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/xlogutils.h"
  55 #include "catalog/catalog.h"
  56 #include "catalog/namespace.h"
  57 #include "miscadmin.h"
  58 #include "pgstat.h"
  59 #include "storage/bufmgr.h"
  60 #include "storage/freespace.h"
  61 #include "storage/lmgr.h"
  62 #include "storage/predicate.h"
  63 #include "storage/procarray.h"
  64 #include "storage/smgr.h"
  65 #include "storage/standby.h"
  66 #include "utils/datum.h"
  67 #include "utils/inval.h"
  68 #include "utils/lsyscache.h"
  69 #include "utils/relcache.h"
  70 #include "utils/snapmgr.h"
  71 #include "utils/syscache.h"
  72 #include "utils/tqual.h"
  73
  74
  75 /* GUC variable */
  76 bool            synchronize_seqscans = true;
  77
  78
  79 static HeapScanDesc heap_beginscan_internal(Relation relation,
  80                                                 Snapshot snapshot,
  81                                                 int nkeys, ScanKey key,
  82                                                 bool allow_strat, bool allow_sync,
  83                                                 bool is_bitmapscan, bool temp_snap);
  84 static HeapTuple heap_prepare_insert(Relation relation, HeapTuple tup,
  85                                         TransactionId xid, CommandId cid, int options);
  86 static XLogRecPtr log_heap_update(Relation reln, Buffer oldbuf,
  87                                 Buffer newbuf, HeapTuple oldtup,
  88                                 HeapTuple newtup, HeapTuple old_key_tup,
  89                                 bool all_visible_cleared, bool new_all_visible_cleared);
  90 static void HeapSatisfiesHOTandKeyUpdate(Relation relation,
  91                                                          Bitmapset *hot_attrs,
  92                                                          Bitmapset *key_attrs, Bitmapset *id_attrs,
  93                                                          bool *satisfies_hot, bool *satisfies_key,
  94                                                          bool *satisfies_id,
  95                                                          HeapTuple oldtup, HeapTuple newtup);
  96 static void compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
  97                                                   uint16 old_infomask2, TransactionId add_to_xmax,
  98                                                   LockTupleMode mode, bool is_update,
  99                                                   TransactionId *result_xmax, uint16 *result_infomask,
 100                                                   uint16 *result_infomask2);
 101 static HTSU_Result heap_lock_updated_tuple(Relation rel, HeapTuple tuple,
 102                                                 ItemPointer ctid, TransactionId xid,
 103                                                 LockTupleMode mode);
 104 static void GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
 105                                            uint16 *new_infomask2);
 106 static TransactionId MultiXactIdGetUpdateXid(TransactionId xmax,
 107                                                 uint16 t_infomask);
 108 static void MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
 109                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
 110                                 int *remaining);
 111 static bool ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
 112                                                    uint16 infomask, Relation rel, int *remaining);
 113 static XLogRecPtr log_heap_new_cid(Relation relation, HeapTuple tup);
 114 static HeapTuple ExtractReplicaIdentity(Relation rel, HeapTuple tup, bool key_modified,
 115                                            bool *copy);
 116
 117
 118 /*
 119  * Each tuple lock mode has a corresponding heavyweight lock, and one or two
 120  * corresponding MultiXactStatuses (one to merely lock tuples, another one to
 121  * update them).  This table (and the macros below) helps us determine the
 122  * heavyweight lock mode and MultiXactStatus values to use for any particular
 123  * tuple lock strength.
 124  *
 125  * Don't look at lockstatus/updstatus directly!  Use get_mxact_status_for_lock
 126  * instead.
 127  */
 128 static const struct
 129 {
 130         LOCKMODE        hwlock;
 131         int                     lockstatus;
 132         int                     updstatus;
 133 }
 134
 135                         tupleLockExtraInfo[MaxLockTupleMode + 1] =
 136 {
 137         {                                                       /* LockTupleKeyShare */
 138                 AccessShareLock,
 139                 MultiXactStatusForKeyShare,
 140                 -1                                              /* KeyShare does not allow updating tuples */
 141         },
 142         {                                                       /* LockTupleShare */
 143                 RowShareLock,
 144                 MultiXactStatusForShare,
 145                 -1                                              /* Share does not allow updating tuples */
 146         },
 147         {                                                       /* LockTupleNoKeyExclusive */
 148                 ExclusiveLock,
 149                 MultiXactStatusForNoKeyUpdate,
 150                 MultiXactStatusNoKeyUpdate
 151         },
 152         {                                                       /* LockTupleExclusive */
 153                 AccessExclusiveLock,
 154                 MultiXactStatusForUpdate,
 155                 MultiXactStatusUpdate
 156         }
 157 };
 158
 159 /* Get the LOCKMODE for a given MultiXactStatus */
 160 #define LOCKMODE_from_mxstatus(status) \
 161                         (tupleLockExtraInfo[TUPLOCK_from_mxstatus((status))].hwlock)
 162
 163 /*
 164  * Acquire heavyweight locks on tuples, using a LockTupleMode strength value.
 165  * This is more readable than having every caller translate it to lock.h's
 166  * LOCKMODE.
 167  */
 168 #define LockTupleTuplock(rel, tup, mode) \
 169         LockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 170 #define UnlockTupleTuplock(rel, tup, mode) \
 171         UnlockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 172 #define ConditionalLockTupleTuplock(rel, tup, mode) \
 173         ConditionalLockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 174
 175 /*
 176  * This table maps tuple lock strength values for each particular
 177  * MultiXactStatus value.
 178  */
 179 static const int MultiXactStatusLock[MaxMultiXactStatus + 1] =
 180 {
 181         LockTupleKeyShare,                      /* ForKeyShare */
 182         LockTupleShare,                         /* ForShare */
 183         LockTupleNoKeyExclusive,        /* ForNoKeyUpdate */
 184         LockTupleExclusive,                     /* ForUpdate */
 185         LockTupleNoKeyExclusive,        /* NoKeyUpdate */
 186         LockTupleExclusive                      /* Update */
 187 };
 188
 189 /* Get the LockTupleMode for a given MultiXactStatus */
 190 #define TUPLOCK_from_mxstatus(status) \
 191                         (MultiXactStatusLock[(status)])
 192
 193 /* ----------------------------------------------------------------
 194  *                                               heap support routines
 195  * ----------------------------------------------------------------
 196  */
 197
 198 /* ----------------
 199  *              initscan - scan code common to heap_beginscan and heap_rescan
 200  * ----------------
 201  */
 202 static void
 203 initscan(HeapScanDesc scan, ScanKey key, bool is_rescan)
 204 {
 205         bool            allow_strat;
 206         bool            allow_sync;
 207
 208         /*
 209          * Determine the number of blocks we have to scan.
 210          *
 211          * It is sufficient to do this once at scan start, since any tuples added
 212          * while the scan is in progress will be invisible to my snapshot anyway.
 213          * (That is not true when using a non-MVCC snapshot.  However, we couldn't
 214          * guarantee to return tuples added after scan start anyway, since they
 215          * might go into pages we already scanned.  To guarantee consistent
 216          * results for a non-MVCC snapshot, the caller must hold some higher-level
 217          * lock that ensures the interesting tuple(s) won't change.)
 218          */
 219         scan->rs_nblocks = RelationGetNumberOfBlocks(scan->rs_rd);
 220
 221         /*
 222          * If the table is large relative to NBuffers, use a bulk-read access
 223          * strategy and enable synchronized scanning (see syncscan.c).  Although
 224          * the thresholds for these features could be different, we make them the
 225          * same so that there are only two behaviors to tune rather than four.
 226          * (However, some callers need to be able to disable one or both of these
 227          * behaviors, independently of the size of the table; also there is a GUC
 228          * variable that can disable synchronized scanning.)
 229          *
 230          * During a rescan, don't make a new strategy object if we don't have to.
 231          */
 232         if (!RelationUsesLocalBuffers(scan->rs_rd) &&
 233                 scan->rs_nblocks > NBuffers / 4)
 234         {
 235                 allow_strat = scan->rs_allow_strat;
 236                 allow_sync = scan->rs_allow_sync;
 237         }
 238         else
 239                 allow_strat = allow_sync = false;
 240
 241         if (allow_strat)
 242         {
 243                 if (scan->rs_strategy == NULL)
 244                         scan->rs_strategy = GetAccessStrategy(BAS_BULKREAD);
 245         }
 246         else
 247         {
 248                 if (scan->rs_strategy != NULL)
 249                         FreeAccessStrategy(scan->rs_strategy);
 250                 scan->rs_strategy = NULL;
 251         }
 252
 253         if (is_rescan)
 254         {
 255                 /*
 256                  * If rescan, keep the previous startblock setting so that rewinding a
 257                  * cursor doesn't generate surprising results.  Reset the syncscan
 258                  * setting, though.
 259                  */
 260                 scan->rs_syncscan = (allow_sync && synchronize_seqscans);
 261         }
 262         else if (allow_sync && synchronize_seqscans)
 263         {
 264                 scan->rs_syncscan = true;
 265                 scan->rs_startblock = ss_get_location(scan->rs_rd, scan->rs_nblocks);
 266         }
 267         else
 268         {
 269                 scan->rs_syncscan = false;
 270                 scan->rs_startblock = 0;
 271         }
 272
 273         scan->rs_inited = false;
 274         scan->rs_ctup.t_data = NULL;
 275         ItemPointerSetInvalid(&scan->rs_ctup.t_self);
 276         scan->rs_cbuf = InvalidBuffer;
 277         scan->rs_cblock = InvalidBlockNumber;
 278
 279         /* we don't have a marked position... */
 280         ItemPointerSetInvalid(&(scan->rs_mctid));
 281
 282         /* page-at-a-time fields are always invalid when not rs_inited */
 283
 284         /*
 285          * copy the scan key, if appropriate
 286          */
 287         if (key != NULL)
 288                 memcpy(scan->rs_key, key, scan->rs_nkeys * sizeof(ScanKeyData));
 289
 290         /*
 291          * Currently, we don't have a stats counter for bitmap heap scans (but the
 292          * underlying bitmap index scans will be counted).
 293          */
 294         if (!scan->rs_bitmapscan)
 295                 pgstat_count_heap_scan(scan->rs_rd);
 296 }
 297
 298 /*
 299  * heapgetpage - subroutine for heapgettup()
 300  *
 301  * This routine reads and pins the specified page of the relation.
 302  * In page-at-a-time mode it performs additional work, namely determining
 303  * which tuples on the page are visible.
 304  */
 305 static void
 306 heapgetpage(HeapScanDesc scan, BlockNumber page)
 307 {
 308         Buffer          buffer;
 309         Snapshot        snapshot;
 310         Page            dp;
 311         int                     lines;
 312         int                     ntup;
 313         OffsetNumber lineoff;
 314         ItemId          lpp;
 315         bool            all_visible;
 316
 317         Assert(page < scan->rs_nblocks);
 318
 319         /* release previous scan buffer, if any */
 320         if (BufferIsValid(scan->rs_cbuf))
 321         {
 322                 ReleaseBuffer(scan->rs_cbuf);
 323                 scan->rs_cbuf = InvalidBuffer;
 324         }
 325
 326         /*
 327          * Be sure to check for interrupts at least once per page.  Checks at
 328          * higher code levels won't be able to stop a seqscan that encounters many
 329          * pages' worth of consecutive dead tuples.
 330          */
 331         CHECK_FOR_INTERRUPTS();
 332
 333         /* read page using selected strategy */
 334         scan->rs_cbuf = ReadBufferExtended(scan->rs_rd, MAIN_FORKNUM, page,
 335                                                                            RBM_NORMAL, scan->rs_strategy);
 336         scan->rs_cblock = page;
 337
 338         if (!scan->rs_pageatatime)
 339                 return;
 340
 341         buffer = scan->rs_cbuf;
 342         snapshot = scan->rs_snapshot;
 343
 344         /*
 345          * Prune and repair fragmentation for the whole page, if possible.
 346          */
 347         heap_page_prune_opt(scan->rs_rd, buffer);
 348
 349         /*
 350          * We must hold share lock on the buffer content while examining tuple
 351          * visibility.  Afterwards, however, the tuples we have found to be
 352          * visible are guaranteed good as long as we hold the buffer pin.
 353          */
 354         LockBuffer(buffer, BUFFER_LOCK_SHARE);
 355
 356         dp = (Page) BufferGetPage(buffer);
 357         lines = PageGetMaxOffsetNumber(dp);
 358         ntup = 0;
 359
 360         /*
 361          * If the all-visible flag indicates that all tuples on the page are
 362          * visible to everyone, we can skip the per-tuple visibility tests.
 363          *
 364          * Note: In hot standby, a tuple that's already visible to all
 365          * transactions in the master might still be invisible to a read-only
 366          * transaction in the standby. We partly handle this problem by tracking
 367          * the minimum xmin of visible tuples as the cut-off XID while marking a
 368          * page all-visible on master and WAL log that along with the visibility
 369          * map SET operation. In hot standby, we wait for (or abort) all
 370          * transactions that can potentially may not see one or more tuples on the
 371          * page. That's how index-only scans work fine in hot standby. A crucial
 372          * difference between index-only scans and heap scans is that the
 373          * index-only scan completely relies on the visibility map where as heap
 374          * scan looks at the page-level PD_ALL_VISIBLE flag. We are not sure if
 375          * the page-level flag can be trusted in the same way, because it might
 376          * get propagated somehow without being explicitly WAL-logged, e.g. via a
 377          * full page write. Until we can prove that beyond doubt, let's check each
 378          * tuple for visibility the hard way.
 379          */
 380         all_visible = PageIsAllVisible(dp) && !snapshot->takenDuringRecovery;
 381
 382         for (lineoff = FirstOffsetNumber, lpp = PageGetItemId(dp, lineoff);
 383                  lineoff <= lines;
 384                  lineoff++, lpp++)
 385         {
 386                 if (ItemIdIsNormal(lpp))
 387                 {
 388                         HeapTupleData loctup;
 389                         bool            valid;
 390
 391                         loctup.t_tableOid = RelationGetRelid(scan->rs_rd);
 392                         loctup.t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 393                         loctup.t_len = ItemIdGetLength(lpp);
 394                         ItemPointerSet(&(loctup.t_self), page, lineoff);
 395
 396                         if (all_visible)
 397                                 valid = true;
 398                         else
 399                                 valid = HeapTupleSatisfiesVisibility(&loctup, snapshot, buffer);
 400
 401                         CheckForSerializableConflictOut(valid, scan->rs_rd, &loctup,
 402                                                                                         buffer, snapshot);
 403
 404                         if (valid)
 405                                 scan->rs_vistuples[ntup++] = lineoff;
 406                 }
 407         }
 408
 409         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
 410
 411         Assert(ntup <= MaxHeapTuplesPerPage);
 412         scan->rs_ntuples = ntup;
 413 }
 414
 415 /* ----------------
 416  *              heapgettup - fetch next heap tuple
 417  *
 418  *              Initialize the scan if not already done; then advance to the next
 419  *              tuple as indicated by "dir"; return the next tuple in scan->rs_ctup,
 420  *              or set scan->rs_ctup.t_data = NULL if no more tuples.
 421  *
 422  * dir == NoMovementScanDirection means "re-fetch the tuple indicated
 423  * by scan->rs_ctup".
 424  *
 425  * Note: the reason nkeys/key are passed separately, even though they are
 426  * kept in the scan descriptor, is that the caller may not want us to check
 427  * the scankeys.
 428  *
 429  * Note: when we fall off the end of the scan in either direction, we
 430  * reset rs_inited.  This means that a further request with the same
 431  * scan direction will restart the scan, which is a bit odd, but a
 432  * request with the opposite scan direction will start a fresh scan
 433  * in the proper direction.  The latter is required behavior for cursors,
 434  * while the former case is generally undefined behavior in Postgres
 435  * so we don't care too much.
 436  * ----------------
 437  */
 438 static void
 439 heapgettup(HeapScanDesc scan,
 440                    ScanDirection dir,
 441                    int nkeys,
 442                    ScanKey key)
 443 {
 444         HeapTuple       tuple = &(scan->rs_ctup);
 445         Snapshot        snapshot = scan->rs_snapshot;
 446         bool            backward = ScanDirectionIsBackward(dir);
 447         BlockNumber page;
 448         bool            finished;
 449         Page            dp;
 450         int                     lines;
 451         OffsetNumber lineoff;
 452         int                     linesleft;
 453         ItemId          lpp;
 454
 455         /*
 456          * calculate next starting lineoff, given scan direction
 457          */
 458         if (ScanDirectionIsForward(dir))
 459         {
 460                 if (!scan->rs_inited)
 461                 {
 462                         /*
 463                          * return null immediately if relation is empty
 464                          */
 465                         if (scan->rs_nblocks == 0)
 466                         {
 467                                 Assert(!BufferIsValid(scan->rs_cbuf));
 468                                 tuple->t_data = NULL;
 469                                 return;
 470                         }
 471                         page = scan->rs_startblock; /* first page */
 472                         heapgetpage(scan, page);
 473                         lineoff = FirstOffsetNumber;            /* first offnum */
 474                         scan->rs_inited = true;
 475                 }
 476                 else
 477                 {
 478                         /* continue from previously returned page/tuple */
 479                         page = scan->rs_cblock;         /* current page */
 480                         lineoff =                       /* next offnum */
 481                                 OffsetNumberNext(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 482                 }
 483
 484                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 485
 486                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 487                 lines = PageGetMaxOffsetNumber(dp);
 488                 /* page and lineoff now reference the physically next tid */
 489
 490                 linesleft = lines - lineoff + 1;
 491         }
 492         else if (backward)
 493         {
 494                 if (!scan->rs_inited)
 495                 {
 496                         /*
 497                          * return null immediately if relation is empty
 498                          */
 499                         if (scan->rs_nblocks == 0)
 500                         {
 501                                 Assert(!BufferIsValid(scan->rs_cbuf));
 502                                 tuple->t_data = NULL;
 503                                 return;
 504                         }
 505
 506                         /*
 507                          * Disable reporting to syncscan logic in a backwards scan; it's
 508                          * not very likely anyone else is doing the same thing at the same
 509                          * time, and much more likely that we'll just bollix things for
 510                          * forward scanners.
 511                          */
 512                         scan->rs_syncscan = false;
 513                         /* start from last page of the scan */
 514                         if (scan->rs_startblock > 0)
 515                                 page = scan->rs_startblock - 1;
 516                         else
 517                                 page = scan->rs_nblocks - 1;
 518                         heapgetpage(scan, page);
 519                 }
 520                 else
 521                 {
 522                         /* continue from previously returned page/tuple */
 523                         page = scan->rs_cblock;         /* current page */
 524                 }
 525
 526                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 527
 528                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 529                 lines = PageGetMaxOffsetNumber(dp);
 530
 531                 if (!scan->rs_inited)
 532                 {
 533                         lineoff = lines;        /* final offnum */
 534                         scan->rs_inited = true;
 535                 }
 536                 else
 537                 {
 538                         lineoff =                       /* previous offnum */
 539                                 OffsetNumberPrev(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 540                 }
 541                 /* page and lineoff now reference the physically previous tid */
 542
 543                 linesleft = lineoff;
 544         }
 545         else
 546         {
 547                 /*
 548                  * ``no movement'' scan direction: refetch prior tuple
 549                  */
 550                 if (!scan->rs_inited)
 551                 {
 552                         Assert(!BufferIsValid(scan->rs_cbuf));
 553                         tuple->t_data = NULL;
 554                         return;
 555                 }
 556
 557                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 558                 if (page != scan->rs_cblock)
 559                         heapgetpage(scan, page);
 560
 561                 /* Since the tuple was previously fetched, needn't lock page here */
 562                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 563                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 564                 lpp = PageGetItemId(dp, lineoff);
 565                 Assert(ItemIdIsNormal(lpp));
 566
 567                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 568                 tuple->t_len = ItemIdGetLength(lpp);
 569
 570                 return;
 571         }
 572
 573         /*
 574          * advance the scan until we find a qualifying tuple or run out of stuff
 575          * to scan
 576          */
 577         lpp = PageGetItemId(dp, lineoff);
 578         for (;;)
 579         {
 580                 while (linesleft > 0)
 581                 {
 582                         if (ItemIdIsNormal(lpp))
 583                         {
 584                                 bool            valid;
 585
 586                                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 587                                 tuple->t_len = ItemIdGetLength(lpp);
 588                                 ItemPointerSet(&(tuple->t_self), page, lineoff);
 589
 590                                 /*
 591                                  * if current tuple qualifies, return it.
 592                                  */
 593                                 valid = HeapTupleSatisfiesVisibility(tuple,
 594                                                                                                          snapshot,
 595                                                                                                          scan->rs_cbuf);
 596
 597                                 CheckForSerializableConflictOut(valid, scan->rs_rd, tuple,
 598                                                                                                 scan->rs_cbuf, snapshot);
 599
 600                                 if (valid && key != NULL)
 601                                         HeapKeyTest(tuple, RelationGetDescr(scan->rs_rd),
 602                                                                 nkeys, key, valid);
 603
 604                                 if (valid)
 605                                 {
 606                                         LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 607                                         return;
 608                                 }
 609                         }
 610
 611                         /*
 612                          * otherwise move to the next item on the page
 613                          */
 614                         --linesleft;
 615                         if (backward)
 616                         {
 617                                 --lpp;                  /* move back in this page's ItemId array */
 618                                 --lineoff;
 619                         }
 620                         else
 621                         {
 622                                 ++lpp;                  /* move forward in this page's ItemId array */
 623                                 ++lineoff;
 624                         }
 625                 }
 626
 627                 /*
 628                  * if we get here, it means we've exhausted the items on this page and
 629                  * it's time to move to the next.
 630                  */
 631                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 632
 633                 /*
 634                  * advance to next/prior page and detect end of scan
 635                  */
 636                 if (backward)
 637                 {
 638                         finished = (page == scan->rs_startblock);
 639                         if (page == 0)
 640                                 page = scan->rs_nblocks;
 641                         page--;
 642                 }
 643                 else
 644                 {
 645                         page++;
 646                         if (page >= scan->rs_nblocks)
 647                                 page = 0;
 648                         finished = (page == scan->rs_startblock);
 649
 650                         /*
 651                          * Report our new scan position for synchronization purposes. We
 652                          * don't do that when moving backwards, however. That would just
 653                          * mess up any other forward-moving scanners.
 654                          *
 655                          * Note: we do this before checking for end of scan so that the
 656                          * final state of the position hint is back at the start of the
 657                          * rel.  That's not strictly necessary, but otherwise when you run
 658                          * the same query multiple times the starting position would shift
 659                          * a little bit backwards on every invocation, which is confusing.
 660                          * We don't guarantee any specific ordering in general, though.
