflicting members, and the infomask is
6105  * used to optimize multixact access in case it's a lock-only multi); 'nowait'
6106  * indicates whether to use conditional lock acquisition, to allow callers to
6107  * fail if lock is unavailable.  'rel', 'ctid' and 'oper' are used to set up
6108  * context information for error messages.  'remaining', if not NULL, receives
6109  * the number of members that are still running, including any (non-aborted)
6110  * subtransactions of our own transaction.
6111  *
6112  * We do this by sleeping on each member using XactLockTableWait.  Any
6113  * members that belong to the current backend are *not* waited for, however;
6114  * this would not merely be useless but would lead to Assert failure inside
6115  * XactLockTableWait.  By the time this returns, it is certain that all
6116  * transactions *of other backends* that were members of the MultiXactId
6117  * that conflict with the requested status are dead (and no new ones can have
6118  * been added, since it is not legal to add members to an existing
6119  * MultiXactId).
6120  *
6121  * But by the time we finish sleeping, someone else may have changed the Xmax
6122  * of the containing tuple, so the caller needs to iterate on us somehow.
6123  *
6124  * Note that in case we return false, the number of remaining members is
6125  * not to be trusted.
6126  */
6127 static bool
6128 Do_MultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6129                                    uint16 infomask, bool nowait,
6130                                    Relation rel, ItemPointer ctid, XLTW_Oper oper,
6131                                    int *remaining)
6132 {
6133         bool            allow_old;
6134         bool            result = true;
6135         MultiXactMember *members;
6136         int                     nmembers;
6137         int                     remain = 0;
6138
6139         allow_old = !(infomask & HEAP_LOCK_MASK) && HEAP_XMAX_IS_LOCKED_ONLY(infomask);
6140         nmembers = GetMultiXactIdMembers(multi, &members, allow_old,
6141                                                                          HEAP_XMAX_IS_LOCKED_ONLY(infomask));
6142
6143         if (nmembers >= 0)
6144         {
6145                 int                     i;
6146
6147                 for (i = 0; i < nmembers; i++)
6148                 {
6149                         TransactionId memxid = members[i].xid;
6150                         MultiXactStatus memstatus = members[i].status;
6151
6152                         if (TransactionIdIsCurrentTransactionId(memxid))
6153                         {
6154                                 remain++;
6155                                 continue;
6156                         }
6157
6158                         if (!DoLockModesConflict(LOCKMODE_from_mxstatus(memstatus),
6159                                                                          LOCKMODE_from_mxstatus(status)))
6160                         {
6161                                 if (remaining && TransactionIdIsInProgress(memxid))
6162                                         remain++;
6163                                 continue;
6164                         }
6165
6166                         /*
6167                          * This member conflicts with our multi, so we have to sleep (or
6168                          * return failure, if asked to avoid waiting.)
6169                          *
6170                          * Note that we don't set up an error context callback ourselves,
6171                          * but instead we pass the info down to XactLockTableWait.  This
6172                          * might seem a bit wasteful because the context is set up and
6173                          * tore down for each member of the multixact, but in reality it
6174                          * should be barely noticeable, and it avoids duplicate code.
6175                          */
6176                         if (nowait)
6177                         {
6178                                 result = ConditionalXactLockTableWait(memxid);
6179                                 if (!result)
6180                                         break;
6181                         }
6182                         else
6183                                 XactLockTableWait(memxid, rel, ctid, oper);
6184                 }
6185
6186                 pfree(members);
6187         }
6188
6189         if (remaining)
6190                 *remaining = remain;
6191
6192         return result;
6193 }
6194
6195 /*
6196  * MultiXactIdWait
6197  *              Sleep on a MultiXactId.
6198  *
6199  * By the time we finish sleeping, someone else may have changed the Xmax
6200  * of the containing tuple, so the caller needs to iterate on us somehow.
6201  *
6202  * We return (in *remaining, if not NULL) the number of members that are still
6203  * running, including any (non-aborted) subtransactions of our own transaction.
6204  */
6205 static void
6206 MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
6207                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
6208                                 int *remaining)
6209 {
6210         (void) Do_MultiXactIdWait(multi, status, infomask, false,
6211                                                           rel, ctid, oper, remaining);
6212 }
6213
6214 /*
6215  * ConditionalMultiXactIdWait
6216  *              As above, but only lock if we can get the lock without blocking.
6217  *
6218  * By the time we finish sleeping, someone else may have changed the Xmax
6219  * of the containing tuple, so the caller needs to iterate on us somehow.
6220  *
6221  * If the multixact is now all gone, return true.  Returns false if some
6222  * transactions might still be running.
6223  *
6224  * We return (in *remaining, if not NULL) the number of members that are still
6225  * running, including any (non-aborted) subtransactions of our own transaction.
6226  */
6227 static bool
6228 ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6229                                                    uint16 infomask, Relation rel, int *remaining)
6230 {
6231         return Do_MultiXactIdWait(multi, status, infomask, true,
6232                                                           rel, NULL, XLTW_None, remaining);
6233 }
6234
6235 /*
6236  * heap_tuple_needs_freeze
6237  *
6238  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6239  * are older than the specified cutoff XID or MultiXactId.  If so, return TRUE.
6240  *
6241  * It doesn't matter whether the tuple is alive or dead, we are checking
6242  * to see if a tuple needs to be removed or frozen to avoid wraparound.
6243  *
6244  * NB: Cannot rely on hint bits here, they might not be set after a crash or
6245  * on a standby.
6246  */
6247 bool
6248 heap_tuple_needs_freeze(HeapTupleHeader tuple, TransactionId cutoff_xid,
6249                                                 MultiXactId cutoff_multi, Buffer buf)
6250 {
6251         TransactionId xid;
6252
6253         xid = HeapTupleHeaderGetXmin(tuple);
6254         if (TransactionIdIsNormal(xid) &&
6255                 TransactionIdPrecedes(xid, cutoff_xid))
6256                 return true;
6257
6258         /*
6259          * The considerations for multixacts are complicated; look at
6260          * heap_freeze_tuple for justifications.  This routine had better be in
6261          * sync with that one!
6262          */
6263         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
6264         {
6265                 MultiXactId multi;
6266
6267                 multi = HeapTupleHeaderGetRawXmax(tuple);
6268                 if (!MultiXactIdIsValid(multi))
6269                 {
6270                         /* no xmax set, ignore */
6271                         ;
6272                 }
6273                 else if (MultiXactIdPrecedes(multi, cutoff_multi))
6274                         return true;
6275                 else
6276                 {
6277                         MultiXactMember *members;
6278                         int                     nmembers;
6279                         int                     i;
6280                         bool            allow_old;
6281
6282                         /* need to check whether any member of the mxact is too old */
6283
6284                         allow_old = !(tuple->t_infomask & HEAP_LOCK_MASK) &&
6285                                 HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask);
6286                         nmembers = GetMultiXactIdMembers(multi, &members, allow_old,
6287                                                                 HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask));
6288
6289                         for (i = 0; i < nmembers; i++)
6290                         {
6291                                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
6292                                 {
6293                                         pfree(members);
6294                                         return true;
6295                                 }
6296                         }
6297                         if (nmembers > 0)
6298                                 pfree(members);
6299                 }
6300         }
6301         else
6302         {
6303                 xid = HeapTupleHeaderGetRawXmax(tuple);
6304                 if (TransactionIdIsNormal(xid) &&
6305                         TransactionIdPrecedes(xid, cutoff_xid))
6306                         return true;
6307         }
6308
6309         if (tuple->t_infomask & HEAP_MOVED)
6310         {
6311                 xid = HeapTupleHeaderGetXvac(tuple);
6312                 if (TransactionIdIsNormal(xid) &&
6313                         TransactionIdPrecedes(xid, cutoff_xid))
6314                         return true;
6315         }
6316
6317         return false;
6318 }
6319
6320 /* ----------------
6321  *              heap_markpos    - mark scan position
6322  * ----------------
6323  */
6324 void
6325 heap_markpos(HeapScanDesc scan)
6326 {
6327         /* Note: no locking manipulations needed */
6328
6329         if (scan->rs_ctup.t_data != NULL)
6330         {
6331                 scan->rs_mctid = scan->rs_ctup.t_self;
6332                 if (scan->rs_pageatatime)
6333                         scan->rs_mindex = scan->rs_cindex;
6334         }
6335         else
6336                 ItemPointerSetInvalid(&scan->rs_mctid);
6337 }
6338
6339 /* ----------------
6340  *              heap_restrpos   - restore position to marked location
6341  * ----------------
6342  */
6343 void
6344 heap_restrpos(HeapScanDesc scan)
6345 {
6346         /* XXX no amrestrpos checking that ammarkpos called */
6347
6348         if (!ItemPointerIsValid(&scan->rs_mctid))
6349         {
6350                 scan->rs_ctup.t_data = NULL;
6351
6352                 /*
6353                  * unpin scan buffers
6354                  */
6355                 if (BufferIsValid(scan->rs_cbuf))
6356                         ReleaseBuffer(scan->rs_cbuf);
6357                 scan->rs_cbuf = InvalidBuffer;
6358                 scan->rs_cblock = InvalidBlockNumber;
6359                 scan->rs_inited = false;
6360         }
6361         else
6362         {
6363                 /*
6364                  * If we reached end of scan, rs_inited will now be false.  We must
6365                  * reset it to true to keep heapgettup from doing the wrong thing.
6366                  */
6367                 scan->rs_inited = true;
6368                 scan->rs_ctup.t_self = scan->rs_mctid;
6369                 if (scan->rs_pageatatime)
6370                 {
6371                         scan->rs_cindex = scan->rs_mindex;
6372                         heapgettup_pagemode(scan,
6373                                                                 NoMovementScanDirection,
6374                                                                 0,              /* needn't recheck scan keys */
6375                                                                 NULL);
6376                 }
6377                 else
6378                         heapgettup(scan,
6379                                            NoMovementScanDirection,
6380                                            0,           /* needn't recheck scan keys */
6381                                            NULL);
6382         }
6383 }
6384
6385 /*
6386  * If 'tuple' contains any visible XID greater than latestRemovedXid,
6387  * ratchet forwards latestRemovedXid to the greatest one found.