 661                          */
 662                         if (scan->rs_syncscan)
 663                                 ss_report_location(scan->rs_rd, page);
 664                 }
 665
 666                 /*
 667                  * return NULL if we've exhausted all the pages
 668                  */
 669                 if (finished)
 670                 {
 671                         if (BufferIsValid(scan->rs_cbuf))
 672                                 ReleaseBuffer(scan->rs_cbuf);
 673                         scan->rs_cbuf = InvalidBuffer;
 674                         scan->rs_cblock = InvalidBlockNumber;
 675                         tuple->t_data = NULL;
 676                         scan->rs_inited = false;
 677                         return;
 678                 }
 679
 680                 heapgetpage(scan, page);
 681
 682                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 683
 684                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 685                 lines = PageGetMaxOffsetNumber((Page) dp);
 686                 linesleft = lines;
 687                 if (backward)
 688                 {
 689                         lineoff = lines;
 690                         lpp = PageGetItemId(dp, lines);
 691                 }
 692                 else
 693                 {
 694                         lineoff = FirstOffsetNumber;
 695                         lpp = PageGetItemId(dp, FirstOffsetNumber);
 696                 }
 697         }
 698 }
 699
 700 /* ----------------
 701  *              heapgettup_pagemode - fetch next heap tuple in page-at-a-time mode
 702  *
 703  *              Same API as heapgettup, but used in page-at-a-time mode
 704  *
 705  * The internal logic is much the same as heapgettup's too, but there are some
 706  * differences: we do not take the buffer content lock (that only needs to
 707  * happen inside heapgetpage), and we iterate through just the tuples listed
 708  * in rs_vistuples[] rather than all tuples on the page.  Notice that
 709  * lineindex is 0-based, where the corresponding loop variable lineoff in
 710  * heapgettup is 1-based.
 711  * ----------------
 712  */
 713 static void
 714 heapgettup_pagemode(HeapScanDesc scan,
 715                                         ScanDirection dir,
 716                                         int nkeys,
 717                                         ScanKey key)
 718 {
 719         HeapTuple       tuple = &(scan->rs_ctup);
 720         bool            backward = ScanDirectionIsBackward(dir);
 721         BlockNumber page;
 722         bool            finished;
 723         Page            dp;
 724         int                     lines;
 725         int                     lineindex;
 726         OffsetNumber lineoff;
 727         int                     linesleft;
 728         ItemId          lpp;
 729
 730         /*
 731          * calculate next starting lineindex, given scan direction
 732          */
 733         if (ScanDirectionIsForward(dir))
 734         {
 735                 if (!scan->rs_inited)
 736                 {
 737                         /*
 738                          * return null immediately if relation is empty
 739                          */
 740                         if (scan->rs_nblocks == 0)
 741                         {
 742                                 Assert(!BufferIsValid(scan->rs_cbuf));
 743                                 tuple->t_data = NULL;
 744                                 return;
 745                         }
 746                         page = scan->rs_startblock; /* first page */
 747                         heapgetpage(scan, page);
 748                         lineindex = 0;
 749                         scan->rs_inited = true;
 750                 }
 751                 else
 752                 {
 753                         /* continue from previously returned page/tuple */
 754                         page = scan->rs_cblock;         /* current page */
 755                         lineindex = scan->rs_cindex + 1;
 756                 }
 757
 758                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 759                 lines = scan->rs_ntuples;
 760                 /* page and lineindex now reference the next visible tid */
 761
 762                 linesleft = lines - lineindex;
 763         }
 764         else if (backward)
 765         {
 766                 if (!scan->rs_inited)
 767                 {
 768                         /*
 769                          * return null immediately if relation is empty
 770                          */
 771                         if (scan->rs_nblocks == 0)
 772                         {
 773                                 Assert(!BufferIsValid(scan->rs_cbuf));
 774                                 tuple->t_data = NULL;
 775                                 return;
 776                         }
 777
 778                         /*
 779                          * Disable reporting to syncscan logic in a backwards scan; it's
 780                          * not very likely anyone else is doing the same thing at the same
 781                          * time, and much more likely that we'll just bollix things for
 782                          * forward scanners.
 783                          */
 784                         scan->rs_syncscan = false;
 785                         /* start from last page of the scan */
 786                         if (scan->rs_startblock > 0)
 787                                 page = scan->rs_startblock - 1;
 788                         else
 789                                 page = scan->rs_nblocks - 1;
 790                         heapgetpage(scan, page);
 791                 }
 792                 else
 793                 {
 794                         /* continue from previously returned page/tuple */
 795                         page = scan->rs_cblock;         /* current page */
 796                 }
 797
 798                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 799                 lines = scan->rs_ntuples;
 800
 801                 if (!scan->rs_inited)
 802                 {
 803                         lineindex = lines - 1;
 804                         scan->rs_inited = true;
 805                 }
 806                 else
 807                 {
 808                         lineindex = scan->rs_cindex - 1;
 809                 }
 810                 /* page and lineindex now reference the previous visible tid */
 811
 812                 linesleft = lineindex + 1;
 813         }
 814         else
 815         {
 816                 /*
 817                  * ``no movement'' scan direction: refetch prior tuple
 818                  */
 819                 if (!scan->rs_inited)
 820                 {
 821                         Assert(!BufferIsValid(scan->rs_cbuf));
 822                         tuple->t_data = NULL;
 823                         return;
 824                 }
 825
 826                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 827                 if (page != scan->rs_cblock)
 828                         heapgetpage(scan, page);
 829
 830                 /* Since the tuple was previously fetched, needn't lock page here */
 831                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 832                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 833                 lpp = PageGetItemId(dp, lineoff);
 834                 Assert(ItemIdIsNormal(lpp));
 835
 836                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 837                 tuple->t_len = ItemIdGetLength(lpp);
 838
 839                 /* check that rs_cindex is in sync */
 840                 Assert(scan->rs_cindex < scan->rs_ntuples);
 841                 Assert(lineoff == scan->rs_vistuples[scan->rs_cindex]);
 842
 843                 return;
 844         }
 845
 846         /*
 847          * advance the scan until we find a qualifying tuple or run out of stuff
 848          * to scan
 849          */
 850         for (;;)
 851         {
 852                 while (linesleft > 0)
 853                 {
 854                         lineoff = scan->rs_vistuples[lineindex];
 855                         lpp = PageGetItemId(dp, lineoff);
 856                         Assert(ItemIdIsNormal(lpp));
 857
 858                         tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 859                         tuple->t_len = ItemIdGetLength(lpp);
 860                         ItemPointerSet(&(tuple->t_self), page, lineoff);
 861
 862                         /*
 863                          * if current tuple qualifies, return it.
 864                          */
 865                         if (key != NULL)
 866                         {
 867                                 bool            valid;
 868
 869                                 HeapKeyTest(tuple, RelationGetDescr(scan->rs_rd),
 870                                                         nkeys, key, valid);
 871                                 if (valid)
 872                                 {
 873                                         scan->rs_cindex = lineindex;
 874                                         return;
 875                                 }
 876                         }
 877                         else
 878                         {
 879                                 scan->rs_cindex = lineindex;
 880                                 return;
 881                         }
 882
 883                         /*
 884                          * otherwise move to the next item on the page
 885                          */
 886                         --linesleft;
 887                         if (backward)
 888                                 --lineindex;
 889                         else
 890                                 ++lineindex;
 891                 }
 892
 893                 /*
 894                  * if we get here, it means we've exhausted the items on this page and
 895                  * it's time to move to the next.
 896                  */
 897                 if (backward)
 898                 {
 899                         finished = (page == scan->rs_startblock);
 900                         if (page == 0)
 901                                 page = scan->rs_nblocks;
 902                         page--;
 903                 }
 904                 else
 905                 {
 906                         page++;
 907                         if (page >= scan->rs_nblocks)
 908                                 page = 0;
 909                         finished = (page == scan->rs_startblock);
 910
 911                         /*
 912                          * Report our new scan position for synchronization purposes. We
 913                          * don't do that when moving backwards, however. That would just
 914                          * mess up any other forward-moving scanners.
 915                          *
 916                          * Note: we do this before checking for end of scan so that the
 917                          * final state of the position hint is back at the start of the
 918                          * rel.  That's not strictly necessary, but otherwise when you run
 919                          * the same query multiple times the starting position would shift
 920                          * a little bit backwards on every invocation, which is confusing.
 921                          * We don't guarantee any specific ordering in general, though.
 922                          */
 923                         if (scan->rs_syncscan)
 924                                 ss_report_location(scan->rs_rd, page);
 925                 }
 926
 927                 /*
 928                  * return NULL if we've exhausted all the pages
 929                  */
 930                 if (finished)
 931                 {
 932                         if (BufferIsValid(scan->rs_cbuf))
 933                                 ReleaseBuffer(scan->rs_cbuf);
 934                         scan->rs_cbuf = InvalidBuffer;
 935                         scan->rs_cblock = InvalidBlockNumber;
 936                         tuple->t_data = NULL;
 937                         scan->rs_inited = false;
 938                         return;
 939                 }
 940
 941                 heapgetpage(scan, page);
 942
 943                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 944                 lines = scan->rs_ntuples;
 945                 linesleft = lines;
 946                 if (backward)
 947                         lineindex = lines - 1;
 948                 else
 949                         lineindex = 0;
 950         }
 951 }
 952
 953
 954 #if defined(DISABLE_COMPLEX_MACRO)
 955 /*
 956  * This is formatted so oddly so that the correspondence to the macro
 957  * definition in access/htup_details.h is maintained.
 958  */
 959 Datum
 960 fastgetattr(HeapTuple tup, int attnum, TupleDesc tupleDesc,
 961                         bool *isnull)
 962 {
 963         return (
 964                         (attnum) > 0 ?
 965                         (
 966                          (*(isnull) = false),
 967                          HeapTupleNoNulls(tup) ?
 968                          (
 969                           (tupleDesc)->attrs[(attnum) - 1]->attcacheoff >= 0 ?
 970                           (
 971                            fetchatt((tupleDesc)->attrs[(attnum) - 1],
 972                                                 (char *) (tup)->t_data + (tup)->t_data->t_hoff +
 973                                                 (tupleDesc)->attrs[(attnum) - 1]->attcacheoff)
 974                            )
 975                           :
 976                           nocachegetattr((tup), (attnum), (tupleDesc))
 977                           )
 978                          :
 979                          (
 980                           att_isnull((attnum) - 1, (tup)->t_data->t_bits) ?
 981                           (
 982                            (*(isnull) = true),
 983                            (Datum) NULL
 984                            )
 985                           :
 986                           (
 987                            nocachegetattr((tup), (attnum), (tupleDesc))
 988                            )
 989                           )
 990                          )
 991                         :
 992                         (
 993                          (Datum) NULL
 994                          )
 995                 );
 996 }
 997 #endif   /* defined(DISABLE_COMPLEX_MACRO) */
 998
 999
1000 /* ----------------------------------------------------------------
1001  *                                       heap access method interface
1002  * ----------------------------------------------------------------
1003  */
1004
1005 /* ----------------
1006  *              relation_open - open any relation by relation OID
1007  *
1008  *              If lockmode is not "NoLock", the specified kind of lock is
1009  *              obtained on the relation.  (Generally, NoLock should only be
1010  *              used if the caller knows it has some appropriate lock on the
1011  *              relation already.)
1012  *
1013  *              An error is raised if the relation does not exist.
1014  *
1015  *              NB: a "relation" is anything with a pg_class entry.  The caller is
1016  *              expected to check whether the relkind is something it can handle.
1017  * ----------------
1018  */
1019 Relation
1020 relation_open(Oid relationId, LOCKMODE lockmode)
1021 {
1022         Relation        r;
1023
1024         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1025
1026         /* Get the lock before trying to open the relcache entry */
1027         if (lockmode != NoLock)
1028                 LockRelationOid(relationId, lockmode);
1029
1030         /* The relcache does all the real work... */
1031         r = RelationIdGetRelation(relationId);
1032
1033         if (!RelationIsValid(r))
1034                 elog(ERROR, "could not open relation with OID %u", relationId);
1035
1036         /* Make note that we've accessed a temporary relation */
1037         if (RelationUsesLocalBuffers(r))
1038                 MyXactAccessedTempRel = true;
1039
1040         pgstat_initstats(r);
1041
1042         return r;
1043 }
1044
1045 /* ----------------
1046  *              try_relation_open - open any relation by relation OID
1047  *
1048  *              Same as relation_open, except return NULL instead of failing
1049  *              if the relation does not exist.
1050  * ----------------
1051  */
1052 Relation
1053 try_relation_open(Oid relationId, LOCKMODE lockmode)
1054 {
1055         Relation        r;
1056
1057         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1058
1059         /* Get the lock first */
1060         if (lockmode != NoLock)
1061                 LockRelationOid(relationId, lockmode);
1062
1063         /*
1064          * Now that we have the lock, probe to see if the relation really exists
1065          * or not.
1066          */
1067         if (!SearchSysCacheExists1(RELOID, ObjectIdGetDatum(relationId)))
1068         {
1069                 /* Release useless lock */
1070                 if (lockmode != NoLock)
1071                         UnlockRelationOid(relationId, lockmode);
1072
1073                 return NULL;
1074         }
1075
1076         /* Should be safe to do a relcache load */
1077         r = RelationIdGetRelation(relationId);
1078
1079         if (!RelationIsValid(r))
1080                 elog(ERROR, "could not open relation with OID %u", relationId);
1081
1082         /* Make note that we've accessed a temporary relation */
1083         if (RelationUsesLocalBuffers(r))
1084                 MyXactAccessedTempRel = true;
1085
1086         pgstat_initstats(r);
1087
1088         return r;
1089 }
1090
1091 /* ----------------
1092  *              relation_openrv - open any relation specified by a RangeVar
1093  *
1094  *              Same as relation_open, but the relation is specified by a RangeVar.
1095  * ----------------
1096  */
1097 Relation
1098 relation_openrv(const RangeVar *relation, LOCKMODE lockmode)
1099 {
1100         Oid                     relOid;
1101
1102         /*
1103          * Check for shared-cache-inval messages before trying to open the
1104          * relation.  This is needed even if we already hold a lock on the
1105          * relation, because GRANT/REVOKE are executed without taking any lock on
1106          * the target relation, and we want to be sure we see current ACL
1107          * information.  We can skip this if asked for NoLock, on the assumption
1108          * that such a call is not the first one in the current command, and so we
1109          * should be reasonably up-to-date already.  (XXX this all could stand to
1110          * be redesigned, but for the moment we'll keep doing this like it's been
1111          * done historically.)
1112          */
1113         if (lockmode != NoLock)
1114                 AcceptInvalidationMessages();
1115
1116         /* Look up and lock the appropriate relation using namespace search */
1117         relOid = RangeVarGetRelid(relation, lockmode, false);
1118
1119         /* Let relation_open do the rest */
1120         return relation_open(relOid, NoLock);
1121 }
1122
1123 /* ----------------
1124  *              relation_openrv_extended - open any relation specified by a RangeVar
1125  *
1126  *              Same as relation_openrv, but with an additional missing_ok argument
1127  *              allowing a NULL return rather than an error if the relation is not
1128  *              found.  (Note that some other causes, such as permissions problems,
1129  *              will still result in an ereport.)
1130  * ----------------
1131  */
1132 Relation
1133 relation_openrv_extended(const RangeVar *relation, LOCKMODE lockmode,
1134                                                  bool missing_ok)
1135 {
1136         Oid                     relOid;
1137
1138         /*
1139          * Check for shared-cache-inval messages before trying to open the
1140          * relation.  See comments in relation_openrv().
1141          */
1142         if (lockmode != NoLock)
1143                 AcceptInvalidationMessages();
1144
1145         /* Look up and lock the appropriate relation using namespace search */
1146         relOid = RangeVarGetRelid(relation, lockmode, missing_ok);
1147
1148         /* Return NULL on not-found */
1149         if (!OidIsValid(relOid))
1150                 return NULL;
1151
1152         /* Let relation_open do the rest */
1153         return relation_open(relOid, NoLock);
1154 }
1155
1156 /* ----------------
1157  *              relation_close - close any relation
1158  *
1159  *              If lockmode is not "NoLock", we then release the specified lock.
1160  *
1161  *              Note that it is often sensible to hold a lock beyond relation_close;
1162  *              in that case, the lock is released automatically at xact end.
1163  * ----------------
1164  */
1165 void
1166 relation_close(Relation relation, LOCKMODE lockmode)
1167 {
1168         LockRelId       relid = relation->rd_lockInfo.lockRelId;
1169
1170         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1171
1172         /* The relcache does the real work... */
1173         RelationClose(relation);
1174
1175         if (lockmode != NoLock)
1176                 UnlockRelationId(&relid, lockmode);
1177 }
1178
1179
1180 /* ----------------
1181  *              heap_open - open a heap relation by relation OID
1182  *
1183  *              This is essentially relation_open plus check that the relation
1184  *              is not an index nor a composite type.  (The caller should also
1185  *              check that it's not a view or foreign table before assuming it has
1186  *              storage.)
1187  * ----------------
1188  */
1189 Relation
1190 heap_open(Oid relationId, LOCKMODE lockmode)
1191 {
1192         Relation        r;
1193
1194         r = relation_open(relationId, lockmode);
1195
1196         if (r->rd_rel->relkind == RELKIND_INDEX)
1197                 ereport(ERROR,
1198                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1199                                  errmsg("\"%s\" is an index",
1200                                                 RelationGetRelationName(r))));
1201         else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1202                 ereport(ERROR,
1203                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1204                                  errmsg("\"%s\" is a composite type",
1205                                                 RelationGetRelationName(r))));
1206
1207         return r;
1208 }
1209
1210 /* ----------------
1211  *              heap_openrv - open a heap relation specified
1212  *              by a RangeVar node
1213  *
1214  *              As above, but relation is specified by a RangeVar.
1215  * ----------------
1216  */
1217 Relation
1218 heap_openrv(const RangeVar *relation, LOCKMODE lockmode)
1219 {
1220         Relation        r;
1221
1222         r = relation_openrv(relation, lockmode);
1223
1224         if (r->rd_rel->relkind == RELKIND_INDEX)
1225                 ereport(ERROR,
1226                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1227                                  errmsg("\"%s\" is an index",
1228                                                 RelationGetRelationName(r))));
1229         else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1230                 ereport(ERROR,
1231                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1232                                  errmsg("\"%s\" is a composite type",
1233                                                 RelationGetRelationName(r))));
1234
1235         return r;
1236 }
1237
1238 /* ----------------
1239  *              heap_openrv_extended - open a heap relation specified
1240  *              by a RangeVar node
1241  *
1242  *              As above, but optionally return NULL instead of failing for
1243  *              relation-not-found.
1244  * ----------------
1245  */
1246 Relation
1247 heap_openrv_extended(const RangeVar *relation, LOCKMODE lockmode,
1248                                          bool missing_ok)
1249 {
1250         Relation        r;
1251
1252         r = relation_openrv_extended(relation, lockmode, missing_ok);
1253
1254         if (r)
1255         {
1256                 if (r->rd_rel->relkind == RELKIND_INDEX)
1257                         ereport(ERROR,
1258                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1259                                          errmsg("\"%s\" is an index",
1260                                                         RelationGetRelationName(r))));
1261                 else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1262                         ereport(ERROR,
1263                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1264                                          errmsg("\"%s\" is a composite type",
1265                                                         RelationGetRelationName(r))));
1266         }
1267
1268         return r;
1269 }
1270
1271
1272 /* ----------------
1273  *              heap_beginscan  - begin relation scan
1274  *
1275  * heap_beginscan_strat offers an extended API that lets the caller control
1276  * whether a nondefault buffer access strategy can be used, and whether
1277  * syncscan can be chosen (possibly resulting in the scan not starting from
1278  * block zero).  Both of these default to TRUE with plain heap_beginscan.
1279  *
1280  * heap_beginscan_bm is an alternative entry point for setting up a
1281  * HeapScanDesc for a bitmap heap scan.  Although that scan technology is
1282  * really quite unlike a standard seqscan, there is just enough commonality
1283  * to make it worth using the same data structure.
1284  * ----------------
1285  */
1286 HeapScanDesc
1287 heap_beginscan(Relation relation, Snapshot snapshot,
1288                            int nkeys, ScanKey key)
1289 {
1290         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1291                                                                    true, true, false, false);
1292 }
1293
1294 HeapScanDesc
1295 heap_beginscan_catalog(Relation relation, int nkeys, ScanKey key)
1296 {
1297         Oid                     relid = RelationGetRelid(relation);
1298         Snapshot        snapshot = RegisterSnapshot(GetCatalogSnapshot(relid));
1299
1300         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1301                                                                    true, true, false, true);
1302 }
1303
1304 HeapScanDesc
1305 heap_beginscan_strat(Relation relation, Snapshot snapshot,
1306                                          int nkeys, ScanKey key,
1307                                          bool allow_strat, bool allow_sync)
1308 {
1309         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1310                                                                    allow_strat, allow_sync, false, false);
1311 }
1312
1313 HeapScanDesc
1314 heap_beginscan_bm(Relation relation, Snapshot snapshot,
1315                                   int nkeys, ScanKey key)
1316 {
1317         return heap_beginscan_internal(relation, snapshot, nkeys, key,
1318                                                                    false, false, true, false);
1319 }
1320
1321 static HeapScanDesc
1322 heap_beginscan_internal(Relation relation, Snapshot snapshot,
1323                                                 int nkeys, ScanKey key,
1324                                                 bool allow_strat, bool allow_sync,
1325                                                 bool is_bitmapscan, bool temp_snap)
1326 {
1327         HeapScanDesc scan;
1328
1329         /*
1330          * increment relation ref count while scanning relation
1331          *
1332          * This is just to make really sure the relcache entry won't go away while
1333          * the scan has a pointer to it.  Caller should be holding the rel open
1334          * anyway, so this is redundant in all normal scenarios...
1335          */
1336         RelationIncrementReferenceCount(relation);
1337
1338         /*
1339          * allocate and initialize scan descriptor
1340          */
1341         scan = (HeapScanDesc) palloc(sizeof(HeapScanDescData));
1342
1343         scan->rs_rd = relation;
1344         scan->rs_snapshot = snapshot;
1345         scan->rs_nkeys = nkeys;
1346         scan->rs_bitmapscan = is_bitmapscan;
1347         scan->rs_strategy = NULL;       /* set in initscan */
1348         scan->rs_allow_strat = allow_strat;
1349         scan->rs_allow_sync = allow_sync;
1350         scan->rs_temp_snap = temp_snap;
1351
1352         /*
1353          * we can use page-at-a-time mode if it's an MVCC-safe snapshot
1354          */
1355         scan->rs_pageatatime = IsMVCCSnapshot(snapshot);
1356
1357         /*
1358          * For a seqscan in a serializable transaction, acquire a predicate lock
1359          * on the entire relation. This is required not only to lock all the
1360          * matching tuples, but also to conflict with new insertions into the
1361          * table. In an indexscan, we take page locks on the index pages covering
1362          * the range specified in the scan qual, but in a heap scan there is
1363          * nothing more fine-grained to lock. A bitmap scan is a different story,
1364          * there we have already scanned the index and locked the index pages
1365          * covering the predicate. But in that case we still have to lock any
1366          * matching heap tuples.
1367          */
1368         if (!is_bitmapscan)
1369                 PredicateLockRelation(relation, snapshot);
1370
1371         /* we only need to set this up once */
1372         scan->rs_ctup.t_tableOid = RelationGetRelid(relation);
1373
1374         /*
1375          * we do this here instead of in initscan() because heap_rescan also calls
1376          * initscan() and we don't want to allocate memory again
1377          */
1378         if (nkeys > 0)
1379                 scan->rs_key = (ScanKey) palloc(sizeof(ScanKeyData) * nkeys);
1380         else
1381                 scan->rs_key = NULL;
1382
1383         initscan(scan, key, false);
1384
1385         return scan;
1386 }
1387
1388 /* ----------------
1389  *              heap_rescan             - restart a relation scan
1390  * ----------------
1391  */
1392 void
1393 heap_rescan(HeapScanDesc scan,
1394                         ScanKey key)
1395 {
1396         /*
1397          * unpin scan buffers
1398          */
1399         if (BufferIsValid(scan->rs_cbuf))
1400                 ReleaseBuffer(scan->rs_cbuf);
1401
1402         /*
1403          * reinitialize scan descriptor
1404          */
1405         initscan(scan, key, true);
1406 }
1407
1408 /* ----------------
1409  *              heap_endscan    - end relation scan
1410  *
1411  *              See how to integrate with index scans.
1412  *              Check handling if reldesc caching.
1413  * ----------------
1414  */
1415 void
1416 heap_endscan(HeapScanDesc scan)
1417 {
1418         /* Note: no locking manipulations needed */
1419
1420         /*
1421          * unpin scan buffers
1422          */
1423         if (BufferIsValid(scan->rs_cbuf))
1424                 ReleaseBuffer(scan->rs_cbuf);
1425
1426         /*
1427          * decrement relation reference count and free scan descriptor storage
1428          */
1429         RelationDecrementReferenceCount(scan->rs_rd);
1430
1431         if (scan->rs_key)
1432                 pfree(scan->rs_key);
1433
1434         if (scan->rs_strategy != NULL)
1435                 FreeAccessStrategy(scan->rs_strategy);
1436
1437         if (scan->rs_temp_snap)
1438                 UnregisterSnapshot(scan->rs_snapshot);
1439
1440         pfree(scan);
1441 }
1442
1443 /* ----------------
1444  *              heap_getnext    - retrieve next tuple in scan
1445  *
1446  *              Fix to work with index relations.
1447  *              We don't return the buffer anymore, but you can get it from the
1448  *              returned HeapTuple.
1449  * ----------------
1450  */
1451
1452 #ifdef HEAPDEBUGALL
1453 #define HEAPDEBUG_1 \
1454         elog(DEBUG2, "heap_getnext([%s,nkeys=%d],dir=%d) called", \
1455                  RelationGetRelationName(scan->rs_rd), scan->rs_nkeys, (int) direction)
1456 #define HEAPDEBUG_2 \
1457         elog(DEBUG2, "heap_getnext returning EOS")
1458 #define HEAPDEBUG_3 \
1459         elog(DEBUG2, "heap_getnext returning tuple")
1460 #else
1461 #define HEAPDEBUG_1
1462 #define HEAPDEBUG_2
1463 #define HEAPDEBUG_3
1464 #endif   /* !defined(HEAPDEBUGALL) */
1465
1466
1467 HeapTuple
1468 heap_getnext(HeapScanDesc scan, ScanDirection direction)
1469 {
1470         /* Note: no locking manipulations needed */
1471
1472         HEAPDEBUG_1;                            /* heap_getnext( info ) */
1473
1474         if (scan->rs_pageatatime)
1475                 heapgettup_pagemode(scan, direction,
1476                                                         scan->rs_nkeys, scan->rs_key);
1477         else
1478                 heapgettup(scan, direction, scan->rs_nkeys, scan->rs_key);
1479
1480         if (scan->rs_ctup.t_data == NULL)
1481         {
1482                 HEAPDEBUG_2;                    /* heap_getnext returning EOS */
1483                 return NULL;
1484         }
1485
1486         /*
1487          * if we get here it means we have a new current scan tuple, so point to
1488          * the proper return buffer and return the tuple.
1489          */
1490         HEAPDEBUG_3;                            /* heap_getnext returning tuple */
1491
1492         pgstat_count_heap_getnext(scan->rs_rd);
1493
1494         return &(scan->rs_ctup);
1495 }
1496
1497 /*
1498  *      heap_fetch              - retrieve tuple with given tid
1499  *
1500  * On entry, tuple->t_self is the TID to fetch.  We pin the buffer holding
1501  * the tuple, fill in the remaining fields of *tuple, and check the tuple
1502  * against the specified snapshot.
1503  *
1504  * If successful (tuple found and passes snapshot time qual), then *userbuf
1505  * is set to the buffer holding the tuple and TRUE is returned.  The caller
1506  * must unpin the buffer when done with the tuple.
1507  *
1508  * If the tuple is not found (ie, item number references a deleted slot),
1509  * then tuple->t_data is set to NULL and FALSE is returned.
1510  *
1511  * If the tuple is found but fails the time qual check, then FALSE is returned
1512  * but tuple->t_data is left pointing to the tuple.
1513  *
1514  * keep_buf determines what is done with the buffer in the FALSE-result cases.
1515  * When the caller specifies keep_buf = true, we retain the pin on the buffer
1516  * and return it in *userbuf (so the caller must eventually unpin it); when
1517  * keep_buf = false, the pin is released and *userbuf is set to InvalidBuffer.
1518  *
1519  * stats_relation is the relation to charge the heap_fetch operation against
1520  * for statistical purposes.  (This could be the heap rel itself, an
1521  * associated index, or NULL to not count the fetch at all.)
1522  *
1523  * heap_fetch does not follow HOT chains: only the exact TID requested will
1524  * be fetched.
1525  *
1526  * It is somewhat inconsistent that we ereport() on invalid block number but
1527  * return false on invalid item number.  There are a couple of reasons though.
1528  * One is that the caller can relatively easily check the block number for
1529  * validity, but cannot check the item number without reading the page
1530  * himself.  Another is that when we are following a t_ctid link, we can be
1531  * reasonably confident that the page number is valid (since VACUUM shouldn't
1532  * truncate off the destination page without having killed the referencing
1533  * tuple first), but the item number might well not be good.
1534  */
1535 bool
1536 heap_fetch(Relation relation,
1537                    Snapshot snapshot,
1538                    HeapTuple tuple,
1539                    Buffer *userbuf,
1540                    bool keep_buf,
1541                    Relation stats_relation)
1542 {
1543         ItemPointer tid = &(tuple->t_self);
1544         ItemId          lp;
1545         Buffer          buffer;
1546         Page            page;
1547         OffsetNumber offnum;
1548         bool            valid;
1549
1550         /*
1551          * Fetch and pin the appropriate page of the relation.
1552          */
1553         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1554
1555         /*
1556          * Need share lock on buffer to examine tuple commit status.
1557          */
1558         LockBuffer(buffer, BUFFER_LOCK_SHARE);
1559         page = BufferGetPage(buffer);
1560
1561         /*
1562          * We'd better check for out-of-range offnum in case of VACUUM since the
1563          * TID was obtained.
1564          */
1565         offnum = ItemPointerGetOffsetNumber(tid);
1566         if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1567         {
1568                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1569                 if (keep_buf)
1570                         *userbuf = buffer;
1571                 else
1572                 {
1573                         ReleaseBuffer(buffer);
1574                         *userbuf = InvalidBuffer;
1575                 }
1576                 tuple->t_data = NULL;
1577                 return false;
1578         }
1579
1580         /*
1581          * get the item line pointer corresponding to the requested tid
1582          */
1583         lp = PageGetItemId(page, offnum);
1584
1585         /*
1586          * Must check for deleted tuple.
1587          */
1588         if (!ItemIdIsNormal(lp))
1589         {
1590                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1591                 if (keep_buf)
1592                         *userbuf = buffer;
1593                 else
1594                 {
1595                         ReleaseBuffer(buffer);
1596                         *userbuf = InvalidBuffer;
1597                 }
1598                 tuple->t_data = NULL;
1599                 return false;
1600         }
1601
1602         /*
1603          * fill in *tuple fields
1604          */
1605         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1606         tuple->t_len = ItemIdGetLength(lp);
1607         tuple->t_tableOid = RelationGetRelid(relation);
1608
1609         /*
1610          * check time qualification of tuple, then release lock
1611          */
1612         valid = HeapTupleSatisfiesVisibility(tuple, snapshot, buffer);
1613
1614         if (valid)
1615                 PredicateLockTuple(relation, tuple, snapshot);
1616
1617         CheckForSerializableConflictOut(valid, relation, tuple, buffer, snapshot);
1618
1619         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1620
1621         if (valid)
1622         {
1623                 /*
1624                  * All checks passed, so return the tuple as valid. Caller is now
1625                  * responsible for releasing the buffer.
1626                  */
1627                 *userbuf = buffer;
1628
1629                 /* Count the successful fetch against appropriate rel, if any */
1630                 if (stats_relation != NULL)
1631                         pgstat_count_heap_fetch(stats_relation);
1632
1633                 return true;
1634         }
1635
1636         /* Tuple failed time qual, but maybe caller wants to see it anyway. */
1637         if (keep_buf)
1638                 *userbuf = buffer;
1639         else
1640         {
1641                 ReleaseBuffer(buffer);
1642                 *userbuf = InvalidBuffer;
1643         }
1644
1645         return false;
1646 }
1647
1648 /*
1649  *      heap_hot_search_buffer  - search HOT chain for tuple satisfying snapshot
1650  *
1651  * On entry, *tid is the TID of a tuple (either a simple tuple, or the root
1652  * of a HOT chain), and buffer is the buffer holding this tuple.  We search
1653  * for the first chain member satisfying the given snapshot.  If one is
1654  * found, we update *tid to reference that tuple's offset number, and
1655  * return TRUE.  If no match, return FALSE without modifying *tid.
1656  *
1657  * heapTuple is a caller-supplied buffer.  When a match is found, we return
1658  * the tuple here, in addition to updating *tid.  If no match is found, the
1659  * contents of this buffer on return are undefined.
1660  *
1661  * If all_dead is not NULL, we check non-visible tuples to see if they are
1662  * globally dead; *all_dead is set TRUE if all members of the HOT chain
1663  * are vacuumable, FALSE if not.
1664  *
1665  * Unlike heap_fetch, the caller must already have pin and (at least) share
1666  * lock on the buffer; it is still pinned/locked at exit.  Also unlike
1667  * heap_fetch, we do not report any pgstats count; caller may do so if wanted.
1668  */
1669 bool
1670 heap_hot_search_buffer(ItemPointer tid, Relation relation, Buffer buffer,
1671                                            Snapshot snapshot, HeapTuple heapTuple,
1672                                            bool *all_dead, bool first_call)
1673 {
1674         Page            dp = (Page) BufferGetPage(buffer);
1675         TransactionId prev_xmax = InvalidTransactionId;
1676         OffsetNumber offnum;
1677         bool            at_chain_start;
1678         bool            valid;
1679         bool            skip;
1680
1681         /* If this is not the first call, previous call returned a (live!) tuple */
1682         if (all_dead)
1683                 *all_dead = first_call;
1684
1685         Assert(TransactionIdIsValid(RecentGlobalXmin));
1686
1687         Assert(ItemPointerGetBlockNumber(tid) == BufferGetBlockNumber(buffer));
1688         offnum = ItemPointerGetOffsetNumber(tid);
1689         at_chain_start = first_call;
1690         skip = !first_call;
1691
1692         heapTuple->t_self = *tid;
1693
1694         /* Scan through possible multiple members of HOT-chain */
1695         for (;;)
1696         {
1697                 ItemId          lp;
1698
1699                 /* check for bogus TID */
1700                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(dp))
1701                         break;
1702
1703                 lp = PageGetItemId(dp, offnum);
1704
1705                 /* check for unused, dead, or redirected items */
1706                 if (!ItemIdIsNormal(lp))
1707                 {
1708                         /* We should only see a redirect at start of chain */
1709                         if (ItemIdIsRedirected(lp) && at_chain_start)
1710                         {
1711                                 /* Follow the redirect */
1712                                 offnum = ItemIdGetRedirect(lp);
1713                                 at_chain_start = false;
1714                                 continue;
1715                         }
1716                         /* else must be end of chain */
1717                         break;
1718                 }
1719
1720                 heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
1721                 heapTuple->t_len = ItemIdGetLength(lp);
1722                 heapTuple->t_tableOid = RelationGetRelid(relation);
1723                 ItemPointerSetOffsetNumber(&heapTuple->t_self, offnum);
1724
1725                 /*
1726                  * Shouldn't see a HEAP_ONLY tuple at chain start.
1727                  */
1728                 if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
1729                         break;
1730
1731                 /*
1732                  * The xmin should match the previous xmax value, else chain is
1733                  * broken.
1734                  */
1735                 if (TransactionIdIsValid(prev_xmax) &&
1736                         !TransactionIdEquals(prev_xmax,
1737                                                                  HeapTupleHeaderGetXmin(heapTuple->t_data)))
1738                         break;
1739
1740                 /*
1741                  * When first_call is true (and thus, skip is initially false) we'll
1742                  * return the first tuple we find.  But on later passes, heapTuple
1743                  * will initially be pointing to the tuple we returned last time.
1744                  * Returning it again would be incorrect (and would loop forever), so
1745                  * we skip it and return the next match we find.
1746                  */
1747                 if (!skip)
1748                 {
1749                         /*
1750                          * For the benefit of logical decoding, have t_self point at the
1751                          * element of the HOT chain we're currently investigating instead
1752                          * of the root tuple of the HOT chain. This is important because
1753                          * the *Satisfies routine for historical mvcc snapshots needs the
1754                          * correct tid to decide about the visibility in some cases.
1755                          */
1756                         ItemPointerSet(&(heapTuple->t_self), BufferGetBlockNumber(buffer), offnum);
1757
1758                         /* If it's visible per the snapshot, we must return it */
1759                         valid = HeapTupleSatisfiesVisibility(heapTuple, snapshot, buffer);
1760                         CheckForSerializableConflictOut(valid, relation, heapTuple,
1761                                                                                         buffer, snapshot);
1762                         /* reset to original, non-redirected, tid */
1763                         heapTuple->t_self = *tid;
1764
1765                         if (valid)
1766                         {
1767                                 ItemPointerSetOffsetNumber(tid, offnum);
1768                                 PredicateLockTuple(relation, heapTuple, snapshot);
1769                                 if (all_dead)
1770                                         *all_dead = false;
1771                                 return true;
1772                         }
1773                 }
1774                 skip = false;
1775
1776                 /*
1777                  * If we can't see it, maybe no one else can either.  At caller
1778                  * request, check whether all chain members are dead to all
1779                  * transactions.
1780                  */
1781                 if (all_dead && *all_dead &&
1782                         !HeapTupleIsSurelyDead(heapTuple, RecentGlobalXmin))
1783                         *all_dead = false;
1784
1785                 /*
1786                  * Check to see if HOT chain continues past this tuple; if so fetch
1787                  * the next offnum and loop around.
1788                  */
1789                 if (HeapTupleIsHotUpdated(heapTuple))
1790                 {
1791                         Assert(ItemPointerGetBlockNumber(&heapTuple->t_data->t_ctid) ==
1792                                    ItemPointerGetBlockNumber(tid));
1793                         offnum = ItemPointerGetOffsetNumber(&heapTuple->t_data->t_ctid);
1794                         at_chain_start = false;
1795                         prev_xmax = HeapTupleHeaderGetUpdateXid(heapTuple->t_data);
1796                 }
1797                 else
1798                         break;                          /* end of chain */
1799         }
1800
1801         return false;
1802 }
1803
1804 /*
1805  *      heap_hot_search         - search HOT chain for tuple satisfying snapshot
1806  *
1807  * This has the same API as heap_hot_search_buffer, except that the caller
1808  * does not provide the buffer containing the page, rather we access it
1809  * locally.
1810  */
1811 bool
1812 heap_hot_search(ItemPointer tid, Relation relation, Snapshot snapshot,
1813                                 bool *all_dead)
1814 {
1815         bool            result;
1816         Buffer          buffer;
1817         HeapTupleData heapTuple;
1818
1819         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1820         LockBuffer(buffer, BUFFER_LOCK_SHARE);
1821         result = heap_hot_search_buffer(tid, relation, buffer, snapshot,
1822                                                                         &heapTuple, all_dead, true);
1823         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1824         ReleaseBuffer(buffer);
1825         return result;
1826 }
1827
1828 /*
1829  *      heap_get_latest_tid -  get the latest tid of a specified tuple
1830  *
1831  * Actually, this gets the latest version that is visible according to
1832  * the passed snapshot.  You can pass SnapshotDirty to get the very latest,
1833  * possibly uncommitted version.
1834  *
1835  * *tid is both an input and an output parameter: it is updated to
1836  * show the latest version of the row.  Note that it will not be changed
1837  * if no version of the row passes the snapshot test.
1838  */
1839 void
1840 heap_get_latest_tid(Relation relation,
1841                                         Snapshot snapshot,
1842                                         ItemPointer tid)
1843 {
1844         BlockNumber blk;
1845         ItemPointerData ctid;
1846         TransactionId priorXmax;
1847
1848         /* this is to avoid Assert failures on bad input */
1849         if (!ItemPointerIsValid(tid))
1850                 return;
1851
1852         /*
1853          * Since this can be called with user-supplied TID, don't trust the input
1854          * too much.  (RelationGetNumberOfBlocks is an expensive check, so we
1855          * don't check t_ctid links again this way.  Note that it would not do to
1856          * call it just once and save the result, either.)
1857          */
1858         blk = ItemPointerGetBlockNumber(tid);
1859         if (blk >= RelationGetNumberOfBlocks(relation))
1860                 elog(ERROR, "block number %u is out of range for relation \"%s\"",
1861                          blk, RelationGetRelationName(relation));
1862
1863         /*
1864          * Loop to chase down t_ctid links.  At top of loop, ctid is the tuple we
1865          * need to examine, and *tid is the TID we will return if ctid turns out
1866          * to be bogus.
1867          *
1868          * Note that we will loop until we reach the end of the t_ctid chain.
1869          * Depending on the snapshot passed, there might be at most one visible
1870          * version of the row, but we don't try to optimize for that.
1871          */
1872         ctid = *tid;
1873         priorXmax = InvalidTransactionId;       /* cannot check first XMIN */
1874         for (;;)
1875         {
1876                 Buffer          buffer;
1877                 Page            page;
1878                 OffsetNumber offnum;
1879                 ItemId          lp;
1880                 HeapTupleData tp;
1881                 bool            valid;
1882
1883                 /*
1884                  * Read, pin, and lock the page.
1885                  */
1886                 buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&ctid));
1887                 LockBuffer(buffer, BUFFER_LOCK_SHARE);
1888                 page = BufferGetPage(buffer);
1889
1890                 /*
1891                  * Check for bogus item number.  This is not treated as an error
1892                  * condition because it can happen while following a t_ctid link. We
1893                  * just assume that the prior tid is OK and return it unchanged.
1894                  */
1895                 offnum = ItemPointerGetOffsetNumber(&ctid);
1896                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1897                 {
1898                         UnlockReleaseBuffer(buffer);
1899                         break;
1900                 }
1901                 lp = PageGetItemId(page, offnum);
1902                 if (!ItemIdIsNormal(lp))
1903                 {
1904                         UnlockReleaseBuffer(buffer);
1905                         break;
1906                 }
1907
1908                 /* OK to access the tuple */
1909                 tp.t_self = ctid;
1910                 tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
1911                 tp.t_len = ItemIdGetLength(lp);
1912                 tp.t_tableOid = RelationGetRelid(relation);
1913
1914                 /*
1915                  * After following a t_ctid link, we might arrive at an unrelated
1916                  * tuple.  Check for XMIN match.
1917                  */
1918                 if (TransactionIdIsValid(priorXmax) &&
1919                   !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(tp.t_data)))
1920                 {
1921                         UnlockReleaseBuffer(buffer);
1922                         break;
1923                 }
1924
1925                 /*
1926                  * Check time qualification of tuple; if visible, set it as the new
1927                  * result candidate.
1928                  */
1929                 valid = HeapTupleSatisfiesVisibility(&tp, snapshot, buffer);
1930                 CheckForSerializableConflictOut(valid, relation, &tp, buffer, snapshot);
1931                 if (valid)
1932                         *tid = ctid;
1933
1934                 /*
1935                  * If there's a valid t_ctid link, follow it, else we're done.
1936                  */
1937                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
1938                         HeapTupleHeaderIsOnlyLocked(tp.t_data) ||
1939                         ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
1940                 {
1941                         UnlockReleaseBuffer(buffer);
1942                         break;
1943                 }
1944
1945                 ctid = tp.t_data->t_ctid;
1946                 priorXmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
1947                 UnlockReleaseBuffer(buffer);
1948         }                                                       /* end of loop */
1949 }
1950
1951
1952 /*
1953  * UpdateXmaxHintBits - update tuple hint bits after xmax transaction ends
1954  *
1955  * This is called after we have waited for the XMAX transaction to terminate.
1956  * If the transaction aborted, we guarantee the XMAX_INVALID hint bit will
1957  * be set on exit.  If the transaction committed, we set the XMAX_COMMITTED
1958  * hint bit if possible --- but beware that that may not yet be possible,
1959  * if the transaction committed asynchronously.
1960  *
1961  * Note that if the transaction was a locker only, we set HEAP_XMAX_INVALID
1962  * even if it commits.
1963  *
1964  * Hence callers should look only at XMAX_INVALID.
1965  *
1966  * Note this is not allowed for tuples whose xmax is a multixact.
1967  */
1968 static void
1969 UpdateXmaxHintBits(HeapTupleHeader tuple, Buffer buffer, TransactionId xid)
1970 {
1971         Assert(TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple), xid));
1972         Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));
1973
1974         if (!(tuple->t_infomask & (HEAP_XMAX_COMMITTED | HEAP_XMAX_INVALID)))
1975         {
1976                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
1977                         TransactionIdDidCommit(xid))
1978                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_COMMITTED,
1979                                                                  xid);
1980                 else
1981                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_INVALID,
1982                                                                  InvalidTransactionId);
1983         }
1984 }
1985
1986
1987 /*
1988  * GetBulkInsertState - prepare status object for a bulk insert
1989  */
1990 BulkInsertState
1991 GetBulkInsertState(void)
1992 {
1993         BulkInsertState bistate;
1994
1995         bistate = (BulkInsertState) palloc(sizeof(BulkInsertStateData));
1996         bistate->strategy = GetAccessStrategy(BAS_BULKWRITE);
1997         bistate->current_buf = InvalidBuffer;
1998         return bistate;
1999 }
2000
2001 /*
2002  * FreeBulkInsertState - clean up after finishing a bulk insert
2003  */
2004 void
2005 FreeBulkInsertState(BulkInsertState bistate)
2006 {
2007         if (bistate->current_buf != InvalidBuffer)
2008                 ReleaseBuffer(bistate->current_buf);
2009         FreeAccessStrategy(bistate->strategy);
2010         pfree(bistate);
2011 }
2012
2013
2014 /*
2015  *      heap_insert             - insert tuple into a heap
2016  *
2017  * The new tuple is stamped with current transaction ID and the specified
2018  * command ID.
2019  *
2020  * If the HEAP_INSERT_SKIP_WAL option is specified, the new tuple is not
2021  * logged in WAL, even for a non-temp relation.  Safe usage of this behavior
2022  * requires that we arrange that all new tuples go into new pages not
2023  * containing any tuples from other transactions, and that the relation gets
2024  * fsync'd before commit.  (See also heap_sync() comments)
2025  *
2026  * The HEAP_INSERT_SKIP_FSM option is passed directly to
2027  * RelationGetBufferForTuple, which see for more info.
2028  *
2029  * HEAP_INSERT_FROZEN should only be specified for inserts into
2030  * relfilenodes created during the current subtransaction and when
2031  * there are no prior snapshots or pre-existing portals open.
2032  * This causes rows to be frozen, which is an MVCC violation and
2033  * requires explicit options chosen by user.
2034  *
2035  * Note that these options will be applied when inserting into the heap's
2036  * TOAST table, too, if the tuple requires any out-of-line data.
2037  *
2038  * The BulkInsertState object (if any; bistate can be NULL for default
2039  * behavior) is also just passed through to RelationGetBufferForTuple.
2040  *
2041  * The return value is the OID assigned to the tuple (either here or by the
2042  * caller), or InvalidOid if no OID.  The header fields of *tup are updated
2043  * to match the stored tuple; in particular tup->t_self receives the actual
2044  * TID where the tuple was stored.  But note that any toasting of fields
2045  * within the tuple data is NOT reflected into *tup.