6388  * This is used as the basis for generating Hot Standby conflicts, so
6389  * if a tuple was never visible then removing it should not conflict
6390  * with queries.
6391  */
6392 void
6393 HeapTupleHeaderAdvanceLatestRemovedXid(HeapTupleHeader tuple,
6394                                                                            TransactionId *latestRemovedXid)
6395 {
6396         TransactionId xmin = HeapTupleHeaderGetXmin(tuple);
6397         TransactionId xmax = HeapTupleHeaderGetUpdateXid(tuple);
6398         TransactionId xvac = HeapTupleHeaderGetXvac(tuple);
6399
6400         if (tuple->t_infomask & HEAP_MOVED)
6401         {
6402                 if (TransactionIdPrecedes(*latestRemovedXid, xvac))
6403                         *latestRemovedXid = xvac;
6404         }
6405
6406         /*
6407          * Ignore tuples inserted by an aborted transaction or if the tuple was
6408          * updated/deleted by the inserting transaction.
6409          *
6410          * Look for a committed hint bit, or if no xmin bit is set, check clog.
6411          * This needs to work on both master and standby, where it is used to
6412          * assess btree delete records.
6413          */
6414         if (HeapTupleHeaderXminCommitted(tuple) ||
6415                 (!HeapTupleHeaderXminInvalid(tuple) && TransactionIdDidCommit(xmin)))
6416         {
6417                 if (xmax != xmin &&
6418                         TransactionIdFollows(xmax, *latestRemovedXid))
6419                         *latestRemovedXid = xmax;
6420         }
6421
6422         /* *latestRemovedXid may still be invalid at end */
6423 }
6424
6425 /*
6426  * Perform XLogInsert to register a heap cleanup info message. These
6427  * messages are sent once per VACUUM and are required because
6428  * of the phasing of removal operations during a lazy VACUUM.
6429  * see comments for vacuum_log_cleanup_info().
6430  */
6431 XLogRecPtr
6432 log_heap_cleanup_info(RelFileNode rnode, TransactionId latestRemovedXid)
6433 {
6434         xl_heap_cleanup_info xlrec;
6435         XLogRecPtr      recptr;
6436
6437         xlrec.node = rnode;
6438         xlrec.latestRemovedXid = latestRemovedXid;
6439
6440         XLogBeginInsert();
6441         XLogRegisterData((char *) &xlrec, SizeOfHeapCleanupInfo);
6442
6443         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_CLEANUP_INFO);
6444
6445         return recptr;
6446 }
6447
6448 /*
6449  * Perform XLogInsert for a heap-clean operation.  Caller must already
6450  * have modified the buffer and marked it dirty.
6451  *
6452  * Note: prior to Postgres 8.3, the entries in the nowunused[] array were
6453  * zero-based tuple indexes.  Now they are one-based like other uses
6454  * of OffsetNumber.
6455  *
6456  * We also include latestRemovedXid, which is the greatest XID present in
6457  * the removed tuples. That allows recovery processing to cancel or wait
6458  * for long standby queries that can still see these tuples.
6459  */
6460 XLogRecPtr
6461 log_heap_clean(Relation reln, Buffer buffer,
6462                            OffsetNumber *redirected, int nredirected,
6463                            OffsetNumber *nowdead, int ndead,
6464                            OffsetNumber *nowunused, int nunused,
6465                            TransactionId latestRemovedXid)
6466 {
6467         xl_heap_clean xlrec;
6468         XLogRecPtr      recptr;
6469
6470         /* Caller should not call me on a non-WAL-logged relation */
6471         Assert(RelationNeedsWAL(reln));
6472
6473         xlrec.latestRemovedXid = latestRemovedXid;
6474         xlrec.nredirected = nredirected;
6475         xlrec.ndead = ndead;
6476
6477         XLogBeginInsert();
6478         XLogRegisterData((char *) &xlrec, SizeOfHeapClean);
6479
6480         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6481
6482         /*
6483          * The OffsetNumber arrays are not actually in the buffer, but we pretend
6484          * that they are.  When XLogInsert stores the whole buffer, the offset
6485          * arrays need not be stored too.  Note that even if all three arrays are
6486          * empty, we want to expose the buffer as a candidate for whole-page
6487          * storage, since this record type implies a defragmentation operation
6488          * even if no item pointers changed state.
6489          */
6490         if (nredirected > 0)
6491                 XLogRegisterBufData(0, (char *) redirected,
6492                                                         nredirected * sizeof(OffsetNumber) * 2);
6493
6494         if (ndead > 0)
6495                 XLogRegisterBufData(0, (char *) nowdead,
6496                                                         ndead * sizeof(OffsetNumber));
6497
6498         if (nunused > 0)
6499                 XLogRegisterBufData(0, (char *) nowunused,
6500                                                         nunused * sizeof(OffsetNumber));
6501
6502         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_CLEAN);
6503
6504         return recptr;
6505 }
6506
6507 /*
6508  * Perform XLogInsert for a heap-freeze operation.  Caller must have already
6509  * modified the buffer and marked it dirty.
6510  */
6511 XLogRecPtr
6512 log_heap_freeze(Relation reln, Buffer buffer, TransactionId cutoff_xid,
6513                                 xl_heap_freeze_tuple *tuples, int ntuples)
6514 {
6515         xl_heap_freeze_page xlrec;
6516         XLogRecPtr      recptr;
6517
6518         /* Caller should not call me on a non-WAL-logged relation */
6519         Assert(RelationNeedsWAL(reln));
6520         /* nor when there are no tuples to freeze */
6521         Assert(ntuples > 0);
6522
6523         xlrec.cutoff_xid = cutoff_xid;
6524         xlrec.ntuples = ntuples;
6525
6526         XLogBeginInsert();
6527         XLogRegisterData((char *) &xlrec, SizeOfHeapFreezePage);
6528
6529         /*
6530          * The freeze plan array is not actually in the buffer, but pretend that
6531          * it is.  When XLogInsert stores the whole buffer, the freeze plan need
6532          * not be stored too.
6533          */
6534         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6535         XLogRegisterBufData(0, (char *) tuples,
6536                                                 ntuples * sizeof(xl_heap_freeze_tuple));
6537
6538         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_FREEZE_PAGE);
6539
6540         return recptr;
6541 }
6542
6543 /*
6544  * Perform XLogInsert for a heap-visible operation.  'block' is the block
6545  * being marked all-visible, and vm_buffer is the buffer containing the
6546  * corresponding visibility map block.  Both should have already been modified
6547  * and dirtied.
6548  *
6549  * If checksums are enabled, we also generate a full-page image of
6550  * heap_buffer, if necessary.
6551  */
6552 XLogRecPtr
6553 log_heap_visible(RelFileNode rnode, Buffer heap_buffer, Buffer vm_buffer,
6554                                  TransactionId cutoff_xid)
6555 {
6556         xl_heap_visible xlrec;
6557         XLogRecPtr      recptr;
6558         uint8           flags;
6559
6560         Assert(BufferIsValid(heap_buffer));
6561         Assert(BufferIsValid(vm_buffer));
6562
6563         xlrec.cutoff_xid = cutoff_xid;
6564         XLogBeginInsert();
6565         XLogRegisterData((char *) &xlrec, SizeOfHeapVisible);
6566
6567         XLogRegisterBuffer(0, vm_buffer, 0);
6568
6569         flags = REGBUF_STANDARD;
6570         if (!XLogHintBitIsNeeded())
6571                 flags |= REGBUF_NO_IMAGE;
6572         XLogRegisterBuffer(1, heap_buffer, flags);
6573
6574         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_VISIBLE);
6575
6576         return recptr;
6577 }
6578
6579 /*
6580  * Perform XLogInsert for a heap-update operation.  Caller must already
6581  * have modified the buffer(s) and marked them dirty.
6582  */
6583 static XLogRecPtr
6584 log_heap_update(Relation reln, Buffer oldbuf,
6585                                 Buffer newbuf, HeapTuple oldtup, HeapTuple newtup,
6586                                 HeapTuple old_key_tuple,
6587                                 bool all_visible_cleared, bool new_all_visible_cleared)
6588 {
6589         xl_heap_update xlrec;
6590         xl_heap_header xlhdr;
6591         xl_heap_header xlhdr_idx;
6592         uint8           info;
6593         uint16          prefix_suffix[2];
6594         uint16          prefixlen = 0,
6595                                 suffixlen = 0;
6596         XLogRecPtr      recptr;
6597         Page            page = BufferGetPage(newbuf);
6598         bool            need_tuple_data = RelationIsLogicallyLogged(reln);
6599         bool            init;
6600         int                     bufflags;
6601
6602         /* Caller should not call me on a non-WAL-logged relation */
6603         Assert(RelationNeedsWAL(reln));
6604
6605         XLogBeginInsert();
6606
6607         if (HeapTupleIsHeapOnly(newtup))
6608                 info = XLOG_HEAP_HOT_UPDATE;
6609         else
6610                 info = XLOG_HEAP_UPDATE;
6611
6612         /*
6613          * If the old and new tuple are on the same page, we only need to log the
6614          * parts of the new tuple that were changed.  That saves on the amount of
6615          * WAL we need to write.  Currently, we just count any unchanged bytes in
6616          * the beginning and end of the tuple.  That's quick to check, and
6617          * perfectly covers the common case that only one field is updated.
6618          *
6619          * We could do this even if the old and new tuple are on different pages,
6620          * but only if we don't make a full-page image of the old page, which is
6621          * difficult to know in advance.  Also, if the old tuple is corrupt for
6622          * some reason, it would allow the corruption to propagate the new page,
6623          * so it seems best to avoid.  Under the general assumption that most
6624          * updates tend to create the new tuple version on the same page, there
6625          * isn't much to be gained by doing this across pages anyway.