2046  */
2047 Oid
2048 heap_insert(Relation relation, HeapTuple tup, CommandId cid,
2049                         int options, BulkInsertState bistate)
2050 {
2051         TransactionId xid = GetCurrentTransactionId();
2052         HeapTuple       heaptup;
2053         Buffer          buffer;
2054         Buffer          vmbuffer = InvalidBuffer;
2055         bool            all_visible_cleared = false;
2056
2057         /*
2058          * Fill in tuple header fields, assign an OID, and toast the tuple if
2059          * necessary.
2060          *
2061          * Note: below this point, heaptup is the data we actually intend to store
2062          * into the relation; tup is the caller's original untoasted data.
2063          */
2064         heaptup = heap_prepare_insert(relation, tup, xid, cid, options);
2065
2066         /*
2067          * We're about to do the actual insert -- but check for conflict first, to
2068          * avoid possibly having to roll back work we've just done.
2069          *
2070          * For a heap insert, we only need to check for table-level SSI locks. Our
2071          * new tuple can't possibly conflict with existing tuple locks, and heap
2072          * page locks are only consolidated versions of tuple locks; they do not
2073          * lock "gaps" as index page locks do.  So we don't need to identify a
2074          * buffer before making the call.
2075          */
2076         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2077
2078         /*
2079          * Find buffer to insert this tuple into.  If the page is all visible,
2080          * this will also pin the requisite visibility map page.
2081          */
2082         buffer = RelationGetBufferForTuple(relation, heaptup->t_len,
2083                                                                            InvalidBuffer, options, bistate,
2084                                                                            &vmbuffer, NULL);
2085
2086         /* NO EREPORT(ERROR) from here till changes are logged */
2087         START_CRIT_SECTION();
2088
2089         RelationPutHeapTuple(relation, buffer, heaptup);
2090
2091         if (PageIsAllVisible(BufferGetPage(buffer)))
2092         {
2093                 all_visible_cleared = true;
2094                 PageClearAllVisible(BufferGetPage(buffer));
2095                 visibilitymap_clear(relation,
2096                                                         ItemPointerGetBlockNumber(&(heaptup->t_self)),
2097                                                         vmbuffer);
2098         }
2099
2100         /*
2101          * XXX Should we set PageSetPrunable on this page ?
2102          *
2103          * The inserting transaction may eventually abort thus making this tuple
2104          * DEAD and hence available for pruning. Though we don't want to optimize
2105          * for aborts, if no other tuple in this page is UPDATEd/DELETEd, the
2106          * aborted tuple will never be pruned until next vacuum is triggered.
2107          *
2108          * If you do add PageSetPrunable here, add it in heap_xlog_insert too.
2109          */
2110
2111         MarkBufferDirty(buffer);
2112
2113         /* XLOG stuff */
2114         if (!(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation))
2115         {
2116                 xl_heap_insert xlrec;
2117                 xl_heap_header xlhdr;
2118                 XLogRecPtr      recptr;
2119                 XLogRecData rdata[4];
2120                 Page            page = BufferGetPage(buffer);
2121                 uint8           info = XLOG_HEAP_INSERT;
2122                 bool            need_tuple_data;
2123
2124                 /*
2125                  * For logical decoding, we need the tuple even if we're doing a full
2126                  * page write, so make sure to log it separately. (XXX We could
2127                  * alternatively store a pointer into the FPW).
2128                  *
2129                  * Also, if this is a catalog, we need to transmit combocids to
2130                  * properly decode, so log that as well.
2131                  */
2132                 need_tuple_data = RelationIsLogicallyLogged(relation);
2133                 if (RelationIsAccessibleInLogicalDecoding(relation))
2134                         log_heap_new_cid(relation, heaptup);
2135
2136                 xlrec.flags = all_visible_cleared ? XLOG_HEAP_ALL_VISIBLE_CLEARED : 0;
2137                 xlrec.target.node = relation->rd_node;
2138                 xlrec.target.tid = heaptup->t_self;
2139                 rdata[0].data = (char *) &xlrec;
2140                 rdata[0].len = SizeOfHeapInsert;
2141                 rdata[0].buffer = InvalidBuffer;
2142                 rdata[0].next = &(rdata[1]);
2143
2144                 xlhdr.t_infomask2 = heaptup->t_data->t_infomask2;
2145                 xlhdr.t_infomask = heaptup->t_data->t_infomask;
2146                 xlhdr.t_hoff = heaptup->t_data->t_hoff;
2147
2148                 /*
2149                  * note we mark rdata[1] as belonging to buffer; if XLogInsert decides
2150                  * to write the whole page to the xlog, we don't need to store
2151                  * xl_heap_header in the xlog.
2152                  */
2153                 rdata[1].data = (char *) &xlhdr;
2154                 rdata[1].len = SizeOfHeapHeader;
2155                 rdata[1].buffer = need_tuple_data ? InvalidBuffer : buffer;
2156                 rdata[1].buffer_std = true;
2157                 rdata[1].next = &(rdata[2]);
2158
2159                 /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
2160                 rdata[2].data = (char *) heaptup->t_data + offsetof(HeapTupleHeaderData, t_bits);
2161                 rdata[2].len = heaptup->t_len - offsetof(HeapTupleHeaderData, t_bits);
2162                 rdata[2].buffer = need_tuple_data ? InvalidBuffer : buffer;
2163                 rdata[2].buffer_std = true;
2164                 rdata[2].next = NULL;
2165
2166                 /*
2167                  * Make a separate rdata entry for the tuple's buffer if we're doing
2168                  * logical decoding, so that an eventual FPW doesn't remove the
2169                  * tuple's data.
2170                  */
2171                 if (need_tuple_data)
2172                 {
2173                         rdata[2].next = &(rdata[3]);
2174
2175                         rdata[3].data = NULL;
2176                         rdata[3].len = 0;
2177                         rdata[3].buffer = buffer;
2178                         rdata[3].buffer_std = true;
2179                         rdata[3].next = NULL;
2180
2181                         xlrec.flags |= XLOG_HEAP_CONTAINS_NEW_TUPLE;
2182                 }
2183
2184                 /*
2185                  * If this is the single and first tuple on page, we can reinit the
2186                  * page instead of restoring the whole thing.  Set flag, and hide
2187                  * buffer references from XLogInsert.
2188                  */
2189                 if (ItemPointerGetOffsetNumber(&(heaptup->t_self)) == FirstOffsetNumber &&
2190                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
2191                 {
2192                         info |= XLOG_HEAP_INIT_PAGE;
2193                         rdata[1].buffer = rdata[2].buffer = rdata[3].buffer = InvalidBuffer;
2194                 }
2195
2196                 recptr = XLogInsert(RM_HEAP_ID, info, rdata);
2197
2198                 PageSetLSN(page, recptr);
2199         }
2200
2201         END_CRIT_SECTION();
2202
2203         UnlockReleaseBuffer(buffer);
2204         if (vmbuffer != InvalidBuffer)
2205                 ReleaseBuffer(vmbuffer);
2206
2207         /*
2208          * If tuple is cachable, mark it for invalidation from the caches in case
2209          * we abort.  Note it is OK to do this after releasing the buffer, because
2210          * the heaptup data structure is all in local memory, not in the shared
2211          * buffer.
2212          */
2213         CacheInvalidateHeapTuple(relation, heaptup, NULL);
2214
2215         pgstat_count_heap_insert(relation, 1);
2216
2217         /*
2218          * If heaptup is a private copy, release it.  Don't forget to copy t_self
2219          * back to the caller's image, too.
2220          */
2221         if (heaptup != tup)
2222         {
2223                 tup->t_self = heaptup->t_self;
2224                 heap_freetuple(heaptup);
2225         }
2226
2227         return HeapTupleGetOid(tup);
2228 }
2229
2230 /*
2231  * Subroutine for heap_insert(). Prepares a tuple for insertion. This sets the
2232  * tuple header fields, assigns an OID, and toasts the tuple if necessary.
2233  * Returns a toasted version of the tuple if it was toasted, or the original
2234  * tuple if not. Note that in any case, the header fields are also set in
2235  * the original tuple.
2236  */
2237 static HeapTuple
2238 heap_prepare_insert(Relation relation, HeapTuple tup, TransactionId xid,
2239                                         CommandId cid, int options)
2240 {
2241         if (relation->rd_rel->relhasoids)
2242         {
2243 #ifdef NOT_USED
2244                 /* this is redundant with an Assert in HeapTupleSetOid */
2245                 Assert(tup->t_data->t_infomask & HEAP_HASOID);
2246 #endif
2247
2248                 /*
2249                  * If the object id of this tuple has already been assigned, trust the
2250                  * caller.  There are a couple of ways this can happen.  At initial db
2251                  * creation, the backend program sets oids for tuples. When we define
2252                  * an index, we set the oid.  Finally, in the future, we may allow
2253                  * users to set their own object ids in order to support a persistent
2254                  * object store (objects need to contain pointers to one another).
2255                  */
2256                 if (!OidIsValid(HeapTupleGetOid(tup)))
2257                         HeapTupleSetOid(tup, GetNewOid(relation));
2258         }
2259         else
2260         {
2261                 /* check there is not space for an OID */
2262                 Assert(!(tup->t_data->t_infomask & HEAP_HASOID));
2263         }
2264
2265         tup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
2266         tup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
2267         tup->t_data->t_infomask |= HEAP_XMAX_INVALID;
2268         HeapTupleHeaderSetXmin(tup->t_data, xid);
2269         if (options & HEAP_INSERT_FROZEN)
2270                 HeapTupleHeaderSetXminFrozen(tup->t_data);
2271
2272         HeapTupleHeaderSetCmin(tup->t_data, cid);
2273         HeapTupleHeaderSetXmax(tup->t_data, 0);         /* for cleanliness */
2274         tup->t_tableOid = RelationGetRelid(relation);
2275
2276         /*
2277          * If the new tuple is too big for storage or contains already toasted
2278          * out-of-line attributes from some other relation, invoke the toaster.
2279          */
2280         if (relation->rd_rel->relkind != RELKIND_RELATION &&
2281                 relation->rd_rel->relkind != RELKIND_MATVIEW)
2282         {
2283                 /* toast table entries should never be recursively toasted */
2284                 Assert(!HeapTupleHasExternal(tup));
2285                 return tup;
2286         }
2287         else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
2288                 return toast_insert_or_update(relation, tup, NULL, options);
2289         else
2290                 return tup;
2291 }
2292
2293 /*
2294  *      heap_multi_insert       - insert multiple tuple into a heap
2295  *
2296  * This is like heap_insert(), but inserts multiple tuples in one operation.
2297  * That's faster than calling heap_insert() in a loop, because when multiple
2298  * tuples can be inserted on a single page, we can write just a single WAL
2299  * record covering all of them, and only need to lock/unlock the page once.
2300  *
2301  * Note: this leaks memory into the current memory context. You can create a
2302  * temporary context before calling this, if that's a problem.
2303  */
2304 void
2305 heap_multi_insert(Relation relation, HeapTuple *tuples, int ntuples,
2306                                   CommandId cid, int options, BulkInsertState bistate)
2307 {
2308         TransactionId xid = GetCurrentTransactionId();
2309         HeapTuple  *heaptuples;
2310         int                     i;
2311         int                     ndone;
2312         char       *scratch = NULL;
2313         Page            page;
2314         bool            needwal;
2315         Size            saveFreeSpace;
2316         bool            need_tuple_data = RelationIsLogicallyLogged(relation);
2317         bool            need_cids = RelationIsAccessibleInLogicalDecoding(relation);
2318
2319         needwal = !(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation);
2320         saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
2321                                                                                                    HEAP_DEFAULT_FILLFACTOR);
2322
2323         /* Toast and set header data in all the tuples */
2324         heaptuples = palloc(ntuples * sizeof(HeapTuple));
2325         for (i = 0; i < ntuples; i++)
2326                 heaptuples[i] = heap_prepare_insert(relation, tuples[i],
2327                                                                                         xid, cid, options);
2328
2329         /*
2330          * Allocate some memory to use for constructing the WAL record. Using
2331          * palloc() within a critical section is not safe, so we allocate this
2332          * beforehand.
2333          */
2334         if (needwal)
2335                 scratch = palloc(BLCKSZ);
2336
2337         /*
2338          * We're about to do the actual inserts -- but check for conflict first,
2339          * to avoid possibly having to roll back work we've just done.
2340          *
2341          * For a heap insert, we only need to check for table-level SSI locks. Our
2342          * new tuple can't possibly conflict with existing tuple locks, and heap
2343          * page locks are only consolidated versions of tuple locks; they do not
2344          * lock "gaps" as index page locks do.  So we don't need to identify a
2345          * buffer before making the call.
2346          */
2347         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2348
2349         ndone = 0;
2350         while (ndone < ntuples)
2351         {
2352                 Buffer          buffer;
2353                 Buffer          vmbuffer = InvalidBuffer;
2354                 bool            all_visible_cleared = false;
2355                 int                     nthispage;
2356
2357                 CHECK_FOR_INTERRUPTS();
2358
2359                 /*
2360                  * Find buffer where at least the next tuple will fit.  If the page is
2361                  * all-visible, this will also pin the requisite visibility map page.
2362                  */
2363                 buffer = RelationGetBufferForTuple(relation, heaptuples[ndone]->t_len,
2364                                                                                    InvalidBuffer, options, bistate,
2365                                                                                    &vmbuffer, NULL);
2366                 page = BufferGetPage(buffer);
2367
2368                 /* NO EREPORT(ERROR) from here till changes are logged */
2369                 START_CRIT_SECTION();
2370
2371                 /*
2372                  * RelationGetBufferForTuple has ensured that the first tuple fits.
2373                  * Put that on the page, and then as many other tuples as fit.
2374                  */
2375                 RelationPutHeapTuple(relation, buffer, heaptuples[ndone]);
2376                 for (nthispage = 1; ndone + nthispage < ntuples; nthispage++)
2377                 {
2378                         HeapTuple       heaptup = heaptuples[ndone + nthispage];
2379
2380                         if (PageGetHeapFreeSpace(page) < MAXALIGN(heaptup->t_len) + saveFreeSpace)
2381                                 break;
2382
2383                         RelationPutHeapTuple(relation, buffer, heaptup);
2384                 }
2385
2386                 if (PageIsAllVisible(page))
2387                 {
2388                         all_visible_cleared = true;
2389                         PageClearAllVisible(page);
2390                         visibilitymap_clear(relation,
2391                                                                 BufferGetBlockNumber(buffer),
2392                                                                 vmbuffer);
2393                 }
2394
2395                 /*
2396                  * XXX Should we set PageSetPrunable on this page ? See heap_insert()
2397                  */
2398
2399                 MarkBufferDirty(buffer);
2400
2401                 /* XLOG stuff */
2402                 if (needwal)
2403                 {
2404                         XLogRecPtr      recptr;
2405                         xl_heap_multi_insert *xlrec;
2406                         XLogRecData rdata[3];
2407                         uint8           info = XLOG_HEAP2_MULTI_INSERT;
2408                         char       *tupledata;
2409                         int                     totaldatalen;
2410                         char       *scratchptr = scratch;
2411                         bool            init;
2412
2413                         /*
2414                          * If the page was previously empty, we can reinit the page
2415                          * instead of restoring the whole thing.
2416                          */
2417                         init = (ItemPointerGetOffsetNumber(&(heaptuples[ndone]->t_self)) == FirstOffsetNumber &&
2418                                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber + nthispage - 1);
2419
2420                         /* allocate xl_heap_multi_insert struct from the scratch area */
2421                         xlrec = (xl_heap_multi_insert *) scratchptr;
2422                         scratchptr += SizeOfHeapMultiInsert;
2423
2424                         /*
2425                          * Allocate offsets array. Unless we're reinitializing the page,
2426                          * in that case the tuples are stored in order starting at
2427                          * FirstOffsetNumber and we don't need to store the offsets
2428                          * explicitly.
2429                          */
2430                         if (!init)
2431                                 scratchptr += nthispage * sizeof(OffsetNumber);
2432
2433                         /* the rest of the scratch space is used for tuple data */
2434                         tupledata = scratchptr;
2435
2436                         xlrec->flags = all_visible_cleared ? XLOG_HEAP_ALL_VISIBLE_CLEARED : 0;
2437                         xlrec->node = relation->rd_node;
2438                         xlrec->blkno = BufferGetBlockNumber(buffer);
2439                         xlrec->ntuples = nthispage;
2440
2441                         /*
2442                          * Write out an xl_multi_insert_tuple and the tuple data itself
2443                          * for each tuple.
2444                          */
2445                         for (i = 0; i < nthispage; i++)
2446                         {
2447                                 HeapTuple       heaptup = heaptuples[ndone + i];
2448                                 xl_multi_insert_tuple *tuphdr;
2449                                 int                     datalen;
2450
2451                                 if (!init)
2452                                         xlrec->offsets[i] = ItemPointerGetOffsetNumber(&heaptup->t_self);
2453                                 /* xl_multi_insert_tuple needs two-byte alignment. */
2454                                 tuphdr = (xl_multi_insert_tuple *) SHORTALIGN(scratchptr);
2455                                 scratchptr = ((char *) tuphdr) + SizeOfMultiInsertTuple;
2456
2457                                 tuphdr->t_infomask2 = heaptup->t_data->t_infomask2;
2458                                 tuphdr->t_infomask = heaptup->t_data->t_infomask;
2459                                 tuphdr->t_hoff = heaptup->t_data->t_hoff;
2460
2461                                 /* write bitmap [+ padding] [+ oid] + data */
2462                                 datalen = heaptup->t_len - offsetof(HeapTupleHeaderData, t_bits);
2463                                 memcpy(scratchptr,
2464                                            (char *) heaptup->t_data + offsetof(HeapTupleHeaderData, t_bits),
2465                                            datalen);
2466                                 tuphdr->datalen = datalen;
2467                                 scratchptr += datalen;
2468
2469                                 /*
2470                                  * We don't use heap_multi_insert for catalog tuples yet, but
2471                                  * better be prepared...
2472                                  */
2473                                 if (need_cids)
2474                                         log_heap_new_cid(relation, heaptup);
2475                         }
2476                         totaldatalen = scratchptr - tupledata;
2477                         Assert((scratchptr - scratch) < BLCKSZ);
2478
2479                         rdata[0].data = (char *) xlrec;
2480                         rdata[0].len = tupledata - scratch;
2481                         rdata[0].buffer = InvalidBuffer;
2482                         rdata[0].next = &rdata[1];
2483
2484                         rdata[1].data = tupledata;
2485                         rdata[1].len = totaldatalen;
2486                         rdata[1].buffer = need_tuple_data ? InvalidBuffer : buffer;
2487                         rdata[1].buffer_std = true;
2488                         rdata[1].next = NULL;
2489
2490                         /*
2491                          * Make a separate rdata entry for the tuple's buffer if we're
2492                          * doing logical decoding, so that an eventual FPW doesn't remove
2493                          * the tuple's data.
2494                          */
2495                         if (need_tuple_data)
2496                         {
2497                                 rdata[1].next = &(rdata[2]);
2498
2499                                 rdata[2].data = NULL;
2500                                 rdata[2].len = 0;
2501                                 rdata[2].buffer = buffer;
2502                                 rdata[2].buffer_std = true;
2503                                 rdata[2].next = NULL;
2504                                 xlrec->flags |= XLOG_HEAP_CONTAINS_NEW_TUPLE;
2505                         }
2506
2507                         /*
2508                          * If we're going to reinitialize the whole page using the WAL
2509                          * record, hide buffer reference from XLogInsert.
2510                          */
2511                         if (init)
2512                         {
2513                                 rdata[1].buffer = rdata[2].buffer = InvalidBuffer;
2514                                 info |= XLOG_HEAP_INIT_PAGE;
2515                         }
2516
2517                         /*
2518                          * Signal that this is the last xl_heap_multi_insert record
2519                          * emitted by this call to heap_multi_insert(). Needed for logical
2520                          * decoding so it knows when to cleanup temporary data.
2521                          */
2522                         if (ndone + nthispage == ntuples)
2523                                 xlrec->flags |= XLOG_HEAP_LAST_MULTI_INSERT;
2524
2525                         recptr = XLogInsert(RM_HEAP2_ID, info, rdata);
2526
2527                         PageSetLSN(page, recptr);
2528                 }
2529
2530                 END_CRIT_SECTION();
2531
2532                 UnlockReleaseBuffer(buffer);
2533                 if (vmbuffer != InvalidBuffer)
2534                         ReleaseBuffer(vmbuffer);
2535
2536                 ndone += nthispage;
2537         }
2538
2539         /*
2540          * If tuples are cachable, mark them for invalidation from the caches in
2541          * case we abort.  Note it is OK to do this after releasing the buffer,
2542          * because the heaptuples data structure is all in local memory, not in
2543          * the shared buffer.
2544          */
2545         if (IsCatalogRelation(relation))
2546         {
2547                 for (i = 0; i < ntuples; i++)
2548                         CacheInvalidateHeapTuple(relation, heaptuples[i], NULL);
2549         }
2550
2551         /*
2552          * Copy t_self fields back to the caller's original tuples. This does
2553          * nothing for untoasted tuples (tuples[i] == heaptuples[i)], but it's
2554          * probably faster to always copy than check.
2555          */
2556         for (i = 0; i < ntuples; i++)
2557                 tuples[i]->t_self = heaptuples[i]->t_self;
2558
2559         pgstat_count_heap_insert(relation, ntuples);
2560 }
2561
2562 /*
2563  *      simple_heap_insert - insert a tuple
2564  *
2565  * Currently, this routine differs from heap_insert only in supplying
2566  * a default command ID and not allowing access to the speedup options.
2567  *
2568  * This should be used rather than using heap_insert directly in most places
2569  * where we are modifying system catalogs.
2570  */
2571 Oid
2572 simple_heap_insert(Relation relation, HeapTuple tup)
2573 {
2574         return heap_insert(relation, tup, GetCurrentCommandId(true), 0, NULL);
2575 }
2576
2577 /*
2578  * Given infomask/infomask2, compute the bits that must be saved in the
2579  * "infobits" field of xl_heap_delete, xl_heap_update, xl_heap_lock,
2580  * xl_heap_lock_updated WAL records.
2581  *
2582  * See fix_infomask_from_infobits.
2583  */
2584 static uint8
2585 compute_infobits(uint16 infomask, uint16 infomask2)
2586 {
2587         return
2588                 ((infomask & HEAP_XMAX_IS_MULTI) != 0 ? XLHL_XMAX_IS_MULTI : 0) |
2589                 ((infomask & HEAP_XMAX_LOCK_ONLY) != 0 ? XLHL_XMAX_LOCK_ONLY : 0) |
2590                 ((infomask & HEAP_XMAX_EXCL_LOCK) != 0 ? XLHL_XMAX_EXCL_LOCK : 0) |
2591         /* note we ignore HEAP_XMAX_SHR_LOCK here */
2592                 ((infomask & HEAP_XMAX_KEYSHR_LOCK) != 0 ? XLHL_XMAX_KEYSHR_LOCK : 0) |
2593                 ((infomask2 & HEAP_KEYS_UPDATED) != 0 ?
2594                  XLHL_KEYS_UPDATED : 0);
2595 }
2596
2597 /*
2598  * Given two versions of the same t_infomask for a tuple, compare them and
2599  * return whether the relevant status for a tuple Xmax has changed.  This is
2600  * used after a buffer lock has been released and reacquired: we want to ensure
2601  * that the tuple state continues to be the same it was when we previously
2602  * examined it.