6626          *
6627          * Skip this if we're taking a full-page image of the new page, as we
6628          * don't include the new tuple in the WAL record in that case.  Also
6629          * disable if wal_level='logical', as logical decoding needs to be able to
6630          * read the new tuple in whole from the WAL record alone.
6631          */
6632         if (oldbuf == newbuf && !need_tuple_data &&
6633                 !XLogCheckBufferNeedsBackup(newbuf))
6634         {
6635                 char       *oldp = (char *) oldtup->t_data + oldtup->t_data->t_hoff;
6636                 char       *newp = (char *) newtup->t_data + newtup->t_data->t_hoff;
6637                 int                     oldlen = oldtup->t_len - oldtup->t_data->t_hoff;
6638                 int                     newlen = newtup->t_len - newtup->t_data->t_hoff;
6639
6640                 /* Check for common prefix between old and new tuple */
6641                 for (prefixlen = 0; prefixlen < Min(oldlen, newlen); prefixlen++)
6642                 {
6643                         if (newp[prefixlen] != oldp[prefixlen])
6644                                 break;
6645                 }
6646
6647                 /*
6648                  * Storing the length of the prefix takes 2 bytes, so we need to save
6649                  * at least 3 bytes or there's no point.
6650                  */
6651                 if (prefixlen < 3)
6652                         prefixlen = 0;
6653
6654                 /* Same for suffix */
6655                 for (suffixlen = 0; suffixlen < Min(oldlen, newlen) - prefixlen; suffixlen++)
6656                 {
6657                         if (newp[newlen - suffixlen - 1] != oldp[oldlen - suffixlen - 1])
6658                                 break;
6659                 }
6660                 if (suffixlen < 3)
6661                         suffixlen = 0;
6662         }
6663
6664         /* Prepare main WAL data chain */
6665         xlrec.flags = 0;
6666         if (all_visible_cleared)
6667                 xlrec.flags |= XLOG_HEAP_ALL_VISIBLE_CLEARED;
6668         if (new_all_visible_cleared)
6669                 xlrec.flags |= XLOG_HEAP_NEW_ALL_VISIBLE_CLEARED;
6670         if (prefixlen > 0)
6671                 xlrec.flags |= XLOG_HEAP_PREFIX_FROM_OLD;
6672         if (suffixlen > 0)
6673                 xlrec.flags |= XLOG_HEAP_SUFFIX_FROM_OLD;
6674         if (need_tuple_data)
6675         {
6676                 xlrec.flags |= XLOG_HEAP_CONTAINS_NEW_TUPLE;
6677                 if (old_key_tuple)
6678                 {
6679                         if (reln->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
6680                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_TUPLE;
6681                         else
6682                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_KEY;
6683                 }
6684         }
6685
6686         /* If new tuple is the single and first tuple on page... */
6687         if (ItemPointerGetOffsetNumber(&(newtup->t_self)) == FirstOffsetNumber &&
6688                 PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
6689         {
6690                 info |= XLOG_HEAP_INIT_PAGE;
6691                 init = true;
6692         }
6693         else
6694                 init = false;
6695
6696         /* Prepare WAL data for the old page */
6697         xlrec.old_offnum = ItemPointerGetOffsetNumber(&oldtup->t_self);
6698         xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data);
6699         xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask,
6700                                                                                           oldtup->t_data->t_infomask2);
6701
6702         /* Prepare WAL data for the new page */
6703         xlrec.new_offnum = ItemPointerGetOffsetNumber(&newtup->t_self);
6704         xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data);
6705
6706         bufflags = REGBUF_STANDARD;
6707         if (init)
6708                 bufflags |= REGBUF_WILL_INIT;
6709         if (need_tuple_data)
6710                 bufflags |= REGBUF_KEEP_DATA;
6711
6712         XLogRegisterBuffer(0, newbuf, bufflags);
6713         if (oldbuf != newbuf)
6714                 XLogRegisterBuffer(1, oldbuf, REGBUF_STANDARD);
6715
6716         XLogRegisterData((char *) &xlrec, SizeOfHeapUpdate);
6717
6718         /*
6719          * Prepare WAL data for the new tuple.
6720          */
6721         if (prefixlen > 0 || suffixlen > 0)
6722         {
6723                 if (prefixlen > 0 && suffixlen > 0)
6724                 {
6725                         prefix_suffix[0] = prefixlen;
6726                         prefix_suffix[1] = suffixlen;
6727                         XLogRegisterBufData(0, (char *) &prefix_suffix, sizeof(uint16) * 2);
6728                 }
6729                 else if (prefixlen > 0)
6730                 {
6731                         XLogRegisterBufData(0, (char *) &prefixlen, sizeof(uint16));
6732                 }
6733                 else
6734                 {
6735                         XLogRegisterBufData(0, (char *) &suffixlen, sizeof(uint16));
6736                 }
6737         }
6738
6739         xlhdr.t_infomask2 = newtup->t_data->t_infomask2;
6740         xlhdr.t_infomask = newtup->t_data->t_infomask;
6741         xlhdr.t_hoff = newtup->t_data->t_hoff;
6742         Assert(offsetof(HeapTupleHeaderData, t_bits) + prefixlen + suffixlen <= newtup->t_len);
6743
6744         /*
6745          * PG73FORMAT: write bitmap [+ padding] [+ oid] + data
6746          *
6747          * The 'data' doesn't include the common prefix or suffix.
6748          */
6749         XLogRegisterBufData(0, (char *) &xlhdr, SizeOfHeapHeader);
6750         if (prefixlen == 0)
6751         {
6752                 XLogRegisterBufData(0,
6753                    ((char *) newtup->t_data) + offsetof(HeapTupleHeaderData, t_bits),
6754                    newtup->t_len - offsetof(HeapTupleHeaderData, t_bits) -suffixlen);
6755         }
6756         else
6757         {
6758                 /*
6759                  * Have to write the null bitmap and data after the common prefix as
6760                  * two separate rdata entries.
6761                  */
6762                 /* bitmap [+ padding] [+ oid] */
6763                 if (newtup->t_data->t_hoff - offsetof(HeapTupleHeaderData, t_bits) >0)
6764                 {
6765                         XLogRegisterBufData(0,
6766                         ((char *) newtup->t_data) + offsetof(HeapTupleHeaderData, t_bits),
6767                          newtup->t_data->t_hoff - offsetof(HeapTupleHeaderData, t_bits));
6768                 }
6769
6770                 /* data after common prefix */
6771                 XLogRegisterBufData(0,
6772                           ((char *) newtup->t_data) + newtup->t_data->t_hoff + prefixlen,
6773                          newtup->t_len - newtup->t_data->t_hoff - prefixlen - suffixlen);
6774         }
6775
6776         /* We need to log a tuple identity */
6777         if (need_tuple_data && old_key_tuple)
6778         {
6779                 /* don't really need this, but its more comfy to decode */
6780                 xlhdr_idx.t_infomask2 = old_key_tuple->t_data->t_infomask2;
6781                 xlhdr_idx.t_infomask = old_key_tuple->t_data->t_infomask;
6782                 xlhdr_idx.t_hoff = old_key_tuple->t_data->t_hoff;
6783
6784                 XLogRegisterData((char *) &xlhdr_idx, SizeOfHeapHeader);
6785
6786                 /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
6787                 XLogRegisterData((char *) old_key_tuple->t_data + offsetof(HeapTupleHeaderData, t_bits),
6788                            old_key_tuple->t_len - offsetof(HeapTupleHeaderData, t_bits));
6789         }
6790
6791         recptr = XLogInsert(RM_HEAP_ID, info);
6792
6793         return recptr;
6794 }
6795
6796 /*
6797  * Perform XLogInsert of a XLOG_HEAP2_NEW_CID record
6798  *
6799  * This is only used in wal_level >= WAL_LEVEL_LOGICAL, and only for catalog
6800  * tuples.
6801  */
6802 static XLogRecPtr
6803 log_heap_new_cid(Relation relation, HeapTuple tup)
6804 {
6805         xl_heap_new_cid xlrec;
6806
6807         XLogRecPtr      recptr;
6808         HeapTupleHeader hdr = tup->t_data;
6809
6810         Assert(ItemPointerIsValid(&tup->t_self));
6811         Assert(tup->t_tableOid != InvalidOid);
6812
6813         xlrec.top_xid = GetTopTransactionId();
6814         xlrec.target_node = relation->rd_node;
6815         xlrec.target_tid = tup->t_self;
6816
6817         /*
6818          * If the tuple got inserted & deleted in the same TX we definitely have a
6819          * combocid, set cmin and cmax.
6820          */
6821         if (hdr->t_infomask & HEAP_COMBOCID)
6822         {
6823                 Assert(!(hdr->t_infomask & HEAP_XMAX_INVALID));
6824                 Assert(!HeapTupleHeaderXminInvalid(hdr));
6825                 xlrec.cmin = HeapTupleHeaderGetCmin(hdr);
6826                 xlrec.cmax = HeapTupleHeaderGetCmax(hdr);
6827                 xlrec.combocid = HeapTupleHeaderGetRawCommandId(hdr);
6828         }
6829         /* No combocid, so only cmin or cmax can be set by this TX */
6830         else
6831         {
6832                 /*
6833                  * Tuple inserted.
6834                  *
6835                  * We need to check for LOCK ONLY because multixacts might be
6836                  * transferred to the new tuple in case of FOR KEY SHARE updates in
6837                  * which case there will be a xmax, although the tuple just got
6838                  * inserted.