2603  *
2604  * Note the Xmax field itself must be compared separately.
2605  */
2606 static inline bool
2607 xmax_infomask_changed(uint16 new_infomask, uint16 old_infomask)
2608 {
2609         const uint16 interesting =
2610         HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY | HEAP_LOCK_MASK;
2611
2612         if ((new_infomask & interesting) != (old_infomask & interesting))
2613                 return true;
2614
2615         return false;
2616 }
2617
2618 /*
2619  *      heap_delete - delete a tuple
2620  *
2621  * NB: do not call this directly unless you are prepared to deal with
2622  * concurrent-update conditions.  Use simple_heap_delete instead.
2623  *
2624  *      relation - table to be modified (caller must hold suitable lock)
2625  *      tid - TID of tuple to be deleted
2626  *      cid - delete command ID (used for visibility test, and stored into
2627  *              cmax if successful)
2628  *      crosscheck - if not InvalidSnapshot, also check tuple against this
2629  *      wait - true if should wait for any conflicting update to commit/abort
2630  *      hufd - output parameter, filled in failure cases (see below)
2631  *
2632  * Normal, successful return value is HeapTupleMayBeUpdated, which
2633  * actually means we did delete it.  Failure return codes are
2634  * HeapTupleSelfUpdated, HeapTupleUpdated, or HeapTupleBeingUpdated
2635  * (the last only possible if wait == false).
2636  *
2637  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
2638  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
2639  * (the last only for HeapTupleSelfUpdated, since we
2640  * cannot obtain cmax from a combocid generated by another transaction).
2641  * See comments for struct HeapUpdateFailureData for additional info.
2642  */
2643 HTSU_Result
2644 heap_delete(Relation relation, ItemPointer tid,
2645                         CommandId cid, Snapshot crosscheck, bool wait,
2646                         HeapUpdateFailureData *hufd)
2647 {
2648         HTSU_Result result;
2649         TransactionId xid = GetCurrentTransactionId();
2650         ItemId          lp;
2651         HeapTupleData tp;
2652         Page            page;
2653         BlockNumber block;
2654         Buffer          buffer;
2655         Buffer          vmbuffer = InvalidBuffer;
2656         TransactionId new_xmax;
2657         uint16          new_infomask,
2658                                 new_infomask2;
2659         bool            have_tuple_lock = false;
2660         bool            iscombo;
2661         bool            all_visible_cleared = false;
2662         HeapTuple       old_key_tuple = NULL;   /* replica identity of the tuple */
2663         bool            old_key_copied = false;
2664
2665         Assert(ItemPointerIsValid(tid));
2666
2667         block = ItemPointerGetBlockNumber(tid);
2668         buffer = ReadBuffer(relation, block);
2669         page = BufferGetPage(buffer);
2670
2671         /*
2672          * Before locking the buffer, pin the visibility map page if it appears to
2673          * be necessary.  Since we haven't got the lock yet, someone else might be
2674          * in the middle of changing this, so we'll need to recheck after we have
2675          * the lock.
2676          */
2677         if (PageIsAllVisible(page))
2678                 visibilitymap_pin(relation, block, &vmbuffer);
2679
2680         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2681
2682         /*
2683          * If we didn't pin the visibility map page and the page has become all
2684          * visible while we were busy locking the buffer, we'll have to unlock and
2685          * re-lock, to avoid holding the buffer lock across an I/O.  That's a bit
2686          * unfortunate, but hopefully shouldn't happen often.
2687          */
2688         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
2689         {
2690                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2691                 visibilitymap_pin(relation, block, &vmbuffer);
2692                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2693         }
2694
2695         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
2696         Assert(ItemIdIsNormal(lp));
2697
2698         tp.t_tableOid = RelationGetRelid(relation);
2699         tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
2700         tp.t_len = ItemIdGetLength(lp);
2701         tp.t_self = *tid;
2702
2703 l1:
2704         result = HeapTupleSatisfiesUpdate(&tp, cid, buffer);
2705
2706         if (result == HeapTupleInvisible)
2707         {
2708                 UnlockReleaseBuffer(buffer);
2709                 elog(ERROR, "attempted to delete invisible tuple");
2710         }
2711         else if (result == HeapTupleBeingUpdated && wait)
2712         {
2713                 TransactionId xwait;
2714                 uint16          infomask;
2715
2716                 /* must copy state data before unlocking buffer */
2717                 xwait = HeapTupleHeaderGetRawXmax(tp.t_data);
2718                 infomask = tp.t_data->t_infomask;
2719
2720                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2721
2722                 /*
2723                  * Acquire tuple lock to establish our priority for the tuple (see
2724                  * heap_lock_tuple).  LockTuple will release us when we are
2725                  * next-in-line for the tuple.
2726                  *
2727                  * If we are forced to "start over" below, we keep the tuple lock;
2728                  * this arranges that we stay at the head of the line while rechecking
2729                  * tuple state.
2730                  */
2731                 if (!have_tuple_lock)
2732                 {
2733                         LockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2734                         have_tuple_lock = true;
2735                 }
2736
2737                 /*
2738                  * Sleep until concurrent transaction ends.  Note that we don't care
2739                  * which lock mode the locker has, because we need the strongest one.
2740                  */
2741
2742                 if (infomask & HEAP_XMAX_IS_MULTI)
2743                 {
2744                         /* wait for multixact */
2745                         MultiXactIdWait((MultiXactId) xwait, MultiXactStatusUpdate, infomask,
2746                                                         relation, &tp.t_data->t_ctid, XLTW_Delete,
2747                                                         NULL);
2748                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2749
2750                         /*
2751                          * If xwait had just locked the tuple then some other xact could
2752                          * update this tuple before we get to this point.  Check for xmax
2753                          * change, and start over if so.
2754                          */
2755                         if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2756                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2757                                                                          xwait))
2758                                 goto l1;
2759
2760                         /*
2761                          * You might think the multixact is necessarily done here, but not
2762                          * so: it could have surviving members, namely our own xact or
2763                          * other subxacts of this backend.  It is legal for us to delete
2764                          * the tuple in either case, however (the latter case is
2765                          * essentially a situation of upgrading our former shared lock to
2766                          * exclusive).  We don't bother changing the on-disk hint bits
2767                          * since we are about to overwrite the xmax altogether.
2768                          */
2769                 }
2770                 else
2771                 {
2772                         /* wait for regular transaction to end */
2773                         XactLockTableWait(xwait, relation, &tp.t_data->t_ctid, XLTW_Delete);
2774                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2775
2776                         /*
2777                          * xwait is done, but if xwait had just locked the tuple then some
2778                          * other xact could update this tuple before we get to this point.
2779                          * Check for xmax change, and start over if so.
2780                          */
2781                         if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2782                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2783                                                                          xwait))
2784                                 goto l1;
2785
2786                         /* Otherwise check if it committed or aborted */
2787                         UpdateXmaxHintBits(tp.t_data, buffer, xwait);
2788                 }
2789
2790                 /*
2791                  * We may overwrite if previous xmax aborted, or if it committed but
2792                  * only locked the tuple without updating it.
2793                  */
2794                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
2795                         HEAP_XMAX_IS_LOCKED_ONLY(tp.t_data->t_infomask) ||
2796                         HeapTupleHeaderIsOnlyLocked(tp.t_data))
2797                         result = HeapTupleMayBeUpdated;
2798                 else
2799                         result = HeapTupleUpdated;
2800         }
2801
2802         if (crosscheck != InvalidSnapshot && result == HeapTupleMayBeUpdated)
2803         {
2804                 /* Perform additional check for transaction-snapshot mode RI updates */
2805                 if (!HeapTupleSatisfiesVisibility(&tp, crosscheck, buffer))
2806                         result = HeapTupleUpdated;
2807         }
2808
2809         if (result != HeapTupleMayBeUpdated)
2810         {
2811                 Assert(result == HeapTupleSelfUpdated ||
2812                            result == HeapTupleUpdated ||
2813                            result == HeapTupleBeingUpdated);
2814                 Assert(!(tp.t_data->t_infomask & HEAP_XMAX_INVALID));
2815                 hufd->ctid = tp.t_data->t_ctid;
2816                 hufd->xmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
2817                 if (result == HeapTupleSelfUpdated)
2818                         hufd->cmax = HeapTupleHeaderGetCmax(tp.t_data);
2819                 else
2820                         hufd->cmax = InvalidCommandId;
2821                 UnlockReleaseBuffer(buffer);
2822                 if (have_tuple_lock)
2823                         UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2824                 if (vmbuffer != InvalidBuffer)
2825                         ReleaseBuffer(vmbuffer);
2826                 return result;
2827         }
2828
2829         /*
2830          * We're about to do the actual delete -- check for conflict first, to
2831          * avoid possibly having to roll back work we've just done.
2832          */
2833         CheckForSerializableConflictIn(relation, &tp, buffer);
2834
2835         /* replace cid with a combo cid if necessary */
2836         HeapTupleHeaderAdjustCmax(tp.t_data, &cid, &iscombo);
2837
2838         /*
2839          * Compute replica identity tuple before entering the critical section so
2840          * we don't PANIC upon a memory allocation failure.
2841          */
2842         old_key_tuple = ExtractReplicaIdentity(relation, &tp, true, &old_key_copied);
2843
2844         /*
2845          * If this is the first possibly-multixact-able operation in the current
2846          * transaction, set my per-backend OldestMemberMXactId setting. We can be
2847          * certain that the transaction will never become a member of any older
2848          * MultiXactIds than that.  (We have to do this even if we end up just
2849          * using our own TransactionId below, since some other backend could
2850          * incorporate our XID into a MultiXact immediately afterwards.)
2851          */
2852         MultiXactIdSetOldestMember();
2853
2854         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(tp.t_data),
2855                                                           tp.t_data->t_infomask, tp.t_data->t_infomask2,
2856                                                           xid, LockTupleExclusive, true,
2857                                                           &new_xmax, &new_infomask, &new_infomask2);
2858
2859         START_CRIT_SECTION();
2860
2861         /*
2862          * If this transaction commits, the tuple will become DEAD sooner or
2863          * later.  Set flag that this page is a candidate for pruning once our xid
2864          * falls below the OldestXmin horizon.  If the transaction finally aborts,
2865          * the subsequent page pruning will be a no-op and the hint will be
2866          * cleared.
2867          */
2868         PageSetPrunable(page, xid);
2869
2870         if (PageIsAllVisible(page))
2871         {
2872                 all_visible_cleared = true;
2873                 PageClearAllVisible(page);
2874                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
2875                                                         vmbuffer);
2876         }
2877
2878         /* store transaction information of xact deleting the tuple */
2879         tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
2880         tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
2881         tp.t_data->t_infomask |= new_infomask;
2882         tp.t_data->t_infomask2 |= new_infomask2;
2883         HeapTupleHeaderClearHotUpdated(tp.t_data);
2884         HeapTupleHeaderSetXmax(tp.t_data, new_xmax);
2885         HeapTupleHeaderSetCmax(tp.t_data, cid, iscombo);
2886         /* Make sure there is no forward chain link in t_ctid */
2887         tp.t_data->t_ctid = tp.t_self;
2888
2889         MarkBufferDirty(buffer);
2890
2891         /* XLOG stuff */
2892         if (RelationNeedsWAL(relation))
2893         {
2894                 xl_heap_delete xlrec;
2895                 XLogRecPtr      recptr;
2896                 XLogRecData rdata[4];
2897
2898                 /* For logical decode we need combocids to properly decode the catalog */
2899                 if (RelationIsAccessibleInLogicalDecoding(relation))
2900                         log_heap_new_cid(relation, &tp);
2901
2902                 xlrec.flags = all_visible_cleared ? XLOG_HEAP_ALL_VISIBLE_CLEARED : 0;
2903                 xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
2904                                                                                           tp.t_data->t_infomask2);
2905                 xlrec.target.node = relation->rd_node;
2906                 xlrec.target.tid = tp.t_self;
2907                 xlrec.xmax = new_xmax;
2908                 rdata[0].data = (char *) &xlrec;
2909                 rdata[0].len = SizeOfHeapDelete;
2910                 rdata[0].buffer = InvalidBuffer;
2911                 rdata[0].next = &(rdata[1]);
2912
2913                 rdata[1].data = NULL;
2914                 rdata[1].len = 0;
2915                 rdata[1].buffer = buffer;
2916                 rdata[1].buffer_std = true;
2917                 rdata[1].next = NULL;
2918
2919                 /*
2920                  * Log replica identity of the deleted tuple if there is one
2921                  */
2922                 if (old_key_tuple != NULL)
2923                 {
2924                         xl_heap_header xlhdr;
2925
2926                         xlhdr.t_infomask2 = old_key_tuple->t_data->t_infomask2;
2927                         xlhdr.t_infomask = old_key_tuple->t_data->t_infomask;
2928                         xlhdr.t_hoff = old_key_tuple->t_data->t_hoff;
2929
2930                         rdata[1].next = &(rdata[2]);
2931                         rdata[2].data = (char *) &xlhdr;
2932                         rdata[2].len = SizeOfHeapHeader;
2933                         rdata[2].buffer = InvalidBuffer;
2934                         rdata[2].next = NULL;
2935
2936                         rdata[2].next = &(rdata[3]);
2937                         rdata[3].data = (char *) old_key_tuple->t_data
2938                                 + offsetof(HeapTupleHeaderData, t_bits);
2939                         rdata[3].len = old_key_tuple->t_len
2940                                 - offsetof(HeapTupleHeaderData, t_bits);
2941                         rdata[3].buffer = InvalidBuffer;
2942                         rdata[3].next = NULL;
2943
2944                         if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
2945                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_TUPLE;
2946                         else
2947                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_KEY;
2948                 }
2949
2950                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE, rdata);
2951
2952                 PageSetLSN(page, recptr);
2953         }
2954
2955         END_CRIT_SECTION();
2956
2957         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2958
2959         if (vmbuffer != InvalidBuffer)
2960                 ReleaseBuffer(vmbuffer);
2961
2962         /*
2963          * If the tuple has toasted out-of-line attributes, we need to delete
2964          * those items too.  We have to do this before releasing the buffer
2965          * because we need to look at the contents of the tuple, but it's OK to
2966          * release the content lock on the buffer first.
2967          */
2968         if (relation->rd_rel->relkind != RELKIND_RELATION &&
2969                 relation->rd_rel->relkind != RELKIND_MATVIEW)
2970         {
2971                 /* toast table entries should never be recursively toasted */
2972                 Assert(!HeapTupleHasExternal(&tp));
2973         }
2974         else if (HeapTupleHasExternal(&tp))
2975                 toast_delete(relation, &tp);
2976
2977         /*
2978          * Mark tuple for invalidation from system caches at next command
2979          * boundary. We have to do this before releasing the buffer because we
2980          * need to look at the contents of the tuple.
2981          */
2982         CacheInvalidateHeapTuple(relation, &tp, NULL);
2983
2984         /* Now we can release the buffer */
2985         ReleaseBuffer(buffer);
2986
2987         /*
2988          * Release the lmgr tuple lock, if we had it.
2989          */
2990         if (have_tuple_lock)
2991                 UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2992
2993         pgstat_count_heap_delete(relation);
2994
2995         if (old_key_tuple != NULL && old_key_copied)
2996                 heap_freetuple(old_key_tuple);
2997
2998         return HeapTupleMayBeUpdated;
2999 }
3000
3001 /*
3002  *      simple_heap_delete - delete a tuple
3003  *
3004  * This routine may be used to delete a tuple when concurrent updates of
3005  * the target tuple are not expected (for example, because we have a lock
3006  * on the relation associated with the tuple).  Any failure is reported
3007  * via ereport().
3008  */
3009 void
3010 simple_heap_delete(Relation relation, ItemPointer tid)
3011 {
3012         HTSU_Result result;
3013         HeapUpdateFailureData hufd;
3014
3015         result = heap_delete(relation, tid,
3016                                                  GetCurrentCommandId(true), InvalidSnapshot,
3017                                                  true /* wait for commit */ ,
3018                                                  &hufd);
3019         switch (result)
3020         {
3021                 case HeapTupleSelfUpdated:
3022                         /* Tuple was already updated in current command? */
3023                         elog(ERROR, "tuple already updated by self");
3024                         break;
3025
3026                 case HeapTupleMayBeUpdated:
3027                         /* done successfully */
3028                         break;
3029
3030                 case HeapTupleUpdated:
3031                         elog(ERROR, "tuple concurrently updated");
3032                         break;
3033
3034                 default:
3035                         elog(ERROR, "unrecognized heap_delete status: %u", result);
3036                         break;
3037         }
3038 }
3039
3040 /*
3041  *      heap_update - replace a tuple
3042  *
3043  * NB: do not call this directly unless you are prepared to deal with
3044  * concurrent-update conditions.  Use simple_heap_update instead.
3045  *
3046  *      relation - table to be modified (caller must hold suitable lock)
3047  *      otid - TID of old tuple to be replaced
3048  *      newtup - newly constructed tuple data to store
3049  *      cid - update command ID (used for visibility test, and stored into
3050  *              cmax/cmin if successful)
3051  *      crosscheck - if not InvalidSnapshot, also check old tuple against this
3052  *      wait - true if should wait for any conflicting update to commit/abort
3053  *      hufd - output parameter, filled in failure cases (see below)
3054  *      lockmode - output parameter, filled with lock mode acquired on tuple
3055  *
3056  * Normal, successful return value is HeapTupleMayBeUpdated, which
3057  * actually means we *did* update it.  Failure return codes are
3058  * HeapTupleSelfUpdated, HeapTupleUpdated, or HeapTupleBeingUpdated
3059  * (the last only possible if wait == false).
3060  *
3061  * On success, the header fields of *newtup are updated to match the new
3062  * stored tuple; in particular, newtup->t_self is set to the TID where the
3063  * new tuple was inserted, and its HEAP_ONLY_TUPLE flag is set iff a HOT
3064  * update was done.  However, any TOAST changes in the new tuple's
3065  * data are not reflected into *newtup.
3066  *
3067  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
3068  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
3069  * (the last only for HeapTupleSelfUpdated, since we
3070  * cannot obtain cmax from a combocid generated by another transaction).
3071  * See comments for struct HeapUpdateFailureData for additional info.
3072  */
3073 HTSU_Result
3074 heap_update(Relation relation, ItemPointer otid, HeapTuple newtup,
3075                         CommandId cid, Snapshot crosscheck, bool wait,
3076                         HeapUpdateFailureData *hufd, LockTupleMode *lockmode)
3077 {
3078         HTSU_Result result;
3079         TransactionId xid = GetCurrentTransactionId();
3080         Bitmapset  *hot_attrs;
3081         Bitmapset  *key_attrs;
3082         Bitmapset  *id_attrs;
3083         ItemId          lp;
3084         HeapTupleData oldtup;
3085         HeapTuple       heaptup;
3086         HeapTuple       old_key_tuple = NULL;
3087         bool            old_key_copied = false;
3088         Page            page;
3089         BlockNumber block;
3090         MultiXactStatus mxact_status;
3091         Buffer          buffer,
3092                                 newbuf,
3093                                 vmbuffer = InvalidBuffer,
3094                                 vmbuffer_new = InvalidBuffer;
3095         bool            need_toast,
3096                                 already_marked;
3097         Size            newtupsize,
3098                                 pagefree;
3099         bool            have_tuple_lock = false;
3100         bool            iscombo;
3101         bool            satisfies_hot;
3102         bool            satisfies_key;
3103         bool            satisfies_id;
3104         bool            use_hot_update = false;
3105         bool            key_intact;
3106         bool            all_visible_cleared = false;
3107         bool            all_visible_cleared_new = false;
3108         bool            checked_lockers;
3109         bool            locker_remains;
3110         TransactionId xmax_new_tuple,
3111                                 xmax_old_tuple;
3112         uint16          infomask_old_tuple,
3113                                 infomask2_old_tuple,
3114                                 infomask_new_tuple,
3115                                 infomask2_new_tuple;
3116
3117         Assert(ItemPointerIsValid(otid));
3118
3119         /*
3120          * Fetch the list of attributes to be checked for HOT update.  This is
3121          * wasted effort if we fail to update or have to put the new tuple on a
3122          * different page.  But we must compute the list before obtaining buffer
3123          * lock --- in the worst case, if we are doing an update on one of the
3124          * relevant system catalogs, we could deadlock if we try to fetch the list
3125          * later.  In any case, the relcache caches the data so this is usually
3126          * pretty cheap.
3127          *
3128          * Note that we get a copy here, so we need not worry about relcache flush
3129          * happening midway through.
3130          */
3131         hot_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_ALL);
3132         key_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_KEY);
3133         id_attrs = RelationGetIndexAttrBitmap(relation,
3134                                                                                   INDEX_ATTR_BITMAP_IDENTITY_KEY);
3135
3136         block = ItemPointerGetBlockNumber(otid);
3137         buffer = ReadBuffer(relation, block);
3138         page = BufferGetPage(buffer);
3139
3140         /*
3141          * Before locking the buffer, pin the visibility map page if it appears to
3142          * be necessary.  Since we haven't got the lock yet, someone else might be
3143          * in the middle of changing this, so we'll need to recheck after we have
3144          * the lock.
3145          */
3146         if (PageIsAllVisible(page))
3147                 visibilitymap_pin(relation, block, &vmbuffer);
3148
3149         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3150
3151         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(otid));
3152         Assert(ItemIdIsNormal(lp));
3153
3154         /*
3155          * Fill in enough data in oldtup for HeapSatisfiesHOTandKeyUpdate to work
3156          * properly.
3157          */
3158         oldtup.t_tableOid = RelationGetRelid(relation);
3159         oldtup.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3160         oldtup.t_len = ItemIdGetLength(lp);
3161         oldtup.t_self = *otid;
3162
3163         /* the new tuple is ready, except for this: */
3164         newtup->t_tableOid = RelationGetRelid(relation);
3165
3166         /* Fill in OID for newtup */
3167         if (relation->rd_rel->relhasoids)
3168         {
3169 #ifdef NOT_USED
3170                 /* this is redundant with an Assert in HeapTupleSetOid */
3171                 Assert(newtup->t_data->t_infomask & HEAP_HASOID);
3172 #endif
3173                 HeapTupleSetOid(newtup, HeapTupleGetOid(&oldtup));
3174         }
3175         else
3176         {
3177                 /* check there is not space for an OID */
3178                 Assert(!(newtup->t_data->t_infomask & HEAP_HASOID));
3179         }
3180
3181         /*
3182          * If we're not updating any "key" column, we can grab a weaker lock type.
3183          * This allows for more concurrency when we are running simultaneously
3184          * with foreign key checks.
3185          *
3186          * Note that if a column gets detoasted while executing the update, but
3187          * the value ends up being the same, this test will fail and we will use
3188          * the stronger lock.  This is acceptable; the important case to optimize
3189          * is updates that don't manipulate key columns, not those that
3190          * serendipitiously arrive at the same key values.
3191          */
3192         HeapSatisfiesHOTandKeyUpdate(relation, hot_attrs, key_attrs, id_attrs,
3193                                                                  &satisfies_hot, &satisfies_key,
3194                                                                  &satisfies_id, &oldtup, newtup);
3195         if (satisfies_key)
3196         {
3197                 *lockmode = LockTupleNoKeyExclusive;
3198                 mxact_status = MultiXactStatusNoKeyUpdate;
3199                 key_intact = true;
3200
3201                 /*
3202                  * If this is the first possibly-multixact-able operation in the
3203                  * current transaction, set my per-backend OldestMemberMXactId
3204                  * setting. We can be certain that the transaction will never become a
3205                  * member of any older MultiXactIds than that.  (We have to do this
3206                  * even if we end up just using our own TransactionId below, since
3207                  * some other backend could incorporate our XID into a MultiXact
3208                  * immediately afterwards.)