6839                  */
6840                 if (hdr->t_infomask & HEAP_XMAX_INVALID ||
6841                         HEAP_XMAX_IS_LOCKED_ONLY(hdr->t_infomask))
6842                 {
6843                         xlrec.cmin = HeapTupleHeaderGetRawCommandId(hdr);
6844                         xlrec.cmax = InvalidCommandId;
6845                 }
6846                 /* Tuple from a different tx updated or deleted. */
6847                 else
6848                 {
6849                         xlrec.cmin = InvalidCommandId;
6850                         xlrec.cmax = HeapTupleHeaderGetRawCommandId(hdr);
6851
6852                 }
6853                 xlrec.combocid = InvalidCommandId;
6854         }
6855
6856         /*
6857          * Note that we don't need to register the buffer here, because this
6858          * operation does not modify the page. The insert/update/delete that
6859          * called us certainly did, but that's WAL-logged separately.
6860          */
6861         XLogBeginInsert();
6862         XLogRegisterData((char *) &xlrec, SizeOfHeapNewCid);
6863
6864         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_NEW_CID);
6865
6866         return recptr;
6867 }
6868
6869 /*
6870  * Build a heap tuple representing the configured REPLICA IDENTITY to represent
6871  * the old tuple in a UPDATE or DELETE.
6872  *
6873  * Returns NULL if there's no need to log a identity or if there's no suitable
6874  * key in the Relation relation.
6875  */
6876 static HeapTuple
6877 ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_changed, bool *copy)
6878 {
6879         TupleDesc       desc = RelationGetDescr(relation);
6880         Oid                     replidindex;
6881         Relation        idx_rel;
6882         TupleDesc       idx_desc;
6883         char            replident = relation->rd_rel->relreplident;
6884         HeapTuple       key_tuple = NULL;
6885         bool            nulls[MaxHeapAttributeNumber];
6886         Datum           values[MaxHeapAttributeNumber];
6887         int                     natt;
6888
6889         *copy = false;
6890
6891         if (!RelationIsLogicallyLogged(relation))
6892                 return NULL;
6893
6894         if (replident == REPLICA_IDENTITY_NOTHING)
6895                 return NULL;
6896
6897         if (replident == REPLICA_IDENTITY_FULL)
6898         {
6899                 /*
6900                  * When logging the entire old tuple, it very well could contain
6901                  * toasted columns. If so, force them to be inlined.
6902                  */
6903                 if (HeapTupleHasExternal(tp))
6904                 {
6905                         *copy = true;
6906                         tp = toast_flatten_tuple(tp, RelationGetDescr(relation));
6907                 }
6908                 return tp;
6909         }
6910
6911         /* if the key hasn't changed and we're only logging the key, we're done */
6912         if (!key_changed)
6913                 return NULL;
6914
6915         /* find the replica identity index */
6916         replidindex = RelationGetReplicaIndex(relation);
6917         if (!OidIsValid(replidindex))
6918         {
6919                 elog(DEBUG4, "could not find configured replica identity for table \"%s\"",
6920                          RelationGetRelationName(relation));
6921                 return NULL;
6922         }
6923
6924         idx_rel = RelationIdGetRelation(replidindex);
6925         idx_desc = RelationGetDescr(idx_rel);
6926
6927         /* deform tuple, so we have fast access to columns */
6928         heap_deform_tuple(tp, desc, values, nulls);
6929
6930         /* set all columns to NULL, regardless of whether they actually are */
6931         memset(nulls, 1, sizeof(nulls));
6932
6933         /*
6934          * Now set all columns contained in the index to NOT NULL, they cannot
6935          * currently be NULL.
6936          */
6937         for (natt = 0; natt < idx_desc->natts; natt++)
6938         {
6939                 int                     attno = idx_rel->rd_index->indkey.values[natt];
6940
6941                 if (attno < 0)
6942                 {
6943                         /*
6944                          * The OID column can appear in an index definition, but that's
6945                          * OK, becuse we always copy the OID if present (see below). Other
6946                          * system columns may not.
6947                          */
6948                         if (attno == ObjectIdAttributeNumber)
6949                                 continue;
6950                         elog(ERROR, "system column in index");
6951                 }
6952                 nulls[attno - 1] = false;
6953         }
6954
6955         key_tuple = heap_form_tuple(desc, values, nulls);
6956         *copy = true;
6957         RelationClose(idx_rel);
6958
6959         /*
6960          * Always copy oids if the table has them, even if not included in the
6961          * index. The space in the logged tuple is used anyway, so there's little
6962          * point in not including the information.
6963          */
6964         if (relation->rd_rel->relhasoids)
6965                 HeapTupleSetOid(key_tuple, HeapTupleGetOid(tp));
6966
6967         /*
6968          * If the tuple, which by here only contains indexed columns, still has
6969          * toasted columns, force them to be inlined. This is somewhat unlikely
6970          * since there's limits on the size of indexed columns, so we don't
6971          * duplicate toast_flatten_tuple()s functionality in the above loop over
6972          * the indexed columns, even if it would be more efficient.
6973          */
6974         if (HeapTupleHasExternal(key_tuple))
6975         {
6976                 HeapTuple       oldtup = key_tuple;
6977
6978                 key_tuple = toast_flatten_tuple(oldtup, RelationGetDescr(relation));
6979                 heap_freetuple(oldtup);
6980         }
6981
6982         return key_tuple;
6983 }
6984
6985 /*
6986  * Handles CLEANUP_INFO
6987  */
6988 static void
6989 heap_xlog_cleanup_info(XLogReaderState *record)
6990 {
6991         xl_heap_cleanup_info *xlrec = (xl_heap_cleanup_info *) XLogRecGetData(record);
6992
6993         if (InHotStandby)
6994                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, xlrec->node);
6995
6996         /*
6997          * Actual operation is a no-op. Record type exists to provide a means for
6998          * conflict processing to occur before we begin index vacuum actions. see
6999          * vacuumlazy.c and also comments in btvacuumpage()
7000          */
7001
7002         /* Backup blocks are not used in cleanup_info records */
7003         Assert(!XLogRecHasAnyBlockRefs(record));
7004 }
7005
7006 /*
7007  * Handles HEAP2_CLEAN record type
7008  */
7009 static void
7010 heap_xlog_clean(XLogReaderState *record)
7011 {
7012         XLogRecPtr      lsn = record->EndRecPtr;
7013         xl_heap_clean *xlrec = (xl_heap_clean *) XLogRecGetData(record);
7014         Buffer          buffer;
7015         Size            freespace = 0;
7016         RelFileNode rnode;
7017         BlockNumber blkno;
7018         XLogRedoAction action;
7019
7020         XLogRecGetBlockTag(record, 0, &rnode, NULL, &blkno);
7021
7022         /*
7023          * We're about to remove tuples. In Hot Standby mode, ensure that there's
7024          * no queries running for which the removed tuples are still visible.
7025          *
7026          * Not all HEAP2_CLEAN records remove tuples with xids, so we only want to
7027          * conflict on the records that cause MVCC failures for user queries. If
7028          * latestRemovedXid is invalid, skip conflict processing.
7029          */
7030         if (InHotStandby && TransactionIdIsValid(xlrec->latestRemovedXid))
7031                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, rnode);
7032
7033         /*
7034          * If we have a full-page image, restore it (using a cleanup lock) and
7035          * we're done.
7036          */
7037         action = XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true,
7038                                                                                    &buffer);
7039         if (action == BLK_NEEDS_REDO)
7040         {
7041                 Page            page = (Page) BufferGetPage(buffer);
7042                 OffsetNumber *end;
7043                 OffsetNumber *redirected;
7044                 OffsetNumber *nowdead;
7045                 OffsetNumber *nowunused;
7046                 int                     nredirected;
7047                 int                     ndead;
7048                 int                     nunused;
7049                 Size            datalen;
7050
7051                 redirected = (OffsetNumber *) XLogRecGetBlockData(record, 0, &datalen);
7052
7053                 nredirected = xlrec->nredirected;
7054                 ndead = xlrec->ndead;
7055                 end = (OffsetNumber *) ((char *) redirected + datalen);
7056                 nowdead = redirected + (nredirected * 2);
7057                 nowunused = nowdead + ndead;
7058                 nunused = (end - nowunused);
7059                 Assert(nunused >= 0);
7060
7061                 /* Update all item pointers per the record, and repair fragmentation */
7062                 heap_page_prune_execute(buffer,
7063                                                                 redirected, nredirected,
7064                                                                 nowdead, ndead,
7065                                                                 nowunused, nunused);
7066
7067                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7068
7069                 /*
7070                  * Note: we don't worry about updating the page's prunability hints.
7071                  * At worst this will cause an extra prune cycle to occur soon.
7072                  */
7073
7074                 PageSetLSN(page, lsn);
7075                 MarkBufferDirty(buffer);
7076         }
7077         if (BufferIsValid(buffer))
7078                 UnlockReleaseBuffer(buffer);
7079
7080         /*
7081          * Update the FSM as well.
7082          *
7083          * XXX: Don't do this if the page was restored from full page image. We
7084          * don't bother to update the FSM in that case, it doesn't need to be
7085          * totally accurate anyway.
7086          */
7087         if (action == BLK_NEEDS_REDO)
7088                 XLogRecordPageWithFreeSpace(rnode, blkno, freespace);
7089 }
7090
7091 /*
7092  * Replay XLOG_HEAP2_VISIBLE record.
7093  *
7094  * The critical integrity requirement here is that we must never end up with
7095  * a situation where the visibility map bit is set, and the page-level
7096  * PD_ALL_VISIBLE bit is clear.  If that were to occur, then a subsequent
7097  * page modification would fail to clear the visibility map bit.
7098  */
7099 static void
7100 heap_xlog_visible(XLogReaderState *record)
7101 {
7102         XLogRecPtr      lsn = record->EndRecPtr;
7103         xl_heap_visible *xlrec = (xl_heap_visible *) XLogRecGetData(record);
7104         Buffer          vmbuffer = InvalidBuffer;
7105         Buffer          buffer;
7106         Page            page;
7107         RelFileNode rnode;
7108         BlockNumber blkno;
7109         XLogRedoAction action;
7110
7111         XLogRecGetBlockTag(record, 1, &rnode, NULL, &blkno);
7112
7113         /*
7114          * If there are any Hot Standby transactions running that have an xmin
7115          * horizon old enough that this page isn't all-visible for them, they
7116          * might incorrectly decide that an index-only scan can skip a heap fetch.
7117          *
7118          * NB: It might be better to throw some kind of "soft" conflict here that
7119          * forces any index-only scan that is in flight to perform heap fetches,
7120          * rather than killing the transaction outright.
7121          */
7122         if (InHotStandby)
7123                 ResolveRecoveryConflictWithSnapshot(xlrec->cutoff_xid, rnode);
7124
7125         /*
7126          * Read the heap page, if it still exists. If the heap file has dropped or
7127          * truncated later in recovery, we don't need to update the page, but we'd
7128          * better still update the visibility map.
7129          */
7130         action = XLogReadBufferForRedo(record, 1, &buffer);
7131         if (action == BLK_NEEDS_REDO)
7132         {
7133                 /*
7134                  * We don't bump the LSN of the heap page when setting the visibility
7135                  * map bit (unless checksums are enabled, in which case we must),
7136                  * because that would generate an unworkable volume of full-page
7137                  * writes.  This exposes us to torn page hazards, but since we're not
7138                  * inspecting the existing page contents in any way, we don't care.
7139                  *
7140                  * However, all operations that clear the visibility map bit *do* bump
7141                  * the LSN, and those operations will only be replayed if the XLOG LSN
7142                  * follows the page LSN.  Thus, if the page LSN has advanced past our
7143                  * XLOG record's LSN, we mustn't mark the page all-visible, because
7144                  * the subsequent update won't be replayed to clear the flag.
7145                  */
7146                 page = BufferGetPage(buffer);
7147                 PageSetAllVisible(page);
7148                 MarkBufferDirty(buffer);
7149         }
7150         else if (action == BLK_RESTORED)
7151         {
7152                 /*
7153                  * If heap block was backed up, restore it. This can only happen with
7154                  * checksums enabled.
7155                  */
7156                 Assert(DataChecksumsEnabled());
7157         }
7158         if (BufferIsValid(buffer))
7159                 UnlockReleaseBuffer(buffer);
7160
7161         /*
7162          * Even if we skipped the heap page update due to the LSN interlock, it's
7163          * still safe to update the visibility map.  Any WAL record that clears
7164          * the visibility map bit does so before checking the page LSN, so any
7165          * bits that need to be cleared will still be cleared.
7166          */
7167         if (XLogReadBufferForRedoExtended(record, 0, RBM_ZERO_ON_ERROR, false,
7168                                                                           &vmbuffer) == BLK_NEEDS_REDO)
7169         {
7170                 Page            vmpage = BufferGetPage(vmbuffer);
7171                 Relation        reln;
7172
7173                 /* initialize the page if it was read as zeros */
7174                 if (PageIsNew(vmpage))
7175                         PageInit(vmpage, BLCKSZ, 0);
7176
7177                 /*
7178                  * XLogReplayBufferExtended locked the buffer. But visibilitymap_set
7179                  * will handle locking itself.
7180                  */
7181                 LockBuffer(vmbuffer, BUFFER_LOCK_UNLOCK);
7182
7183                 reln = CreateFakeRelcacheEntry(rnode);
7184                 visibilitymap_pin(reln, blkno, &vmbuffer);
7185
7186                 /*
7187                  * Don't set the bit if replay has already passed this point.
7188                  *
7189                  * It might be safe to do this unconditionally; if replay has passed
7190                  * this point, we'll replay at least as far this time as we did
7191                  * before, and if this bit needs to be cleared, the record responsible
7192                  * for doing so should be again replayed, and clear it.  For right
7193                  * now, out of an abundance of conservatism, we use the same test here
7194                  * we did for the heap page.  If this results in a dropped bit, no
7195                  * real harm is done; and the next VACUUM will fix it.
7196                  */
7197                 if (lsn > PageGetLSN(vmpage))
7198                         visibilitymap_set(reln, blkno, InvalidBuffer, lsn, vmbuffer,
7199                                                           xlrec->cutoff_xid);
7200
7201                 ReleaseBuffer(vmbuffer);
7202                 FreeFakeRelcacheEntry(reln);
7203         }
7204         else if (BufferIsValid(vmbuffer))
7205                 UnlockReleaseBuffer(vmbuffer);
7206 }
7207
7208 /*
7209  * Replay XLOG_HEAP2_FREEZE_PAGE records
7210  */
7211 static void
7212 heap_xlog_freeze_page(XLogReaderState *record)
7213 {
7214         XLogRecPtr      lsn = record->EndRecPtr;
7215         xl_heap_freeze_page *xlrec = (xl_heap_freeze_page *) XLogRecGetData(record);
7216         TransactionId cutoff_xid = xlrec->cutoff_xid;
7217         Buffer          buffer;
7218         int                     ntup;
7219
7220         /*
7221          * In Hot Standby mode, ensure that there's no queries running which still
7222          * consider the frozen xids as running.
7223          */
7224         if (InHotStandby)
7225         {
7226                 RelFileNode rnode;
7227
7228                 XLogRecGetBlockTag(record, 0, &rnode, NULL, NULL);
7229                 ResolveRecoveryConflictWithSnapshot(cutoff_xid, rnode);
7230         }
7231
7232         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
7233         {
7234                 Page            page = BufferGetPage(buffer);
7235                 xl_heap_freeze_tuple *tuples;
7236
7237                 tuples = (xl_heap_freeze_tuple *) XLogRecGetBlockData(record, 0, NULL);
7238
7239                 /* now execute freeze plan for each frozen tuple */
7240                 for (ntup = 0; ntup < xlrec->ntuples; ntup++)
7241                 {
7242                         xl_heap_freeze_tuple *xlrec_tp;
7243                         ItemId          lp;
7244                         HeapTupleHeader tuple;
7245
7246                         xlrec_tp = &tuples[ntup];
7247                         lp = PageGetItemId(page, xlrec_tp->offset); /* offsets are one-based */
7248                         tuple = (HeapTupleHeader) PageGetItem(page, lp);
7249
7250                         heap_execute_freeze_tuple(tuple, xlrec_tp);
7251                 }
7252
7253                 PageSetLSN(page, lsn);
7254                 MarkBufferDirty(buffer);
7255         }
7256         if (BufferIsValid(buffer))
7257                 UnlockReleaseBuffer(buffer);
7258 }
7259
7260 /*
7261  * Given an "infobits" field from an XLog record, set the correct bits in the
7262  * given infomask and infomask2 for the tuple touched by the record.
7263  *
7264  * (This is the reverse of compute_infobits).
7265  */
7266 static void
7267 fix_infomask_from_infobits(uint8 infobits, uint16 *infomask, uint16 *infomask2)
7268 {
7269         *infomask &= ~(HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY |
7270                                    HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_EXCL_LOCK);
7271         *infomask2 &= ~HEAP_KEYS_UPDATED;
7272
7273         if (infobits & XLHL_XMAX_IS_MULTI)
7274                 *infomask |= HEAP_XMAX_IS_MULTI;
7275         if (infobits & XLHL_XMAX_LOCK_ONLY)
7276                 *infomask |= HEAP_XMAX_LOCK_ONLY;
7277         if (infobits & XLHL_XMAX_EXCL_LOCK)
7278                 *infomask |= HEAP_XMAX_EXCL_LOCK;
7279         /* note HEAP_XMAX_SHR_LOCK isn't considered here */
7280         if (infobits & XLHL_XMAX_KEYSHR_LOCK)
7281                 *infomask |= HEAP_XMAX_KEYSHR_LOCK;
7282
7283         if (infobits & XLHL_KEYS_UPDATED)
7284                 *infomask2 |= HEAP_KEYS_UPDATED;
7285 }
7286
7287 static void
7288 heap_xlog_delete(XLogReaderState *record)
7289 {
7290         XLogRecPtr      lsn = record->EndRecPtr;
7291         xl_heap_delete *xlrec = (xl_heap_delete *) XLogRecGetData(record);
7292         Buffer          buffer;
7293         Page            page;
7294         ItemId          lp = NULL;
7295         HeapTupleHeader htup;
7296         BlockNumber blkno;
7297         RelFileNode target_node;
7298         ItemPointerData target_tid;
7299
7300         XLogRecGetBlockTag(record, 0, &target_node, NULL, &blkno);
7301         ItemPointerSetBlockNumber(&target_tid, blkno);
7302         ItemPointerSetOffsetNumber(&target_tid, xlrec->offnum);
7303
7304         /*
7305          * The visibility map may need to be fixed even if the heap page is
7306          * already up-to-date.