3209                  */
3210                 MultiXactIdSetOldestMember();
3211         }
3212         else
3213         {
3214                 *lockmode = LockTupleExclusive;
3215                 mxact_status = MultiXactStatusUpdate;
3216                 key_intact = false;
3217         }
3218
3219         /*
3220          * Note: beyond this point, use oldtup not otid to refer to old tuple.
3221          * otid may very well point at newtup->t_self, which we will overwrite
3222          * with the new tuple's location, so there's great risk of confusion if we
3223          * use otid anymore.
3224          */
3225
3226 l2:
3227         checked_lockers = false;
3228         locker_remains = false;
3229         result = HeapTupleSatisfiesUpdate(&oldtup, cid, buffer);
3230
3231         /* see below about the "no wait" case */
3232         Assert(result != HeapTupleBeingUpdated || wait);
3233
3234         if (result == HeapTupleInvisible)
3235         {
3236                 UnlockReleaseBuffer(buffer);
3237                 elog(ERROR, "attempted to update invisible tuple");
3238         }
3239         else if (result == HeapTupleBeingUpdated && wait)
3240         {
3241                 TransactionId xwait;
3242                 uint16          infomask;
3243                 bool            can_continue = false;
3244
3245                 checked_lockers = true;
3246
3247                 /*
3248                  * XXX note that we don't consider the "no wait" case here.  This
3249                  * isn't a problem currently because no caller uses that case, but it
3250                  * should be fixed if such a caller is introduced.  It wasn't a
3251                  * problem previously because this code would always wait, but now
3252                  * that some tuple locks do not conflict with one of the lock modes we
3253                  * use, it is possible that this case is interesting to handle
3254                  * specially.
3255                  *
3256                  * This may cause failures with third-party code that calls
3257                  * heap_update directly.
3258                  */
3259
3260                 /* must copy state data before unlocking buffer */
3261                 xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3262                 infomask = oldtup.t_data->t_infomask;
3263
3264                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3265
3266                 /*
3267                  * Acquire tuple lock to establish our priority for the tuple (see
3268                  * heap_lock_tuple).  LockTuple will release us when we are
3269                  * next-in-line for the tuple.
3270                  *
3271                  * If we are forced to "start over" below, we keep the tuple lock;
3272                  * this arranges that we stay at the head of the line while rechecking
3273                  * tuple state.
3274                  */
3275                 if (!have_tuple_lock)
3276                 {
3277                         LockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3278                         have_tuple_lock = true;
3279                 }
3280
3281                 /*
3282                  * Now we have to do something about the existing locker.  If it's a
3283                  * multi, sleep on it; we might be awakened before it is completely
3284                  * gone (or even not sleep at all in some cases); we need to preserve
3285                  * it as locker, unless it is gone completely.
3286                  *
3287                  * If it's not a multi, we need to check for sleeping conditions
3288                  * before actually going to sleep.  If the update doesn't conflict
3289                  * with the locks, we just continue without sleeping (but making sure
3290                  * it is preserved).
3291                  */
3292                 if (infomask & HEAP_XMAX_IS_MULTI)
3293                 {
3294                         TransactionId update_xact;
3295                         int                     remain;
3296
3297                         /* wait for multixact */
3298                         MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
3299                                                         relation, &oldtup.t_data->t_ctid, XLTW_Update,
3300                                                         &remain);
3301                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3302
3303                         /*
3304                          * If xwait had just locked the tuple then some other xact could
3305                          * update this tuple before we get to this point.  Check for xmax
3306                          * change, and start over if so.
3307                          */
3308                         if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3309                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3310                                                                          xwait))
3311                                 goto l2;
3312
3313                         /*
3314                          * Note that the multixact may not be done by now.  It could have
3315                          * surviving members; our own xact or other subxacts of this
3316                          * backend, and also any other concurrent transaction that locked
3317                          * the tuple with KeyShare if we only got TupleLockUpdate.  If
3318                          * this is the case, we have to be careful to mark the updated
3319                          * tuple with the surviving members in Xmax.
3320                          *
3321                          * Note that there could have been another update in the
3322                          * MultiXact. In that case, we need to check whether it committed
3323                          * or aborted. If it aborted we are safe to update it again;
3324                          * otherwise there is an update conflict, and we have to return
3325                          * HeapTupleUpdated below.
3326                          *
3327                          * In the LockTupleExclusive case, we still need to preserve the
3328                          * surviving members: those would include the tuple locks we had
3329                          * before this one, which are important to keep in case this
3330                          * subxact aborts.
3331                          */
3332                         update_xact = InvalidTransactionId;
3333                         if (!HEAP_XMAX_IS_LOCKED_ONLY(oldtup.t_data->t_infomask))
3334                                 update_xact = HeapTupleGetUpdateXid(oldtup.t_data);
3335
3336                         /*
3337                          * There was no UPDATE in the MultiXact; or it aborted. No
3338                          * TransactionIdIsInProgress() call needed here, since we called
3339                          * MultiXactIdWait() above.
3340                          */
3341                         if (!TransactionIdIsValid(update_xact) ||
3342                                 TransactionIdDidAbort(update_xact))
3343                                 can_continue = true;
3344
3345                         locker_remains = remain != 0;
3346                 }
3347                 else
3348                 {
3349                         /*
3350                          * If it's just a key-share locker, and we're not changing the key
3351                          * columns, we don't need to wait for it to end; but we need to
3352                          * preserve it as locker.
3353                          */
3354                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) && key_intact)
3355                         {
3356                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3357
3358                                 /*
3359                                  * recheck the locker; if someone else changed the tuple while
3360                                  * we weren't looking, start over.
3361                                  */
3362                                 if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3363                                         !TransactionIdEquals(
3364                                                                         HeapTupleHeaderGetRawXmax(oldtup.t_data),
3365                                                                                  xwait))
3366                                         goto l2;
3367
3368                                 can_continue = true;
3369                                 locker_remains = true;
3370                         }
3371                         else
3372                         {
3373                                 /* wait for regular transaction to end */
3374                                 XactLockTableWait(xwait, relation, &oldtup.t_data->t_ctid,
3375                                                                   XLTW_Update);
3376                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3377
3378                                 /*
3379                                  * xwait is done, but if xwait had just locked the tuple then
3380                                  * some other xact could update this tuple before we get to
3381                                  * this point. Check for xmax change, and start over if so.
3382                                  */
3383                                 if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3384                                         !TransactionIdEquals(
3385                                                                         HeapTupleHeaderGetRawXmax(oldtup.t_data),
3386                                                                                  xwait))
3387                                         goto l2;
3388
3389                                 /* Otherwise check if it committed or aborted */
3390                                 UpdateXmaxHintBits(oldtup.t_data, buffer, xwait);
3391                                 if (oldtup.t_data->t_infomask & HEAP_XMAX_INVALID)
3392                                         can_continue = true;
3393                         }
3394                 }
3395
3396                 result = can_continue ? HeapTupleMayBeUpdated : HeapTupleUpdated;
3397         }
3398
3399         if (crosscheck != InvalidSnapshot && result == HeapTupleMayBeUpdated)
3400         {
3401                 /* Perform additional check for transaction-snapshot mode RI updates */
3402                 if (!HeapTupleSatisfiesVisibility(&oldtup, crosscheck, buffer))
3403                         result = HeapTupleUpdated;
3404         }
3405
3406         if (result != HeapTupleMayBeUpdated)
3407         {
3408                 Assert(result == HeapTupleSelfUpdated ||
3409                            result == HeapTupleUpdated ||
3410                            result == HeapTupleBeingUpdated);
3411                 Assert(!(oldtup.t_data->t_infomask & HEAP_XMAX_INVALID));
3412                 hufd->ctid = oldtup.t_data->t_ctid;
3413                 hufd->xmax = HeapTupleHeaderGetUpdateXid(oldtup.t_data);
3414                 if (result == HeapTupleSelfUpdated)
3415                         hufd->cmax = HeapTupleHeaderGetCmax(oldtup.t_data);
3416                 else
3417                         hufd->cmax = InvalidCommandId;
3418                 UnlockReleaseBuffer(buffer);
3419                 if (have_tuple_lock)
3420                         UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3421                 if (vmbuffer != InvalidBuffer)
3422                         ReleaseBuffer(vmbuffer);
3423                 bms_free(hot_attrs);
3424                 bms_free(key_attrs);
3425                 return result;
3426         }
3427
3428         /*
3429          * If we didn't pin the visibility map page and the page has become all
3430          * visible while we were busy locking the buffer, or during some
3431          * subsequent window during which we had it unlocked, we'll have to unlock
3432          * and re-lock, to avoid holding the buffer lock across an I/O.  That's a
3433          * bit unfortunate, especially since we'll now have to recheck whether the
3434          * tuple has been locked or updated under us, but hopefully it won't
3435          * happen very often.
3436          */
3437         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3438         {
3439                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3440                 visibilitymap_pin(relation, block, &vmbuffer);
3441                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3442                 goto l2;
3443         }
3444
3445         /*
3446          * We're about to do the actual update -- check for conflict first, to
3447          * avoid possibly having to roll back work we've just done.
3448          */
3449         CheckForSerializableConflictIn(relation, &oldtup, buffer);
3450
3451         /* Fill in transaction status data */
3452
3453         /*
3454          * If the tuple we're updating is locked, we need to preserve the locking
3455          * info in the old tuple's Xmax.  Prepare a new Xmax value for this.
3456          */
3457         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3458                                                           oldtup.t_data->t_infomask,
3459                                                           oldtup.t_data->t_infomask2,
3460                                                           xid, *lockmode, true,
3461                                                           &xmax_old_tuple, &infomask_old_tuple,
3462                                                           &infomask2_old_tuple);
3463
3464         /*
3465          * And also prepare an Xmax value for the new copy of the tuple.  If there
3466          * was no xmax previously, or there was one but all lockers are now gone,
3467          * then use InvalidXid; otherwise, get the xmax from the old tuple.  (In
3468          * rare cases that might also be InvalidXid and yet not have the
3469          * HEAP_XMAX_INVALID bit set; that's fine.)
3470          */
3471         if ((oldtup.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3472                 (checked_lockers && !locker_remains))
3473                 xmax_new_tuple = InvalidTransactionId;
3474         else
3475                 xmax_new_tuple = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3476
3477         if (!TransactionIdIsValid(xmax_new_tuple))
3478         {
3479                 infomask_new_tuple = HEAP_XMAX_INVALID;
3480                 infomask2_new_tuple = 0;
3481         }
3482         else
3483         {
3484                 /*
3485                  * If we found a valid Xmax for the new tuple, then the infomask bits
3486                  * to use on the new tuple depend on what was there on the old one.
3487                  * Note that since we're doing an update, the only possibility is that
3488                  * the lockers had FOR KEY SHARE lock.
3489                  */
3490                 if (oldtup.t_data->t_infomask & HEAP_XMAX_IS_MULTI)
3491                 {
3492                         GetMultiXactIdHintBits(xmax_new_tuple, &infomask_new_tuple,
3493                                                                    &infomask2_new_tuple);
3494                 }
3495                 else
3496                 {
3497                         infomask_new_tuple = HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_LOCK_ONLY;
3498                         infomask2_new_tuple = 0;
3499                 }
3500         }
3501
3502         /*
3503          * Prepare the new tuple with the appropriate initial values of Xmin and
3504          * Xmax, as well as initial infomask bits as computed above.
3505          */
3506         newtup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
3507         newtup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
3508         HeapTupleHeaderSetXmin(newtup->t_data, xid);
3509         HeapTupleHeaderSetCmin(newtup->t_data, cid);
3510         newtup->t_data->t_infomask |= HEAP_UPDATED | infomask_new_tuple;
3511         newtup->t_data->t_infomask2 |= infomask2_new_tuple;
3512         HeapTupleHeaderSetXmax(newtup->t_data, xmax_new_tuple);
3513
3514         /*
3515          * Replace cid with a combo cid if necessary.  Note that we already put
3516          * the plain cid into the new tuple.
3517          */
3518         HeapTupleHeaderAdjustCmax(oldtup.t_data, &cid, &iscombo);
3519
3520         /*
3521          * If the toaster needs to be activated, OR if the new tuple will not fit
3522          * on the same page as the old, then we need to release the content lock
3523          * (but not the pin!) on the old tuple's buffer while we are off doing
3524          * TOAST and/or table-file-extension work.  We must mark the old tuple to
3525          * show that it's already being updated, else other processes may try to
3526          * update it themselves.
3527          *
3528          * We need to invoke the toaster if there are already any out-of-line
3529          * toasted values present, or if the new tuple is over-threshold.
3530          */
3531         if (relation->rd_rel->relkind != RELKIND_RELATION &&
3532                 relation->rd_rel->relkind != RELKIND_MATVIEW)
3533         {
3534                 /* toast table entries should never be recursively toasted */
3535                 Assert(!HeapTupleHasExternal(&oldtup));
3536                 Assert(!HeapTupleHasExternal(newtup));
3537                 need_toast = false;
3538         }
3539         else
3540                 need_toast = (HeapTupleHasExternal(&oldtup) ||
3541                                           HeapTupleHasExternal(newtup) ||
3542                                           newtup->t_len > TOAST_TUPLE_THRESHOLD);
3543
3544         pagefree = PageGetHeapFreeSpace(page);
3545
3546         newtupsize = MAXALIGN(newtup->t_len);
3547
3548         if (need_toast || newtupsize > pagefree)
3549         {
3550                 /* Clear obsolete visibility flags ... */
3551                 oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3552                 oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3553                 HeapTupleClearHotUpdated(&oldtup);
3554                 /* ... and store info about transaction updating this tuple */
3555                 Assert(TransactionIdIsValid(xmax_old_tuple));
3556                 HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
3557                 oldtup.t_data->t_infomask |= infomask_old_tuple;
3558                 oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
3559                 HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3560                 /* temporarily make it look not-updated */
3561                 oldtup.t_data->t_ctid = oldtup.t_self;
3562                 already_marked = true;
3563                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3564
3565                 /*
3566                  * Let the toaster do its thing, if needed.
3567                  *
3568                  * Note: below this point, heaptup is the data we actually intend to
3569                  * store into the relation; newtup is the caller's original untoasted
3570                  * data.
3571                  */
3572                 if (need_toast)
3573                 {
3574                         /* Note we always use WAL and FSM during updates */
3575                         heaptup = toast_insert_or_update(relation, newtup, &oldtup, 0);
3576                         newtupsize = MAXALIGN(heaptup->t_len);
3577                 }
3578                 else
3579                         heaptup = newtup;
3580
3581                 /*
3582                  * Now, do we need a new page for the tuple, or not?  This is a bit
3583                  * tricky since someone else could have added tuples to the page while
3584                  * we weren't looking.  We have to recheck the available space after
3585                  * reacquiring the buffer lock.  But don't bother to do that if the
3586                  * former amount of free space is still not enough; it's unlikely
3587                  * there's more free now than before.
3588                  *
3589                  * What's more, if we need to get a new page, we will need to acquire
3590                  * buffer locks on both old and new pages.  To avoid deadlock against
3591                  * some other backend trying to get the same two locks in the other
3592                  * order, we must be consistent about the order we get the locks in.
3593                  * We use the rule "lock the lower-numbered page of the relation
3594                  * first".  To implement this, we must do RelationGetBufferForTuple
3595                  * while not holding the lock on the old page, and we must rely on it
3596                  * to get the locks on both pages in the correct order.
3597                  */
3598                 if (newtupsize > pagefree)
3599                 {
3600                         /* Assume there's no chance to put heaptup on same page. */
3601                         newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3602                                                                                            buffer, 0, NULL,
3603                                                                                            &vmbuffer_new, &vmbuffer);
3604                 }
3605                 else
3606                 {
3607                         /* Re-acquire the lock on the old tuple's page. */
3608                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3609                         /* Re-check using the up-to-date free space */
3610                         pagefree = PageGetHeapFreeSpace(page);
3611                         if (newtupsize > pagefree)
3612                         {
3613                                 /*
3614                                  * Rats, it doesn't fit anymore.  We must now unlock and
3615                                  * relock to avoid deadlock.  Fortunately, this path should
3616                                  * seldom be taken.
3617                                  */
3618                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3619                                 newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3620                                                                                                    buffer, 0, NULL,
3621                                                                                                    &vmbuffer_new, &vmbuffer);
3622                         }
3623                         else
3624                         {
3625                                 /* OK, it fits here, so we're done. */
3626                                 newbuf = buffer;
3627                         }
3628                 }
3629         }
3630         else
3631         {
3632                 /* No TOAST work needed, and it'll fit on same page */
3633                 already_marked = false;
3634                 newbuf = buffer;
3635                 heaptup = newtup;
3636         }
3637
3638         /*
3639          * We're about to create the new tuple -- check for conflict first, to
3640          * avoid possibly having to roll back work we've just done.
3641          *
3642          * NOTE: For a tuple insert, we only need to check for table locks, since
3643          * predicate locking at the index level will cover ranges for anything
3644          * except a table scan.  Therefore, only provide the relation.
3645          */
3646         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
3647
3648         /*
3649          * At this point newbuf and buffer are both pinned and locked, and newbuf
3650          * has enough space for the new tuple.  If they are the same buffer, only
3651          * one pin is held.
3652          */
3653
3654         if (newbuf == buffer)
3655         {
3656                 /*
3657                  * Since the new tuple is going into the same page, we might be able
3658                  * to do a HOT update.  Check if any of the index columns have been
3659                  * changed.  If not, then HOT update is possible.
3660                  */
3661                 if (satisfies_hot)
3662                         use_hot_update = true;
3663         }
3664         else
3665         {
3666                 /* Set a hint that the old page could use prune/defrag */
3667                 PageSetFull(page);
3668         }
3669
3670         /*
3671          * Compute replica identity tuple before entering the critical section so
3672          * we don't PANIC upon a memory allocation failure.
3673          * ExtractReplicaIdentity() will return NULL if nothing needs to be
3674          * logged.
3675          */
3676         old_key_tuple = ExtractReplicaIdentity(relation, &oldtup, !satisfies_id, &old_key_copied);
3677
3678         /* NO EREPORT(ERROR) from here till changes are logged */
3679         START_CRIT_SECTION();
3680
3681         /*
3682          * If this transaction commits, the old tuple will become DEAD sooner or
3683          * later.  Set flag that this page is a candidate for pruning once our xid
3684          * falls below the OldestXmin horizon.  If the transaction finally aborts,
3685          * the subsequent page pruning will be a no-op and the hint will be
3686          * cleared.
3687          *
3688          * XXX Should we set hint on newbuf as well?  If the transaction aborts,
3689          * there would be a prunable tuple in the newbuf; but for now we choose
3690          * not to optimize for aborts.  Note that heap_xlog_update must be kept in
3691          * sync if this decision changes.
3692          */
3693         PageSetPrunable(page, xid);
3694
3695         if (use_hot_update)
3696         {
3697                 /* Mark the old tuple as HOT-updated */
3698                 HeapTupleSetHotUpdated(&oldtup);
3699                 /* And mark the new tuple as heap-only */
3700                 HeapTupleSetHeapOnly(heaptup);
3701                 /* Mark the caller's copy too, in case different from heaptup */
3702                 HeapTupleSetHeapOnly(newtup);
3703         }
3704         else
3705         {
3706                 /* Make sure tuples are correctly marked as not-HOT */
3707                 HeapTupleClearHotUpdated(&oldtup);
3708                 HeapTupleClearHeapOnly(heaptup);
3709                 HeapTupleClearHeapOnly(newtup);
3710         }
3711
3712         RelationPutHeapTuple(relation, newbuf, heaptup);        /* insert new tuple */
3713
3714         if (!already_marked)
3715         {
3716                 /* Clear obsolete visibility flags ... */
3717                 oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3718                 oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3719                 /* ... and store info about transaction updating this tuple */
3720                 Assert(TransactionIdIsValid(xmax_old_tuple));
3721                 HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
3722                 oldtup.t_data->t_infomask |= infomask_old_tuple;
3723                 oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
3724                 HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3725         }
3726
3727         /* record address of new tuple in t_ctid of old one */
3728         oldtup.t_data->t_ctid = heaptup->t_self;
3729
3730         /* clear PD_ALL_VISIBLE flags */
3731         if (PageIsAllVisible(BufferGetPage(buffer)))
3732         {
3733                 all_visible_cleared = true;
3734                 PageClearAllVisible(BufferGetPage(buffer));
3735                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
3736                                                         vmbuffer);
3737         }
3738         if (newbuf != buffer && PageIsAllVisible(BufferGetPage(newbuf)))
3739         {
3740                 all_visible_cleared_new = true;
3741                 PageClearAllVisible(BufferGetPage(newbuf));
3742                 visibilitymap_clear(relation, BufferGetBlockNumber(newbuf),
3743                                                         vmbuffer_new);
3744         }
3745
3746         if (newbuf != buffer)
3747                 MarkBufferDirty(newbuf);
3748         MarkBufferDirty(buffer);
3749
3750         /* XLOG stuff */
3751         if (RelationNeedsWAL(relation))
3752         {
3753                 XLogRecPtr      recptr;
3754
3755                 /*
3756                  * For logical decoding we need combocids to properly decode the
3757                  * catalog.
3758                  */
3759                 if (RelationIsAccessibleInLogicalDecoding(relation))
3760                 {
3761                         log_heap_new_cid(relation, &oldtup);
3762                         log_heap_new_cid(relation, heaptup);
3763                 }
3764
3765                 recptr = log_heap_update(relation, buffer,
3766                                                                  newbuf, &oldtup, heaptup,
3767                                                                  old_key_tuple,
3768                                                                  all_visible_cleared,
3769                                                                  all_visible_cleared_new);
3770                 if (newbuf != buffer)
3771                 {
3772                         PageSetLSN(BufferGetPage(newbuf), recptr);
3773                 }
3774                 PageSetLSN(BufferGetPage(buffer), recptr);
3775         }
3776
3777         END_CRIT_SECTION();
3778
3779         if (newbuf != buffer)
3780                 LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
3781         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3782
3783         /*
3784          * Mark old tuple for invalidation from system caches at next command
3785          * boundary, and mark the new tuple for invalidation in case we abort. We
3786          * have to do this before releasing the buffer because oldtup is in the
3787          * buffer.  (heaptup is all in local memory, but it's necessary to process
3788          * both tuple versions in one call to inval.c so we can avoid redundant
3789          * sinval messages.)
3790          */
3791         CacheInvalidateHeapTuple(relation, &oldtup, heaptup);
3792
3793         /* Now we can release the buffer(s) */
3794         if (newbuf != buffer)
3795                 ReleaseBuffer(newbuf);
3796         ReleaseBuffer(buffer);
3797         if (BufferIsValid(vmbuffer_new))
3798                 ReleaseBuffer(vmbuffer_new);
3799         if (BufferIsValid(vmbuffer))
3800                 ReleaseBuffer(vmbuffer);
3801
3802         /*
3803          * Release the lmgr tuple lock, if we had it.
3804          */
3805         if (have_tuple_lock)
3806                 UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3807
3808         pgstat_count_heap_update(relation, use_hot_update);
3809
3810         /*
3811          * If heaptup is a private copy, release it.  Don't forget to copy t_self
3812          * back to the caller's image, too.