7307          */
7308         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7309         {
7310                 Relation        reln = CreateFakeRelcacheEntry(target_node);
7311                 Buffer          vmbuffer = InvalidBuffer;
7312
7313                 visibilitymap_pin(reln, blkno, &vmbuffer);
7314                 visibilitymap_clear(reln, blkno, vmbuffer);
7315                 ReleaseBuffer(vmbuffer);
7316                 FreeFakeRelcacheEntry(reln);
7317         }
7318
7319         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
7320         {
7321                 page = BufferGetPage(buffer);
7322
7323                 if (PageGetMaxOffsetNumber(page) >= xlrec->offnum)
7324                         lp = PageGetItemId(page, xlrec->offnum);
7325
7326                 if (PageGetMaxOffsetNumber(page) < xlrec->offnum || !ItemIdIsNormal(lp))
7327                         elog(PANIC, "heap_delete_redo: invalid lp");
7328
7329                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7330
7331                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
7332                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7333                 HeapTupleHeaderClearHotUpdated(htup);
7334                 fix_infomask_from_infobits(xlrec->infobits_set,
7335                                                                    &htup->t_infomask, &htup->t_infomask2);
7336                 HeapTupleHeaderSetXmax(htup, xlrec->xmax);
7337                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
7338
7339                 /* Mark the page as a candidate for pruning */
7340                 PageSetPrunable(page, XLogRecGetXid(record));
7341
7342                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7343                         PageClearAllVisible(page);
7344
7345                 /* Make sure there is no forward chain link in t_ctid */
7346                 htup->t_ctid = target_tid;
7347                 PageSetLSN(page, lsn);
7348                 MarkBufferDirty(buffer);
7349         }
7350         if (BufferIsValid(buffer))
7351                 UnlockReleaseBuffer(buffer);
7352 }
7353
7354 static void
7355 heap_xlog_insert(XLogReaderState *record)
7356 {
7357         XLogRecPtr      lsn = record->EndRecPtr;
7358         xl_heap_insert *xlrec = (xl_heap_insert *) XLogRecGetData(record);
7359         Buffer          buffer;
7360         Page            page;
7361         struct
7362         {
7363                 HeapTupleHeaderData hdr;
7364                 char            data[MaxHeapTupleSize];
7365         }                       tbuf;
7366         HeapTupleHeader htup;
7367         xl_heap_header xlhdr;
7368         uint32          newlen;
7369         Size            freespace = 0;
7370         RelFileNode target_node;
7371         BlockNumber blkno;
7372         ItemPointerData target_tid;
7373         XLogRedoAction action;
7374
7375         XLogRecGetBlockTag(record, 0, &target_node, NULL, &blkno);
7376         ItemPointerSetBlockNumber(&target_tid, blkno);
7377         ItemPointerSetOffsetNumber(&target_tid, xlrec->offnum);
7378
7379         /*
7380          * The visibility map may need to be fixed even if the heap page is
7381          * already up-to-date.
7382          */
7383         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7384         {
7385                 Relation        reln = CreateFakeRelcacheEntry(target_node);
7386                 Buffer          vmbuffer = InvalidBuffer;
7387
7388                 visibilitymap_pin(reln, blkno, &vmbuffer);
7389                 visibilitymap_clear(reln, blkno, vmbuffer);
7390                 ReleaseBuffer(vmbuffer);
7391                 FreeFakeRelcacheEntry(reln);
7392         }
7393
7394         /*
7395          * If we inserted the first and only tuple on the page, re-initialize the
7396          * page from scratch.
7397          */
7398         if (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE)
7399         {
7400                 buffer = XLogInitBufferForRedo(record, 0);
7401                 page = BufferGetPage(buffer);
7402                 PageInit(page, BufferGetPageSize(buffer), 0);
7403                 action = BLK_NEEDS_REDO;
7404         }
7405         else
7406                 action = XLogReadBufferForRedo(record, 0, &buffer);
7407         if (action == BLK_NEEDS_REDO)
7408         {
7409                 Size            datalen;
7410                 char       *data;
7411
7412                 page = BufferGetPage(buffer);
7413
7414                 if (PageGetMaxOffsetNumber(page) + 1 < xlrec->offnum)
7415                         elog(PANIC, "heap_insert_redo: invalid max offset number");
7416
7417                 data = XLogRecGetBlockData(record, 0, &datalen);
7418
7419                 newlen = datalen - SizeOfHeapHeader;
7420                 Assert(datalen > SizeOfHeapHeader && newlen <= MaxHeapTupleSize);
7421                 memcpy((char *) &xlhdr, data, SizeOfHeapHeader);
7422                 data += SizeOfHeapHeader;
7423
7424                 htup = &tbuf.hdr;
7425                 MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData));
7426                 /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
7427                 memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits),
7428                            data,
7429                            newlen);
7430                 newlen += offsetof(HeapTupleHeaderData, t_bits);
7431                 htup->t_infomask2 = xlhdr.t_infomask2;
7432                 htup->t_infomask = xlhdr.t_infomask;
7433                 htup->t_hoff = xlhdr.t_hoff;
7434                 HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
7435                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
7436                 htup->t_ctid = target_tid;
7437
7438                 if (PageAddItem(page, (Item) htup, newlen, xlrec->offnum,
7439                                                 true, true) == InvalidOffsetNumber)
7440                         elog(PANIC, "heap_insert_redo: failed to add tuple");
7441
7442                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7443
7444                 PageSetLSN(page, lsn);
7445
7446                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7447                         PageClearAllVisible(page);
7448
7449                 MarkBufferDirty(buffer);
7450         }
7451         if (BufferIsValid(buffer))
7452                 UnlockReleaseBuffer(buffer);
7453
7454         /*
7455          * If the page is running low on free space, update the FSM as well.
7456          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
7457          * better than that without knowing the fill-factor for the table.
7458          *
7459          * XXX: Don't do this if the page was restored from full page image. We
7460          * don't bother to update the FSM in that case, it doesn't need to be
7461          * totally accurate anyway.
7462          */
7463         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
7464                 XLogRecordPageWithFreeSpace(target_node, blkno, freespace);
7465 }
7466
7467 /*
7468  * Handles MULTI_INSERT record type.
7469  */
7470 static void
7471 heap_xlog_multi_insert(XLogReaderState *record)
7472 {
7473         XLogRecPtr      lsn = record->EndRecPtr;
7474         xl_heap_multi_insert *xlrec;
7475         RelFileNode rnode;
7476         BlockNumber blkno;
7477         Buffer          buffer;
7478         Page            page;
7479         struct
7480         {
7481                 HeapTupleHeaderData hdr;
7482                 char            data[MaxHeapTupleSize];
7483         }                       tbuf;
7484         HeapTupleHeader htup;
7485         uint32          newlen;
7486         Size            freespace = 0;
7487         int                     i;
7488         bool            isinit = (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE) != 0;
7489         XLogRedoAction action;
7490
7491         /*
7492          * Insertion doesn't overwrite MVCC data, so no conflict processing is
7493          * required.
7494          */
7495         xlrec = (xl_heap_multi_insert *) XLogRecGetData(record);
7496
7497         XLogRecGetBlockTag(record, 0, &rnode, NULL, &blkno);
7498
7499         /*
7500          * The visibility map may need to be fixed even if the heap page is
7501          * already up-to-date.
7502          */
7503         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7504         {
7505                 Relation        reln = CreateFakeRelcacheEntry(rnode);
7506                 Buffer          vmbuffer = InvalidBuffer;
7507
7508                 visibilitymap_pin(reln, blkno, &vmbuffer);
7509                 visibilitymap_clear(reln, blkno, vmbuffer);
7510                 ReleaseBuffer(vmbuffer);
7511                 FreeFakeRelcacheEntry(reln);
7512         }
7513
7514         if (isinit)
7515         {
7516                 buffer = XLogInitBufferForRedo(record, 0);
7517                 page = BufferGetPage(buffer);
7518                 PageInit(page, BufferGetPageSize(buffer), 0);
7519                 action = BLK_NEEDS_REDO;
7520         }
7521         else
7522                 action = XLogReadBufferForRedo(record, 0, &buffer);
7523         if (action == BLK_NEEDS_REDO)
7524         {
7525                 char       *tupdata;
7526                 char       *endptr;
7527                 Size            len;
7528
7529                 /* Tuples are stored as block data */
7530                 tupdata = XLogRecGetBlockData(record, 0, &len);
7531                 endptr = tupdata + len;
7532
7533                 page = (Page) BufferGetPage(buffer);
7534
7535                 for (i = 0; i < xlrec->ntuples; i++)
7536                 {
7537                         OffsetNumber offnum;
7538                         xl_multi_insert_tuple *xlhdr;
7539
7540                         /*
7541                          * If we're reinitializing the page, the tuples are stored in
7542                          * order from FirstOffsetNumber. Otherwise there's an array of
7543                          * offsets in the WAL record, and the tuples come after that.
7544                          */
7545                         if (isinit)
7546                                 offnum = FirstOffsetNumber + i;
7547                         else
7548                                 offnum = xlrec->offsets[i];
7549                         if (PageGetMaxOffsetNumber(page) + 1 < offnum)
7550                                 elog(PANIC, "heap_multi_insert_redo: invalid max offset number");
7551
7552                         xlhdr = (xl_multi_insert_tuple *) SHORTALIGN(tupdata);
7553                         tupdata = ((char *) xlhdr) + SizeOfMultiInsertTuple;
7554
7555                         newlen = xlhdr->datalen;
7556                         Assert(newlen <= MaxHeapTupleSize);
7557                         htup = &tbuf.hdr;
7558                         MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData));
7559                         /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
7560                         memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits),
7561                                    (char *) tupdata,
7562                                    newlen);
7563                         tupdata += newlen;
7564
7565                         newlen += offsetof(HeapTupleHeaderData, t_bits);
7566                         htup->t_infomask2 = xlhdr->t_infomask2;
7567                         htup->t_infomask = xlhdr->t_infomask;
7568                         htup->t_hoff = xlhdr->t_hoff;
7569                         HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
7570                         HeapTupleHeaderSetCmin(htup, FirstCommandId);
7571                         ItemPointerSetBlockNumber(&htup->t_ctid, blkno);
7572                         ItemPointerSetOffsetNumber(&htup->t_ctid, offnum);
7573
7574                         offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
7575                         if (offnum == InvalidOffsetNumber)
7576                                 elog(PANIC, "heap_multi_insert_redo: failed to add tuple");
7577                 }
7578                 if (tupdata != endptr)
7579                         elog(PANIC, "heap_multi_insert_redo: total tuple length mismatch");
7580
7581                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7582
7583                 PageSetLSN(page, lsn);
7584
7585                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7586                         PageClearAllVisible(page);
7587
7588                 MarkBufferDirty(buffer);
7589         }
7590         if (BufferIsValid(buffer))
7591                 UnlockReleaseBuffer(buffer);
7592
7593         /*
7594          * If the page is running low on free space, update the FSM as well.
7595          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
7596          * better than that without knowing the fill-factor for the table.
7597          *
7598          * XXX: Don't do this if the page was restored from full page image. We
7599          * don't bother to update the FSM in that case, it doesn't need to be
7600          * totally accurate anyway.
7601          */
7602         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
7603                 XLogRecordPageWithFreeSpace(rnode, blkno, freespace);
7604 }
7605
7606 /*
7607  * Handles UPDATE and HOT_UPDATE
7608  */
7609 static void
7610 heap_xlog_update(XLogReaderState *record, bool hot_update)
7611 {
7612         XLogRecPtr      lsn = record->EndRecPtr;
7613         xl_heap_update *xlrec = (xl_heap_update *) XLogRecGetData(record);
7614         RelFileNode rnode;
7615         BlockNumber oldblk;
7616         BlockNumber newblk;
7617         ItemPointerData newtid;
7618         Buffer          obuffer,
7619                                 nbuffer;
7620         Page            page;
7621         OffsetNumber offnum;
7622         ItemId          lp = NULL;
7623         HeapTupleData oldtup;
7624         HeapTupleHeader htup;
7625         uint16          prefixlen = 0,
7626                                 suffixlen = 0;
7627         char       *newp;
7628         struct
7629         {
7630                 HeapTupleHeaderData hdr;
7631                 char            data[MaxHeapTupleSize];
7632         }                       tbuf;
7633         xl_heap_header xlhdr;
7634         uint32          newlen;
7635         Size            freespace = 0;
7636         XLogRedoAction oldaction;
7637         XLogRedoAction newaction;
7638
7639         /* initialize to keep the compiler quiet */
7640         oldtup.t_data = NULL;
7641         oldtup.t_len = 0;
7642
7643         XLogRecGetBlockTag(record, 0, &rnode, NULL, &newblk);
7644         if (XLogRecGetBlockTag(record, 1, NULL, NULL, &oldblk))
7645         {
7646                 /* HOT updates are never done across pages */
7647                 Assert(!hot_update);
7648         }
7649         else
7650                 oldblk = newblk;
7651
7652         ItemPointerSet(&newtid, newblk, xlrec->new_offnum);
7653
7654         /*
7655          * The visibility map may need to be fixed even if the heap page is
7656          * already up-to-date.
7657          */
7658         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7659         {
7660                 Relation        reln = CreateFakeRelcacheEntry(rnode);
7661                 Buffer          vmbuffer = InvalidBuffer;
7662
7663                 visibilitymap_pin(reln, oldblk, &vmbuffer);
7664                 visibilitymap_clear(reln, oldblk, vmbuffer);
7665                 ReleaseBuffer(vmbuffer);
7666                 FreeFakeRelcacheEntry(reln);
7667         }
7668
7669         /*
7670          * In normal operation, it is important to lock the two pages in
7671          * page-number order, to avoid possible deadlocks against other update
7672          * operations going the other way.  However, during WAL replay there can
7673          * be no other update happening, so we don't need to worry about that. But
7674          * we *do* need to worry that we don't expose an inconsistent state to Hot
7675          * Standby queries --- so the original page can't be unlocked before we've
7676          * added the new tuple to the new page.
7677          */
7678
7679         /* Deal with old tuple version */
7680         oldaction = XLogReadBufferForRedo(record, (oldblk == newblk) ? 0 : 1,
7681                                                                           &obuffer);
7682         if (oldaction == BLK_NEEDS_REDO)
7683         {
7684                 page = BufferGetPage(obuffer);
7685                 offnum = xlrec->old_offnum;
7686                 if (PageGetMaxOffsetNumber(page) >= offnum)
7687                         lp = PageGetItemId(page, offnum);
7688
7689                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
7690                         elog(PANIC, "heap_update_redo: invalid lp");
7691
7692                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7693
7694                 oldtup.t_data = htup;
7695                 oldtup.t_len = ItemIdGetLength(lp);
7696
7697                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
7698                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7699                 if (hot_update)
7700                         HeapTupleHeaderSetHotUpdated(htup);
7701                 else
7702                         HeapTupleHeaderClearHotUpdated(htup);
7703                 fix_infomask_from_infobits(xlrec->old_infobits_set, &htup->t_infomask,
7704                                                                    &htup->t_infomask2);
7705                 HeapTupleHeaderSetXmax(htup, xlrec->old_xmax);
7706                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
7707                 /* Set forward chain link in t_ctid */
7708                 htup->t_ctid = newtid;
7709
7710                 /* Mark the page as a candidate for pruning */
7711                 PageSetPrunable(page, XLogRecGetXid(record));
7712
7713                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7714                         PageClearAllVisible(page);
7715
7716                 PageSetLSN(page, lsn);
7717                 MarkBufferDirty(obuffer);
7718         }
7719
7720         /*
7721          * Read the page the new tuple goes into, if different from old.
7722          */
7723         if (oldblk == newblk)
7724         {
7725                 nbuffer = obuffer;
7726                 newaction = oldaction;
7727         }
7728         else if (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE)
7729         {
7730                 nbuffer = XLogInitBufferForRedo(record, 0);
7731                 page = (Page) BufferGetPage(nbuffer);
7732                 PageInit(page, BufferGetPageSize(nbuffer), 0);
7733                 newaction = BLK_NEEDS_REDO;
7734         }
7735         else
7736                 newaction = XLogReadBufferForRedo(record, 0, &nbuffer);
7737
7738         /*
7739          * The visibility map may need to be fixed even if the heap page is
7740          * already up-to-date.
7741          */
7742         if (xlrec->flags & XLOG_HEAP_NEW_ALL_VISIBLE_CLEARED)
7743         {
7744                 Relation        reln = CreateFakeRelcacheEntry(rnode);
7745                 Buffer          vmbuffer = InvalidBuffer;
7746
7747                 visibilitymap_pin(reln, newblk, &vmbuffer);
7748                 visibilitymap_clear(reln, newblk, vmbuffer);
7749                 ReleaseBuffer(vmbuffer);
7750                 FreeFakeRelcacheEntry(reln);
7751         }
7752
7753         /* Deal with new tuple */
7754         if (newaction == BLK_NEEDS_REDO)
7755         {
7756                 char       *recdata;
7757                 char       *recdata_end;
7758                 Size            datalen;
7759                 Size            tuplen;
7760
7761                 recdata = XLogRecGetBlockData(record, 0, &datalen);
7762                 recdata_end = recdata + datalen;
7763
7764                 page = BufferGetPage(nbuffer);
7765
7766                 offnum = xlrec->new_offnum;
7767                 if (PageGetMaxOffsetNumber(page) + 1 < offnum)
7768                         elog(PANIC, "heap_update_redo: invalid max offset number");
7769
7770                 if (xlrec->flags & XLOG_HEAP_PREFIX_FROM_OLD)
7771                 {
7772                         Assert(newblk == oldblk);
7773                         memcpy(&prefixlen, recdata, sizeof(uint16));
7774                         recdata += sizeof(uint16);
7775                 }
7776                 if (xlrec->flags & XLOG_HEAP_SUFFIX_FROM_OLD)
7777                 {
7778                         Assert(newblk == oldblk);
7779                         memcpy(&suffixlen, recdata, sizeof(uint16));
7780                         recdata += sizeof(uint16);
7781                 }
7782
7783                 memcpy((char *) &xlhdr, recdata, SizeOfHeapHeader);
7784                 recdata += SizeOfHeapHeader;
7785
7786                 tuplen = recdata_end - recdata;
7787                 Assert(tuplen <= MaxHeapTupleSize);
7788
7789                 htup = &tbuf.hdr;
7790                 MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData));
7791
7792                 /*
7793                  * Reconstruct the new tuple using the prefix and/or suffix from the
7794                  * old tuple, and the data stored in the WAL record.
7795                  */
7796                 newp = (char *) htup + offsetof(HeapTupleHeaderData, t_bits);
7797                 if (prefixlen > 0)
7798                 {
7799                         int                     len;
7800
7801                         /* copy bitmap [+ padding] [+ oid] from WAL record */
7802                         len = xlhdr.t_hoff - offsetof(HeapTupleHeaderData, t_bits);
7803                         memcpy(newp, recdata, len);
7804                         recdata += len;
7805                         newp += len;
7806
7807                         /* copy prefix from old tuple */
7808                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_data->t_hoff, prefixlen);
7809                         newp += prefixlen;
7810
7811                         /* copy new tuple data from WAL record */
7812                         len = tuplen - (xlhdr.t_hoff - offsetof(HeapTupleHeaderData, t_bits));
7813                         memcpy(newp, recdata, len);
7814                         recdata += len;
7815                         newp += len;
7816                 }
7817                 else
7818                 {
7819                         /*
7820                          * copy bitmap [+ padding] [+ oid] + data from record, all in one
7821                          * go
7822                          */
7823                         memcpy(newp, recdata, tuplen);
7824                         recdata += tuplen;
7825                         newp += tuplen;
7826                 }
7827                 Assert(recdata == recdata_end);
7828
7829                 /* copy suffix from old tuple */
7830                 if (suffixlen > 0)
7831                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_len - suffixlen, suffixlen);
7832
7833                 newlen = offsetof(HeapTupleHeaderData, t_bits) + tuplen + prefixlen + suffixlen;
7834                 htup->t_infomask2 = xlhdr.t_infomask2;
7835                 htup->t_infomask = xlhdr.t_infomask;
7836                 htup->t_hoff = xlhdr.t_hoff;
7837
7838                 HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
7839                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
7840                 HeapTupleHeaderSetXmax(htup, xlrec->new_xmax);
7841                 /* Make sure there is no forward chain link in t_ctid */
7842                 htup->t_ctid = newtid;
7843
7844                 offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
7845                 if (offnum == InvalidOffsetNumber)
7846                         elog(PANIC, "heap_update_redo: failed to add tuple");
7847
7848                 if (xlrec->flags & XLOG_HEAP_NEW_ALL_VISIBLE_CLEARED)
7849                         PageClearAllVisible(page);
7850
7851                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7852
7853                 PageSetLSN(page, lsn);
7854                 MarkBufferDirty(nbuffer);
7855         }
7856
7857         if (BufferIsValid(nbuffer) && nbuffer != obuffer)
7858                 UnlockReleaseBuffer(nbuffer);
7859         if (BufferIsValid(obuffer))
7860                 UnlockReleaseBuffer(obuffer);
7861
7862         /*
7863          * If the new page is running low on free space, update the FSM as well.