3813          */
3814         if (heaptup != newtup)
3815         {
3816                 newtup->t_self = heaptup->t_self;
3817                 heap_freetuple(heaptup);
3818         }
3819
3820         if (old_key_tuple != NULL && old_key_copied)
3821                 heap_freetuple(old_key_tuple);
3822
3823         bms_free(hot_attrs);
3824         bms_free(key_attrs);
3825
3826         return HeapTupleMayBeUpdated;
3827 }
3828
3829 /*
3830  * Check if the specified attribute's value is same in both given tuples.
3831  * Subroutine for HeapSatisfiesHOTandKeyUpdate.
3832  */
3833 static bool
3834 heap_tuple_attr_equals(TupleDesc tupdesc, int attrnum,
3835                                            HeapTuple tup1, HeapTuple tup2)
3836 {
3837         Datum           value1,
3838                                 value2;
3839         bool            isnull1,
3840                                 isnull2;
3841         Form_pg_attribute att;
3842
3843         /*
3844          * If it's a whole-tuple reference, say "not equal".  It's not really
3845          * worth supporting this case, since it could only succeed after a no-op
3846          * update, which is hardly a case worth optimizing for.
3847          */
3848         if (attrnum == 0)
3849                 return false;
3850
3851         /*
3852          * Likewise, automatically say "not equal" for any system attribute other
3853          * than OID and tableOID; we cannot expect these to be consistent in a HOT
3854          * chain, or even to be set correctly yet in the new tuple.
3855          */
3856         if (attrnum < 0)
3857         {
3858                 if (attrnum != ObjectIdAttributeNumber &&
3859                         attrnum != TableOidAttributeNumber)
3860                         return false;
3861         }
3862
3863         /*
3864          * Extract the corresponding values.  XXX this is pretty inefficient if
3865          * there are many indexed columns.  Should HeapSatisfiesHOTandKeyUpdate do
3866          * a single heap_deform_tuple call on each tuple, instead?      But that
3867          * doesn't work for system columns ...
3868          */
3869         value1 = heap_getattr(tup1, attrnum, tupdesc, &isnull1);
3870         value2 = heap_getattr(tup2, attrnum, tupdesc, &isnull2);
3871
3872         /*
3873          * If one value is NULL and other is not, then they are certainly not
3874          * equal
3875          */
3876         if (isnull1 != isnull2)
3877                 return false;
3878
3879         /*
3880          * If both are NULL, they can be considered equal.
3881          */
3882         if (isnull1)
3883                 return true;
3884
3885         /*
3886          * We do simple binary comparison of the two datums.  This may be overly
3887          * strict because there can be multiple binary representations for the
3888          * same logical value.  But we should be OK as long as there are no false
3889          * positives.  Using a type-specific equality operator is messy because
3890          * there could be multiple notions of equality in different operator
3891          * classes; furthermore, we cannot safely invoke user-defined functions
3892          * while holding exclusive buffer lock.
3893          */
3894         if (attrnum <= 0)
3895         {
3896                 /* The only allowed system columns are OIDs, so do this */
3897                 return (DatumGetObjectId(value1) == DatumGetObjectId(value2));
3898         }
3899         else
3900         {
3901                 Assert(attrnum <= tupdesc->natts);
3902                 att = tupdesc->attrs[attrnum - 1];
3903                 return datumIsEqual(value1, value2, att->attbyval, att->attlen);
3904         }
3905 }
3906
3907 /*
3908  * Check which columns are being updated.
3909  *
3910  * This simultaneously checks conditions for HOT updates, for FOR KEY
3911  * SHARE updates, and REPLICA IDENTITY concerns.  Since much of the time they
3912  * will be checking very similar sets of columns, and doing the same tests on
3913  * them, it makes sense to optimize and do them together.
3914  *
3915  * We receive three bitmapsets comprising the three sets of columns we're
3916  * interested in.  Note these are destructively modified; that is OK since
3917  * this is invoked at most once in heap_update.
3918  *
3919  * hot_result is set to TRUE if it's okay to do a HOT update (i.e. it does not
3920  * modified indexed columns); key_result is set to TRUE if the update does not
3921  * modify columns used in the key; id_result is set to TRUE if the update does
3922  * not modify columns in any index marked as the REPLICA IDENTITY.
3923  */
3924 static void
3925 HeapSatisfiesHOTandKeyUpdate(Relation relation, Bitmapset *hot_attrs,
3926                                                          Bitmapset *key_attrs, Bitmapset *id_attrs,
3927                                                          bool *satisfies_hot, bool *satisfies_key,
3928                                                          bool *satisfies_id,
3929                                                          HeapTuple oldtup, HeapTuple newtup)
3930 {
3931         int                     next_hot_attnum;
3932         int                     next_key_attnum;
3933         int                     next_id_attnum;
3934         bool            hot_result = true;
3935         bool            key_result = true;
3936         bool            id_result = true;
3937
3938         /* If REPLICA IDENTITY is set to FULL, id_attrs will be empty. */
3939         Assert(bms_is_subset(id_attrs, key_attrs));
3940         Assert(bms_is_subset(key_attrs, hot_attrs));
3941
3942         /*
3943          * If one of these sets contains no remaining bits, bms_first_member will
3944          * return -1, and after adding FirstLowInvalidHeapAttributeNumber (which
3945          * is negative!)  we'll get an attribute number that can't possibly be
3946          * real, and thus won't match any actual attribute number.
3947          */
3948         next_hot_attnum = bms_first_member(hot_attrs);
3949         next_hot_attnum += FirstLowInvalidHeapAttributeNumber;
3950         next_key_attnum = bms_first_member(key_attrs);
3951         next_key_attnum += FirstLowInvalidHeapAttributeNumber;
3952         next_id_attnum = bms_first_member(id_attrs);
3953         next_id_attnum += FirstLowInvalidHeapAttributeNumber;
3954
3955         for (;;)
3956         {
3957                 bool            changed;
3958                 int                     check_now;
3959
3960                 /*
3961                  * Since the HOT attributes are a superset of the key attributes and
3962                  * the key attributes are a superset of the id attributes, this logic
3963                  * is guaranteed to identify the next column that needs to be checked.
3964                  */
3965                 if (hot_result && next_hot_attnum > FirstLowInvalidHeapAttributeNumber)
3966                         check_now = next_hot_attnum;
3967                 else if (key_result && next_key_attnum > FirstLowInvalidHeapAttributeNumber)
3968                         check_now = next_key_attnum;
3969                 else if (id_result && next_id_attnum > FirstLowInvalidHeapAttributeNumber)
3970                         check_now = next_id_attnum;
3971                 else
3972                         break;
3973
3974                 /* See whether it changed. */
3975                 changed = !heap_tuple_attr_equals(RelationGetDescr(relation),
3976                                                                                   check_now, oldtup, newtup);
3977                 if (changed)
3978                 {
3979                         if (check_now == next_hot_attnum)
3980                                 hot_result = false;
3981                         if (check_now == next_key_attnum)
3982                                 key_result = false;
3983                         if (check_now == next_id_attnum)
3984                                 id_result = false;
3985
3986                         /* if all are false now, we can stop checking */
3987                         if (!hot_result && !key_result && !id_result)
3988                                 break;
3989                 }
3990
3991                 /*
3992                  * Advance the next attribute numbers for the sets that contain the
3993                  * attribute we just checked.  As we work our way through the columns,
3994                  * the next_attnum values will rise; but when each set becomes empty,
3995                  * bms_first_member() will return -1 and the attribute number will end
3996                  * up with a value less than FirstLowInvalidHeapAttributeNumber.
3997                  */
3998                 if (hot_result && check_now == next_hot_attnum)
3999                 {
4000                         next_hot_attnum = bms_first_member(hot_attrs);
4001                         next_hot_attnum += FirstLowInvalidHeapAttributeNumber;
4002                 }
4003                 if (key_result && check_now == next_key_attnum)
4004                 {
4005                         next_key_attnum = bms_first_member(key_attrs);
4006                         next_key_attnum += FirstLowInvalidHeapAttributeNumber;
4007                 }
4008                 if (id_result && check_now == next_id_attnum)
4009                 {
4010                         next_id_attnum = bms_first_member(id_attrs);
4011                         next_id_attnum += FirstLowInvalidHeapAttributeNumber;
4012                 }
4013         }
4014
4015         *satisfies_hot = hot_result;
4016         *satisfies_key = key_result;
4017         *satisfies_id = id_result;
4018 }
4019
4020 /*
4021  *      simple_heap_update - replace a tuple
4022  *
4023  * This routine may be used to update a tuple when concurrent updates of
4024  * the target tuple are not expected (for example, because we have a lock
4025  * on the relation associated with the tuple).  Any failure is reported
4026  * via ereport().
4027  */
4028 void
4029 simple_heap_update(Relation relation, ItemPointer otid, HeapTuple tup)
4030 {
4031         HTSU_Result result;
4032         HeapUpdateFailureData hufd;
4033         LockTupleMode lockmode;
4034
4035         result = heap_update(relation, otid, tup,
4036                                                  GetCurrentCommandId(true), InvalidSnapshot,
4037                                                  true /* wait for commit */ ,
4038                                                  &hufd, &lockmode);
4039         switch (result)
4040         {
4041                 case HeapTupleSelfUpdated:
4042                         /* Tuple was already updated in current command? */
4043                         elog(ERROR, "tuple already updated by self");
4044                         break;
4045
4046                 case HeapTupleMayBeUpdated:
4047                         /* done successfully */
4048                         break;
4049
4050                 case HeapTupleUpdated:
4051                         elog(ERROR, "tuple concurrently updated");
4052                         break;
4053
4054                 default:
4055                         elog(ERROR, "unrecognized heap_update status: %u", result);
4056                         break;
4057         }
4058 }
4059
4060
4061 /*
4062  * Return the MultiXactStatus corresponding to the given tuple lock mode.
4063  */
4064 static MultiXactStatus
4065 get_mxact_status_for_lock(LockTupleMode mode, bool is_update)
4066 {
4067         int                     retval;
4068
4069         if (is_update)
4070                 retval = tupleLockExtraInfo[mode].updstatus;
4071         else
4072                 retval = tupleLockExtraInfo[mode].lockstatus;
4073
4074         if (retval == -1)
4075                 elog(ERROR, "invalid lock tuple mode %d/%s", mode,
4076                          is_update ? "true" : "false");
4077
4078         return (MultiXactStatus) retval;
4079 }
4080
4081
4082 /*
4083  *      heap_lock_tuple - lock a tuple in shared or exclusive mode
4084  *
4085  * Note that this acquires a buffer pin, which the caller must release.
4086  *
4087  * Input parameters:
4088  *      relation: relation containing tuple (caller must hold suitable lock)
4089  *      tuple->t_self: TID of tuple to lock (rest of struct need not be valid)
4090  *      cid: current command ID (used for visibility test, and stored into
4091  *              tuple's cmax if lock is successful)
4092  *      mode: indicates if shared or exclusive tuple lock is desired
4093  *      nowait: if true, ereport rather than blocking if lock not available
4094  *      follow_updates: if true, follow the update chain to also lock descendant
4095  *              tuples.
4096  *
4097  * Output parameters:
4098  *      *tuple: all fields filled in
4099  *      *buffer: set to buffer holding tuple (pinned but not locked at exit)
4100  *      *hufd: filled in failure cases (see below)
4101  *
4102  * Function result may be:
4103  *      HeapTupleMayBeUpdated: lock was successfully acquired
4104  *      HeapTupleSelfUpdated: lock failed because tuple updated by self
4105  *      HeapTupleUpdated: lock failed because tuple updated by other xact
4106  *
4107  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
4108  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
4109  * (the last only for HeapTupleSelfUpdated, since we
4110  * cannot obtain cmax from a combocid generated by another transaction).
4111  * See comments for struct HeapUpdateFailureData for additional info.
4112  *
4113  * See README.tuplock for a thorough explanation of this mechanism.
4114  */
4115 HTSU_Result
4116 heap_lock_tuple(Relation relation, HeapTuple tuple,
4117                                 CommandId cid, LockTupleMode mode, bool nowait,
4118                                 bool follow_updates,
4119                                 Buffer *buffer, HeapUpdateFailureData *hufd)
4120 {
4121         HTSU_Result result;
4122         ItemPointer tid = &(tuple->t_self);
4123         ItemId          lp;
4124         Page            page;
4125         TransactionId xid,
4126                                 xmax;
4127         uint16          old_infomask,
4128                                 new_infomask,
4129                                 new_infomask2;
4130         bool            have_tuple_lock = false;
4131
4132         *buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
4133         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4134
4135         page = BufferGetPage(*buffer);
4136         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
4137         Assert(ItemIdIsNormal(lp));
4138
4139         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
4140         tuple->t_len = ItemIdGetLength(lp);
4141         tuple->t_tableOid = RelationGetRelid(relation);
4142
4143 l3:
4144         result = HeapTupleSatisfiesUpdate(tuple, cid, *buffer);
4145
4146         if (result == HeapTupleInvisible)
4147         {
4148                 UnlockReleaseBuffer(*buffer);
4149                 elog(ERROR, "attempted to lock invisible tuple");
4150         }
4151         else if (result == HeapTupleBeingUpdated)
4152         {
4153                 TransactionId xwait;
4154                 uint16          infomask;
4155                 uint16          infomask2;
4156                 bool            require_sleep;
4157                 ItemPointerData t_ctid;
4158
4159                 /* must copy state data before unlocking buffer */
4160                 xwait = HeapTupleHeaderGetRawXmax(tuple->t_data);
4161                 infomask = tuple->t_data->t_infomask;
4162                 infomask2 = tuple->t_data->t_infomask2;
4163                 ItemPointerCopy(&tuple->t_data->t_ctid, &t_ctid);
4164
4165                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4166
4167                 /*
4168                  * If any subtransaction of the current top transaction already holds
4169                  * a lock as strong or stronger than what we're requesting, we
4170                  * effectively hold the desired lock already.  We *must* succeed
4171                  * without trying to take the tuple lock, else we will deadlock
4172                  * against anyone wanting to acquire a stronger lock.
4173                  */
4174                 if (infomask & HEAP_XMAX_IS_MULTI)
4175                 {
4176                         int                     i;
4177                         int                     nmembers;
4178                         MultiXactMember *members;
4179
4180                         /*
4181                          * We don't need to allow old multixacts here; if that had been
4182                          * the case, HeapTupleSatisfiesUpdate would have returned
4183                          * MayBeUpdated and we wouldn't be here.
4184                          */
4185                         nmembers =
4186                                 GetMultiXactIdMembers(xwait, &members, false,
4187                                                                           HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4188
4189                         for (i = 0; i < nmembers; i++)
4190                         {
4191                                 if (TransactionIdIsCurrentTransactionId(members[i].xid))
4192                                 {
4193                                         LockTupleMode membermode;
4194
4195                                         membermode = TUPLOCK_from_mxstatus(members[i].status);
4196
4197                                         if (membermode >= mode)
4198                                         {
4199                                                 if (have_tuple_lock)
4200                                                         UnlockTupleTuplock(relation, tid, mode);
4201
4202                                                 pfree(members);
4203                                                 return HeapTupleMayBeUpdated;
4204                                         }
4205                                 }
4206                         }
4207
4208                         if (members)
4209                                 pfree(members);
4210                 }
4211
4212                 /*
4213                  * Acquire tuple lock to establish our priority for the tuple.
4214                  * LockTuple will release us when we are next-in-line for the tuple.
4215                  * We must do this even if we are share-locking.
4216                  *
4217                  * If we are forced to "start over" below, we keep the tuple lock;
4218                  * this arranges that we stay at the head of the line while rechecking
4219                  * tuple state.
4220                  */
4221                 if (!have_tuple_lock)
4222                 {
4223                         if (nowait)
4224                         {
4225                                 if (!ConditionalLockTupleTuplock(relation, tid, mode))
4226                                         ereport(ERROR,
4227                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4228                                         errmsg("could not obtain lock on row in relation \"%s\"",
4229                                                    RelationGetRelationName(relation))));
4230                         }
4231                         else
4232                                 LockTupleTuplock(relation, tid, mode);
4233                         have_tuple_lock = true;
4234                 }
4235
4236                 /*
4237                  * Initially assume that we will have to wait for the locking
4238                  * transaction(s) to finish.  We check various cases below in which
4239                  * this can be turned off.
4240                  */
4241                 require_sleep = true;
4242                 if (mode == LockTupleKeyShare)
4243                 {
4244                         /*
4245                          * If we're requesting KeyShare, and there's no update present, we
4246                          * don't need to wait.  Even if there is an update, we can still
4247                          * continue if the key hasn't been modified.
4248                          *
4249                          * However, if there are updates, we need to walk the update chain
4250                          * to mark future versions of the row as locked, too.  That way,
4251                          * if somebody deletes that future version, we're protected
4252                          * against the key going away.  This locking of future versions
4253                          * could block momentarily, if a concurrent transaction is
4254                          * deleting a key; or it could return a value to the effect that
4255                          * the transaction deleting the key has already committed.  So we
4256                          * do this before re-locking the buffer; otherwise this would be
4257                          * prone to deadlocks.
4258                          *
4259                          * Note that the TID we're locking was grabbed before we unlocked
4260                          * the buffer.  For it to change while we're not looking, the
4261                          * other properties we're testing for below after re-locking the
4262                          * buffer would also change, in which case we would restart this
4263                          * loop above.
4264                          */
4265                         if (!(infomask2 & HEAP_KEYS_UPDATED))
4266                         {
4267                                 bool            updated;
4268
4269                                 updated = !HEAP_XMAX_IS_LOCKED_ONLY(infomask);
4270
4271                                 /*
4272                                  * If there are updates, follow the update chain; bail out if
4273                                  * that cannot be done.
4274                                  */
4275                                 if (follow_updates && updated)
4276                                 {
4277                                         HTSU_Result res;
4278
4279                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4280                                                                                                   GetCurrentTransactionId(),
4281                                                                                                   mode);
4282                                         if (res != HeapTupleMayBeUpdated)
4283                                         {
4284                                                 result = res;
4285                                                 /* recovery code expects to have buffer lock held */
4286                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4287                                                 goto failed;
4288                                         }
4289                                 }
4290
4291                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4292
4293                                 /*
4294                                  * Make sure it's still an appropriate lock, else start over.
4295                                  * Also, if it wasn't updated before we released the lock, but
4296                                  * is updated now, we start over too; the reason is that we
4297                                  * now need to follow the update chain to lock the new
4298                                  * versions.
4299                                  */
4300                                 if (!HeapTupleHeaderIsOnlyLocked(tuple->t_data) &&
4301                                         ((tuple->t_data->t_infomask2 & HEAP_KEYS_UPDATED) ||
4302                                          !updated))
4303                                         goto l3;
4304
4305                                 /* Things look okay, so we can skip sleeping */
4306                                 require_sleep = false;
4307
4308                                 /*
4309                                  * Note we allow Xmax to change here; other updaters/lockers
4310                                  * could have modified it before we grabbed the buffer lock.
4311                                  * However, this is not a problem, because with the recheck we
4312                                  * just did we ensure that they still don't conflict with the
4313                                  * lock we want.
4314                                  */
4315                         }
4316                 }
4317                 else if (mode == LockTupleShare)
4318                 {
4319                         /*
4320                          * If we're requesting Share, we can similarly avoid sleeping if
4321                          * there's no update and no exclusive lock present.
4322                          */
4323                         if (HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
4324                                 !HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4325                         {
4326                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4327
4328                                 /*
4329                                  * Make sure it's still an appropriate lock, else start over.
4330                                  * See above about allowing xmax to change.
4331                                  */
4332                                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4333                                         HEAP_XMAX_IS_EXCL_LOCKED(tuple->t_data->t_infomask))
4334                                         goto l3;
4335                                 require_sleep = false;
4336                         }
4337                 }
4338                 else if (mode == LockTupleNoKeyExclusive)
4339                 {
4340                         /*
4341                          * If we're requesting NoKeyExclusive, we might also be able to
4342                          * avoid sleeping; just ensure that there's no other lock type
4343                          * than KeyShare.  Note that this is a bit more involved than just
4344                          * checking hint bits -- we need to expand the multixact to figure
4345                          * out lock modes for each one (unless there was only one such
4346                          * locker).
4347                          */
4348                         if (infomask & HEAP_XMAX_IS_MULTI)
4349                         {
4350                                 int                     nmembers;
4351                                 MultiXactMember *members;
4352
4353                                 /*
4354                                  * We don't need to allow old multixacts here; if that had
4355                                  * been the case, HeapTupleSatisfiesUpdate would have returned
4356                                  * MayBeUpdated and we wouldn't be here.
4357                                  */
4358                                 nmembers =
4359                                         GetMultiXactIdMembers(xwait, &members, false,
4360                                                                                   HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4361
4362                                 if (nmembers <= 0)
4363                                 {
4364                                         /*
4365                                          * No need to keep the previous xmax here. This is
4366                                          * unlikely to happen.
4367                                          */
4368                                         require_sleep = false;
4369                                 }
4370                                 else
4371                                 {
4372                                         int                     i;
4373                                         bool            allowed = true;
4374
4375                                         for (i = 0; i < nmembers; i++)
4376                                         {
4377                                                 if (members[i].status != MultiXactStatusForKeyShare)
4378                                                 {
4379                                                         allowed = false;
4380                                                         break;
4381                                                 }
4382                                         }
4383                                         if (allowed)
4384                                         {
4385                                                 /*
4386                                                  * if the xmax changed under us in the meantime, start
4387                                                  * over.
4388                                                  */
4389                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4390                                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4391                                                         !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4392                                                                                                  xwait))
4393                                                 {
4394                                                         pfree(members);
4395                                                         goto l3;
4396                                                 }
4397                                                 /* otherwise, we're good */
4398                                                 require_sleep = false;
4399                                         }
4400
4401                                         pfree(members);
4402                                 }
4403                         }
4404                         else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
4405                         {
4406                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4407
4408                                 /* if the xmax changed in the meantime, start over */
4409                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4410                                         !TransactionIdEquals(
4411                                                                         HeapTupleHeaderGetRawXmax(tuple->t_data),
4412                                                                                  xwait))
4413                                         goto l3;
4414                                 /* otherwise, we're good */
4415                                 require_sleep = false;
4416                         }
4417                 }
4418
4419                 /*
4420                  * By here, we either have already acquired the buffer exclusive lock,
4421                  * or we must wait for the locking transaction or multixact; so below
4422                  * we ensure that we grab buffer lock after the sleep.
4423                  */
4424
4425                 if (require_sleep)
4426                 {
4427                         if (infomask & HEAP_XMAX_IS_MULTI)
4428                         {
4429                                 MultiXactStatus status = get_mxact_status_for_lock(mode, false);
4430
4431                                 /* We only ever lock tuples, never update them */
4432                                 if (status >= MultiXactStatusNoKeyUpdate)
4433                                         elog(ERROR, "invalid lock mode in heap_lock_tuple");
4434
4435                                 /* wait for multixact to end */
4436                                 if (nowait)
4437                                 {
4438                                         if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4439                                                                                                   status, infomask, relation,
4440                                                                                                         NULL))
4441                                                 ereport(ERROR,
4442                                                                 (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4443                                                                  errmsg("could not obtain lock on row in relation \"%s\"",
4444                                                                                 RelationGetRelationName(relation))));
4445                                 }
4446                                 else
4447                                         MultiXactIdWait((MultiXactId) xwait, status, infomask,
4448                                                                         relation, &tuple->t_data->t_ctid,
4449                                                                         XLTW_Lock, NULL);
4450
4451                                 /* if there are updates, follow the update chain */
4452                                 if (follow_updates &&
4453                                         !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
4454                                 {
4455                                         HTSU_Result res;
4456
4457                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4458                                                                                                   GetCurrentTransactionId(),
4459                                                                                                   mode);
4460                                         if (res != HeapTupleMayBeUpdated)
4461                                         {
4462                                                 result = res;
4463                                                 /* recovery code expects to have buffer lock held */
4464                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4465                                                 goto failed;
4466                                         }
4467                                 }
4468
4469                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4470
4471                                 /*
4472                                  * If xwait had just locked the tuple then some other xact
4473                                  * could update this tuple before we get to this point. Check
4474                                  * for xmax change, and start over if so.