7864          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
7865          * better than that without knowing the fill-factor for the table.
7866          *
7867          * However, don't update the FSM on HOT updates, because after crash
7868          * recovery, either the old or the new tuple will certainly be dead and
7869          * prunable. After pruning, the page will have roughly as much free space
7870          * as it did before the update, assuming the new tuple is about the same
7871          * size as the old one.
7872          *
7873          * XXX: Don't do this if the page was restored from full page image. We
7874          * don't bother to update the FSM in that case, it doesn't need to be
7875          * totally accurate anyway.
7876          */
7877         if (newaction == BLK_NEEDS_REDO && !hot_update && freespace < BLCKSZ / 5)
7878                 XLogRecordPageWithFreeSpace(rnode, newblk, freespace);
7879 }
7880
7881 static void
7882 heap_xlog_lock(XLogReaderState *record)
7883 {
7884         XLogRecPtr      lsn = record->EndRecPtr;
7885         xl_heap_lock *xlrec = (xl_heap_lock *) XLogRecGetData(record);
7886         Buffer          buffer;
7887         Page            page;
7888         OffsetNumber offnum;
7889         ItemId          lp = NULL;
7890         HeapTupleHeader htup;
7891
7892         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
7893         {
7894                 page = (Page) BufferGetPage(buffer);
7895
7896                 offnum = xlrec->offnum;
7897                 if (PageGetMaxOffsetNumber(page) >= offnum)
7898                         lp = PageGetItemId(page, offnum);
7899
7900                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
7901                         elog(PANIC, "heap_lock_redo: invalid lp");
7902
7903                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7904
7905                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
7906                                                                    &htup->t_infomask2);
7907
7908                 /*
7909                  * Clear relevant update flags, but only if the modified infomask says
7910                  * there's no update.
7911                  */
7912                 if (HEAP_XMAX_IS_LOCKED_ONLY(htup->t_infomask))
7913                 {
7914                         HeapTupleHeaderClearHotUpdated(htup);
7915                         /* Make sure there is no forward chain link in t_ctid */
7916                         ItemPointerSet(&htup->t_ctid,
7917                                                    BufferGetBlockNumber(buffer),
7918                                                    offnum);
7919                 }
7920                 HeapTupleHeaderSetXmax(htup, xlrec->locking_xid);
7921                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
7922                 PageSetLSN(page, lsn);
7923                 MarkBufferDirty(buffer);
7924         }
7925         if (BufferIsValid(buffer))
7926                 UnlockReleaseBuffer(buffer);
7927 }
7928
7929 static void
7930 heap_xlog_lock_updated(XLogReaderState *record)
7931 {
7932         XLogRecPtr      lsn = record->EndRecPtr;
7933         xl_heap_lock_updated *xlrec;
7934         Buffer          buffer;
7935         Page            page;
7936         OffsetNumber offnum;
7937         ItemId          lp = NULL;
7938         HeapTupleHeader htup;
7939
7940         xlrec = (xl_heap_lock_updated *) XLogRecGetData(record);
7941
7942         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
7943         {
7944                 page = BufferGetPage(buffer);
7945
7946                 offnum = xlrec->offnum;
7947                 if (PageGetMaxOffsetNumber(page) >= offnum)
7948                         lp = PageGetItemId(page, offnum);
7949
7950                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
7951                         elog(PANIC, "heap_xlog_lock_updated: invalid lp");
7952
7953                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7954
7955                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
7956                                                                    &htup->t_infomask2);
7957                 HeapTupleHeaderSetXmax(htup, xlrec->xmax);
7958
7959                 PageSetLSN(page, lsn);
7960                 MarkBufferDirty(buffer);
7961         }
7962         if (BufferIsValid(buffer))
7963                 UnlockReleaseBuffer(buffer);
7964 }
7965
7966 static void
7967 heap_xlog_inplace(XLogReaderState *record)
7968 {
7969         XLogRecPtr      lsn = record->EndRecPtr;
7970         xl_heap_inplace *xlrec = (xl_heap_inplace *) XLogRecGetData(record);
7971         Buffer          buffer;
7972         Page            page;
7973         OffsetNumber offnum;
7974         ItemId          lp = NULL;
7975         HeapTupleHeader htup;
7976         uint32          oldlen;
7977         Size            newlen;
7978
7979         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
7980         {
7981                 char       *newtup = XLogRecGetBlockData(record, 0, &newlen);
7982
7983                 page = BufferGetPage(buffer);
7984
7985                 offnum = xlrec->offnum;
7986                 if (PageGetMaxOffsetNumber(page) >= offnum)
7987                         lp = PageGetItemId(page, offnum);
7988
7989                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
7990                         elog(PANIC, "heap_inplace_redo: invalid lp");
7991
7992                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7993
7994                 oldlen = ItemIdGetLength(lp) - htup->t_hoff;
7995                 if (oldlen != newlen)
7996                         elog(PANIC, "heap_inplace_redo: wrong tuple length");
7997
7998                 memcpy((char *) htup + htup->t_hoff, newtup, newlen);
7999
8000                 PageSetLSN(page, lsn);
8001                 MarkBufferDirty(buffer);
8002         }
8003         if (BufferIsValid(buffer))
8004                 UnlockReleaseBuffer(buffer);
8005 }
8006
8007 void
8008 heap_redo(XLogReaderState *record)
8009 {
8010         uint8           info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8011
8012         /*
8013          * These operations don't overwrite MVCC data so no conflict processing is
8014          * required. The ones in heap2 rmgr do.
8015          */
8016
8017         switch (info & XLOG_HEAP_OPMASK)
8018         {
8019                 case XLOG_HEAP_INSERT:
8020                         heap_xlog_insert(record);
8021                         break;
8022                 case XLOG_HEAP_DELETE:
8023                         heap_xlog_delete(record);
8024                         break;
8025                 case XLOG_HEAP_UPDATE:
8026                         heap_xlog_update(record, false);
8027                         break;
8028                 case XLOG_HEAP_HOT_UPDATE:
8029                         heap_xlog_update(record, true);
8030                         break;
8031                 case XLOG_HEAP_LOCK:
8032                         heap_xlog_lock(record);
8033                         break;
8034                 case XLOG_HEAP_INPLACE:
8035                         heap_xlog_inplace(record);
8036                         break;
8037                 default:
8038                         elog(PANIC, "heap_redo: unknown op code %u", info);
8039         }
8040 }
8041
8042 void
8043 heap2_redo(XLogReaderState *record)
8044 {
8045         uint8           info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8046
8047         switch (info & XLOG_HEAP_OPMASK)
8048         {
8049                 case XLOG_HEAP2_CLEAN:
8050                         heap_xlog_clean(record);
8051                         break;
8052                 case XLOG_HEAP2_FREEZE_PAGE:
8053                         heap_xlog_freeze_page(record);
8054                         break;
8055                 case XLOG_HEAP2_CLEANUP_INFO:
8056                         heap_xlog_cleanup_info(record);
8057                         break;
8058                 case XLOG_HEAP2_VISIBLE:
8059                         heap_xlog_visible(record);
8060                         break;
8061                 case XLOG_HEAP2_MULTI_INSERT:
8062                         heap_xlog_multi_insert(record);
8063                         break;
8064                 case XLOG_HEAP2_LOCK_UPDATED:
8065                         heap_xlog_lock_updated(record);
8066                         break;
8067                 case XLOG_HEAP2_NEW_CID:
8068
8069                         /*
8070                          * Nothing to do on a real replay, only used during logical
8071                          * decoding.
8072                          */
8073                         break;
8074                 case XLOG_HEAP2_REWRITE:
8075                         heap_xlog_logical_rewrite(record);
8076                         break;
8077                 default:
8078                         elog(PANIC, "heap2_redo: unknown op code %u", info);
8079         }
8080 }
8081
8082 /*
8083  *      heap_sync               - sync a heap, for use when no WAL has been written
8084  *
8085  * This forces the heap contents (including TOAST heap if any) down to disk.
8086  * If we skipped using WAL, and WAL is otherwise needed, we must force the
8087  * relation down to disk before it's safe to commit the transaction.  This
8088  * requires writing out any dirty buffers and then doing a forced fsync.
8089  *
8090  * Indexes are not touched.  (Currently, index operations associated with
8091  * the commands that use this are WAL-logged and so do not need fsync.
8092  * That behavior might change someday, but in any case it's likely that
8093  * any fsync decisions required would be per-index and hence not appropriate
8094  * to be done here.)
8095  */
8096 void
8097 heap_sync(Relation rel)
8098 {
8099         /* non-WAL-logged tables never need fsync */
8100         if (!RelationNeedsWAL(rel))
8101                 return;
8102
8103         /* main heap */
8104         FlushRelationBuffers(rel);
8105         /* FlushRelationBuffers will have opened rd_smgr */
8106         smgrimmedsync(rel->rd_smgr, MAIN_FORKNUM);
8107
8108         /* FSM is not critical, don't bother syncing it */
8109
8110         /* toast heap, if any */
8111         if (OidIsValid(rel->rd_rel->reltoastrelid))
8112         {
8113                 Relation        toastrel;
8114
8115                 toastrel = heap_open(rel->rd_rel->reltoastrelid, AccessShareLock);
8116                 FlushRelationBuffers(toastrel);
8117                 smgrimmedsync(toastrel->rd_smgr, MAIN_FORKNUM);
8118                 heap_close(toastrel, AccessShareLock);
8119         }
8120 }