4475                                  */
4476                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4477                                         !TransactionIdEquals(
4478                                                                         HeapTupleHeaderGetRawXmax(tuple->t_data),
4479                                                                                  xwait))
4480                                         goto l3;
4481
4482                                 /*
4483                                  * Of course, the multixact might not be done here: if we're
4484                                  * requesting a light lock mode, other transactions with light
4485                                  * locks could still be alive, as well as locks owned by our
4486                                  * own xact or other subxacts of this backend.  We need to
4487                                  * preserve the surviving MultiXact members.  Note that it
4488                                  * isn't absolutely necessary in the latter case, but doing so
4489                                  * is simpler.
4490                                  */
4491                         }
4492                         else
4493                         {
4494                                 /* wait for regular transaction to end */
4495                                 if (nowait)
4496                                 {
4497                                         if (!ConditionalXactLockTableWait(xwait))
4498                                                 ereport(ERROR,
4499                                                                 (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4500                                                                  errmsg("could not obtain lock on row in relation \"%s\"",
4501                                                                                 RelationGetRelationName(relation))));
4502                                 }
4503                                 else
4504                                         XactLockTableWait(xwait, relation, &tuple->t_data->t_ctid,
4505                                                                           XLTW_Lock);
4506
4507                                 /* if there are updates, follow the update chain */
4508                                 if (follow_updates &&
4509                                         !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
4510                                 {
4511                                         HTSU_Result res;
4512
4513                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4514                                                                                                   GetCurrentTransactionId(),
4515                                                                                                   mode);
4516                                         if (res != HeapTupleMayBeUpdated)
4517                                         {
4518                                                 result = res;
4519                                                 /* recovery code expects to have buffer lock held */
4520                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4521                                                 goto failed;
4522                                         }
4523                                 }
4524
4525                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4526
4527                                 /*
4528                                  * xwait is done, but if xwait had just locked the tuple then
4529                                  * some other xact could update this tuple before we get to
4530                                  * this point.  Check for xmax change, and start over if so.
4531                                  */
4532                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4533                                         !TransactionIdEquals(
4534                                                                         HeapTupleHeaderGetRawXmax(tuple->t_data),
4535                                                                                  xwait))
4536                                         goto l3;
4537
4538                                 /*
4539                                  * Otherwise check if it committed or aborted.  Note we cannot
4540                                  * be here if the tuple was only locked by somebody who didn't
4541                                  * conflict with us; that should have been handled above.  So
4542                                  * that transaction must necessarily be gone by now.
4543                                  */
4544                                 UpdateXmaxHintBits(tuple->t_data, *buffer, xwait);
4545                         }
4546                 }
4547
4548                 /* By here, we're certain that we hold buffer exclusive lock again */
4549
4550                 /*
4551                  * We may lock if previous xmax aborted, or if it committed but only
4552                  * locked the tuple without updating it; or if we didn't have to wait
4553                  * at all for whatever reason.
4554                  */
4555                 if (!require_sleep ||
4556                         (tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
4557                         HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4558                         HeapTupleHeaderIsOnlyLocked(tuple->t_data))
4559                         result = HeapTupleMayBeUpdated;
4560                 else
4561                         result = HeapTupleUpdated;
4562         }
4563
4564 failed:
4565         if (result != HeapTupleMayBeUpdated)
4566         {
4567                 Assert(result == HeapTupleSelfUpdated || result == HeapTupleUpdated);
4568                 Assert(!(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));
4569                 hufd->ctid = tuple->t_data->t_ctid;
4570                 hufd->xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
4571                 if (result == HeapTupleSelfUpdated)
4572                         hufd->cmax = HeapTupleHeaderGetCmax(tuple->t_data);
4573                 else
4574                         hufd->cmax = InvalidCommandId;
4575                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4576                 if (have_tuple_lock)
4577                         UnlockTupleTuplock(relation, tid, mode);
4578                 return result;
4579         }
4580
4581         xmax = HeapTupleHeaderGetRawXmax(tuple->t_data);
4582         old_infomask = tuple->t_data->t_infomask;
4583
4584         /*
4585          * We might already hold the desired lock (or stronger), possibly under a
4586          * different subtransaction of the current top transaction.  If so, there
4587          * is no need to change state or issue a WAL record.  We already handled
4588          * the case where this is true for xmax being a MultiXactId, so now check
4589          * for cases where it is a plain TransactionId.
4590          *
4591          * Note in particular that this covers the case where we already hold
4592          * exclusive lock on the tuple and the caller only wants key share or
4593          * share lock. It would certainly not do to give up the exclusive lock.
4594          */
4595         if (!(old_infomask & (HEAP_XMAX_INVALID |
4596                                                   HEAP_XMAX_COMMITTED |
4597                                                   HEAP_XMAX_IS_MULTI)) &&
4598                 (mode == LockTupleKeyShare ?
4599                  (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask) ||
4600                   HEAP_XMAX_IS_SHR_LOCKED(old_infomask) ||
4601                   HEAP_XMAX_IS_EXCL_LOCKED(old_infomask)) :
4602                  mode == LockTupleShare ?
4603                  (HEAP_XMAX_IS_SHR_LOCKED(old_infomask) ||
4604                   HEAP_XMAX_IS_EXCL_LOCKED(old_infomask)) :
4605                  (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))) &&
4606                 TransactionIdIsCurrentTransactionId(xmax))
4607         {
4608                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4609                 /* Probably can't hold tuple lock here, but may as well check */
4610                 if (have_tuple_lock)
4611                         UnlockTupleTuplock(relation, tid, mode);
4612                 return HeapTupleMayBeUpdated;
4613         }
4614
4615         /*
4616          * If this is the first possibly-multixact-able operation in the current
4617          * transaction, set my per-backend OldestMemberMXactId setting. We can be
4618          * certain that the transaction will never become a member of any older
4619          * MultiXactIds than that.  (We have to do this even if we end up just
4620          * using our own TransactionId below, since some other backend could
4621          * incorporate our XID into a MultiXact immediately afterwards.)
4622          */
4623         MultiXactIdSetOldestMember();
4624
4625         /*
4626          * Compute the new xmax and infomask to store into the tuple.  Note we do
4627          * not modify the tuple just yet, because that would leave it in the wrong
4628          * state if multixact.c elogs.
4629          */
4630         compute_new_xmax_infomask(xmax, old_infomask, tuple->t_data->t_infomask2,
4631                                                           GetCurrentTransactionId(), mode, false,
4632                                                           &xid, &new_infomask, &new_infomask2);
4633
4634         START_CRIT_SECTION();
4635
4636         /*
4637          * Store transaction information of xact locking the tuple.
4638          *
4639          * Note: Cmax is meaningless in this context, so don't set it; this avoids
4640          * possibly generating a useless combo CID.  Moreover, if we're locking a
4641          * previously updated tuple, it's important to preserve the Cmax.
4642          *
4643          * Also reset the HOT UPDATE bit, but only if there's no update; otherwise
4644          * we would break the HOT chain.
4645          */
4646         tuple->t_data->t_infomask &= ~HEAP_XMAX_BITS;
4647         tuple->t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
4648         tuple->t_data->t_infomask |= new_infomask;
4649         tuple->t_data->t_infomask2 |= new_infomask2;
4650         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
4651                 HeapTupleHeaderClearHotUpdated(tuple->t_data);
4652         HeapTupleHeaderSetXmax(tuple->t_data, xid);
4653
4654         /*
4655          * Make sure there is no forward chain link in t_ctid.  Note that in the
4656          * cases where the tuple has been updated, we must not overwrite t_ctid,
4657          * because it was set by the updater.  Moreover, if the tuple has been
4658          * updated, we need to follow the update chain to lock the new versions of
4659          * the tuple as well.
4660          */
4661         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
4662                 tuple->t_data->t_ctid = *tid;
4663
4664         MarkBufferDirty(*buffer);
4665
4666         /*
4667          * XLOG stuff.  You might think that we don't need an XLOG record because
4668          * there is no state change worth restoring after a crash.  You would be
4669          * wrong however: we have just written either a TransactionId or a
4670          * MultiXactId that may never have been seen on disk before, and we need
4671          * to make sure that there are XLOG entries covering those ID numbers.
4672          * Else the same IDs might be re-used after a crash, which would be
4673          * disastrous if this page made it to disk before the crash.  Essentially
4674          * we have to enforce the WAL log-before-data rule even in this case.
4675          * (Also, in a PITR log-shipping or 2PC environment, we have to have XLOG
4676          * entries for everything anyway.)
4677          */
4678         if (RelationNeedsWAL(relation))
4679         {
4680                 xl_heap_lock xlrec;
4681                 XLogRecPtr      recptr;
4682                 XLogRecData rdata[2];
4683
4684                 xlrec.target.node = relation->rd_node;
4685                 xlrec.target.tid = tuple->t_self;
4686                 xlrec.locking_xid = xid;
4687                 xlrec.infobits_set = compute_infobits(new_infomask,
4688                                                                                           tuple->t_data->t_infomask2);
4689                 rdata[0].data = (char *) &xlrec;
4690                 rdata[0].len = SizeOfHeapLock;
4691                 rdata[0].buffer = InvalidBuffer;
4692                 rdata[0].next = &(rdata[1]);
4693
4694                 rdata[1].data = NULL;
4695                 rdata[1].len = 0;
4696                 rdata[1].buffer = *buffer;
4697                 rdata[1].buffer_std = true;
4698                 rdata[1].next = NULL;
4699
4700                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK, rdata);
4701
4702                 PageSetLSN(page, recptr);
4703         }
4704
4705         END_CRIT_SECTION();
4706
4707         LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4708
4709         /*
4710          * Don't update the visibility map here. Locking a tuple doesn't change
4711          * visibility info.
4712          */
4713
4714         /*
4715          * Now that we have successfully marked the tuple as locked, we can
4716          * release the lmgr tuple lock, if we had it.
4717          */
4718         if (have_tuple_lock)
4719                 UnlockTupleTuplock(relation, tid, mode);
4720
4721         return HeapTupleMayBeUpdated;
4722 }
4723
4724
4725 /*
4726  * Given an original set of Xmax and infomask, and a transaction (identified by
4727  * add_to_xmax) acquiring a new lock of some mode, compute the new Xmax and
4728  * corresponding infomasks to use on the tuple.
4729  *
4730  * Note that this might have side effects such as creating a new MultiXactId.
4731  *
4732  * Most callers will have called HeapTupleSatisfiesUpdate before this function;
4733  * that will have set the HEAP_XMAX_INVALID bit if the xmax was a MultiXactId
4734  * but it was not running anymore. There is a race condition, which is that the
4735  * MultiXactId may have finished since then, but that uncommon case is handled
4736  * either here, or within MultiXactIdExpand.
4737  *
4738  * There is a similar race condition possible when the old xmax was a regular
4739  * TransactionId.  We test TransactionIdIsInProgress again just to narrow the
4740  * window, but it's still possible to end up creating an unnecessary
4741  * MultiXactId.  Fortunately this is harmless.
4742  */
4743 static void
4744 compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
4745                                                   uint16 old_infomask2, TransactionId add_to_xmax,
4746                                                   LockTupleMode mode, bool is_update,
4747                                                   TransactionId *result_xmax, uint16 *result_infomask,
4748                                                   uint16 *result_infomask2)
4749 {
4750         TransactionId new_xmax;
4751         uint16          new_infomask,
4752                                 new_infomask2;
4753
4754         Assert(TransactionIdIsCurrentTransactionId(add_to_xmax));
4755
4756 l5:
4757         new_infomask = 0;
4758         new_infomask2 = 0;
4759         if (old_infomask & HEAP_XMAX_INVALID)
4760         {
4761                 /*
4762                  * No previous locker; we just insert our own TransactionId.
4763                  *
4764                  * Note that it's critical that this case be the first one checked,
4765                  * because there are several blocks below that come back to this one
4766                  * to implement certain optimizations; old_infomask might contain
4767                  * other dirty bits in those cases, but we don't really care.
4768                  */
4769                 if (is_update)
4770                 {
4771                         new_xmax = add_to_xmax;
4772                         if (mode == LockTupleExclusive)
4773                                 new_infomask2 |= HEAP_KEYS_UPDATED;
4774                 }
4775                 else
4776                 {
4777                         new_infomask |= HEAP_XMAX_LOCK_ONLY;
4778                         switch (mode)
4779                         {
4780                                 case LockTupleKeyShare:
4781                                         new_xmax = add_to_xmax;
4782                                         new_infomask |= HEAP_XMAX_KEYSHR_LOCK;
4783                                         break;
4784                                 case LockTupleShare:
4785                                         new_xmax = add_to_xmax;
4786                                         new_infomask |= HEAP_XMAX_SHR_LOCK;
4787                                         break;
4788                                 case LockTupleNoKeyExclusive:
4789                                         new_xmax = add_to_xmax;
4790                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
4791                                         break;
4792                                 case LockTupleExclusive:
4793                                         new_xmax = add_to_xmax;
4794                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
4795                                         new_infomask2 |= HEAP_KEYS_UPDATED;
4796                                         break;
4797                                 default:
4798                                         new_xmax = InvalidTransactionId;        /* silence compiler */
4799                                         elog(ERROR, "invalid lock mode");
4800                         }
4801                 }
4802         }
4803         else if (old_infomask & HEAP_XMAX_IS_MULTI)
4804         {
4805                 MultiXactStatus new_status;
4806
4807                 /*
4808                  * Currently we don't allow XMAX_COMMITTED to be set for multis, so
4809                  * cross-check.
4810                  */
4811                 Assert(!(old_infomask & HEAP_XMAX_COMMITTED));
4812
4813                 /*
4814                  * A multixact together with LOCK_ONLY set but neither lock bit set
4815                  * (i.e. a pg_upgraded share locked tuple) cannot possibly be running
4816                  * anymore.  This check is critical for databases upgraded by
4817                  * pg_upgrade; both MultiXactIdIsRunning and MultiXactIdExpand assume
4818                  * that such multis are never passed.
4819                  */
4820                 if (!(old_infomask & HEAP_LOCK_MASK) &&
4821                         HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
4822                 {
4823                         old_infomask &= ~HEAP_XMAX_IS_MULTI;
4824                         old_infomask |= HEAP_XMAX_INVALID;
4825                         goto l5;
4826                 }
4827
4828                 /*
4829                  * If the XMAX is already a MultiXactId, then we need to expand it to
4830                  * include add_to_xmax; but if all the members were lockers and are
4831                  * all gone, we can do away with the IS_MULTI bit and just set
4832                  * add_to_xmax as the only locker/updater.  If all lockers are gone
4833                  * and we have an updater that aborted, we can also do without a
4834                  * multi.
4835                  *
4836                  * The cost of doing GetMultiXactIdMembers would be paid by
4837                  * MultiXactIdExpand if we weren't to do this, so this check is not
4838                  * incurring extra work anyhow.
4839                  */
4840                 if (!MultiXactIdIsRunning(xmax, HEAP_XMAX_IS_LOCKED_ONLY(old_infomask)))
4841                 {
4842                         if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) ||
4843                                 TransactionIdDidAbort(MultiXactIdGetUpdateXid(xmax,
4844                                                                                                                           old_infomask)))
4845                         {
4846                                 /*
4847                                  * Reset these bits and restart; otherwise fall through to
4848                                  * create a new multi below.
4849                                  */
4850                                 old_infomask &= ~HEAP_XMAX_IS_MULTI;
4851                                 old_infomask |= HEAP_XMAX_INVALID;
4852                                 goto l5;
4853                         }
4854                 }
4855
4856                 new_status = get_mxact_status_for_lock(mode, is_update);
4857
4858                 new_xmax = MultiXactIdExpand((MultiXactId) xmax, add_to_xmax,
4859                                                                          new_status);
4860                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4861         }
4862         else if (old_infomask & HEAP_XMAX_COMMITTED)
4863         {
4864                 /*
4865                  * It's a committed update, so we need to preserve him as updater of
4866                  * the tuple.
4867                  */
4868                 MultiXactStatus status;
4869                 MultiXactStatus new_status;
4870
4871                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4872                         status = MultiXactStatusUpdate;
4873                 else
4874                         status = MultiXactStatusNoKeyUpdate;
4875
4876                 new_status = get_mxact_status_for_lock(mode, is_update);
4877
4878                 /*
4879                  * since it's not running, it's obviously impossible for the old
4880                  * updater to be identical to the current one, so we need not check
4881                  * for that case as we do in the block above.
4882                  */
4883                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
4884                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4885         }
4886         else if (TransactionIdIsInProgress(xmax))
4887         {
4888                 /*
4889                  * If the XMAX is a valid, in-progress TransactionId, then we need to
4890                  * create a new MultiXactId that includes both the old locker or
4891                  * updater and our own TransactionId.
4892                  */
4893                 MultiXactStatus new_status;
4894                 MultiXactStatus old_status;
4895                 LockTupleMode old_mode;
4896
4897                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
4898                 {
4899                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
4900                                 old_status = MultiXactStatusForKeyShare;
4901                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
4902                                 old_status = MultiXactStatusForShare;
4903                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
4904                         {
4905                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4906                                         old_status = MultiXactStatusForUpdate;
4907                                 else
4908                                         old_status = MultiXactStatusForNoKeyUpdate;
4909                         }
4910                         else
4911                         {
4912                                 /*
4913                                  * LOCK_ONLY can be present alone only when a page has been
4914                                  * upgraded by pg_upgrade.  But in that case,
4915                                  * TransactionIdIsInProgress() should have returned false.  We
4916                                  * assume it's no longer locked in this case.
4917                                  */
4918                                 elog(WARNING, "LOCK_ONLY found for Xid in progress %u", xmax);
4919                                 old_infomask |= HEAP_XMAX_INVALID;
4920                                 old_infomask &= ~HEAP_XMAX_LOCK_ONLY;
4921                                 goto l5;
4922                         }
4923                 }
4924                 else
4925                 {
4926                         /* it's an update, but which kind? */
4927                         if (old_infomask2 & HEAP_KEYS_UPDATED)
4928                                 old_status = MultiXactStatusUpdate;
4929                         else
4930                                 old_status = MultiXactStatusNoKeyUpdate;
4931                 }
4932
4933                 old_mode = TUPLOCK_from_mxstatus(old_status);
4934
4935                 /*
4936                  * If the lock to be acquired is for the same TransactionId as the
4937                  * existing lock, there's an optimization possible: consider only the
4938                  * strongest of both locks as the only one present, and restart.
4939                  */
4940                 if (xmax == add_to_xmax)
4941                 {
4942                         /*
4943                          * Note that it's not possible for the original tuple to be
4944                          * updated: we wouldn't be here because the tuple would have been
4945                          * invisible and we wouldn't try to update it.  As a subtlety,
4946                          * this code can also run when traversing an update chain to lock
4947                          * future versions of a tuple.  But we wouldn't be here either,
4948                          * because the add_to_xmax would be different from the original
4949                          * updater.
4950                          */
4951                         Assert(HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
4952
4953                         /* acquire the strongest of both */
4954                         if (mode < old_mode)
4955                                 mode = old_mode;
4956                         /* mustn't touch is_update */
4957
4958                         old_infomask |= HEAP_XMAX_INVALID;
4959                         goto l5;
4960                 }
4961
4962                 /* otherwise, just fall back to creating a new multixact */
4963                 new_status = get_mxact_status_for_lock(mode, is_update);
4964                 new_xmax = MultiXactIdCreate(xmax, old_status,
4965                                                                          add_to_xmax, new_status);
4966                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4967         }
4968         else if (!HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) &&
4969                          TransactionIdDidCommit(xmax))
4970         {
4971                 /*
4972                  * It's a committed update, so we gotta preserve him as updater of the
4973                  * tuple.
4974                  */
4975                 MultiXactStatus status;
4976                 MultiXactStatus new_status;
4977
4978                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4979                         status = MultiXactStatusUpdate;
4980                 else
4981                         status = MultiXactStatusNoKeyUpdate;
4982
4983                 new_status = get_mxact_status_for_lock(mode, is_update);
4984
4985                 /*
4986                  * since it's not running, it's obviously impossible for the old
4987                  * updater to be identical to the current one, so we need not check
4988                  * for that case as we do in the block above.
4989                  */
4990                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
4991                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4992         }
4993         else
4994         {
4995                 /*
4996                  * Can get here iff the locking/updating transaction was running when
4997                  * the infomask was extracted from the tuple, but finished before
4998                  * TransactionIdIsInProgress got to run.  Deal with it as if there was
4999                  * no locker at all in the first place.
5000                  */
5001                 old_infomask |= HEAP_XMAX_INVALID;
5002                 goto l5;
5003         }
5004
5005         *result_infomask = new_infomask;
5006         *result_infomask2 = new_infomask2;
5007         *result_xmax = new_xmax;
5008 }
5009
5010 /*
5011  * Subroutine for heap_lock_updated_tuple_rec.
5012  *
5013  * Given an hypothetical multixact status held by the transaction identified
5014  * with the given xid, does the current transaction need to wait, fail, or can
5015  * it continue if it wanted to acquire a lock of the given mode?  "needwait"
5016  * is set to true if waiting is necessary; if it can continue, then
5017  * HeapTupleMayBeUpdated is returned.  In case of a conflict, a different
5018  * HeapTupleSatisfiesUpdate return code is returned.
5019  *
5020  * The held status is said to be hypothetical because it might correspond to a
5021  * lock held by a single Xid, i.e. not a real MultiXactId; we express it this
5022  * way for simplicity of API.
5023  */
5024 static HTSU_Result
5025 test_lockmode_for_conflict(MultiXactStatus status, TransactionId xid,
5026                                                    LockTupleMode mode, bool *needwait)
5027 {
5028         MultiXactStatus wantedstatus;
5029
5030         *needwait = false;
5031         wantedstatus = get_mxact_status_for_lock(mode, false);
5032
5033         /*
5034          * Note: we *must* check TransactionIdIsInProgress before
5035          * TransactionIdDidAbort/Commit; see comment at top of tqual.c for an
5036          * explanation.
5037          */
5038         if (TransactionIdIsCurrentTransactionId(xid))
5039         {
5040                 /*
5041                  * Updated by our own transaction? Just return failure.  This
5042                  * shouldn't normally happen.
5043                  */
5044                 return HeapTupleSelfUpdated;
5045         }
5046         else if (TransactionIdIsInProgress(xid))
5047         {
5048                 /*
5049                  * If the locking transaction is running, what we do depends on
5050                  * whether the lock modes conflict: if they do, then we must wait for
5051                  * it to finish; otherwise we can fall through to lock this tuple
5052                  * version without waiting.
5053                  */
5054                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5055                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5056                 {
5057                         *needwait = true;
5058                 }
5059
5060                 /*
5061                  * If we set needwait above, then this value doesn't matter;
5062                  * otherwise, this value signals to caller that it's okay to proceed.
5063                  */
5064                 return HeapTupleMayBeUpdated;
5065         }
5066         else if (TransactionIdDidAbort(xid))
5067                 return HeapTupleMayBeUpdated;
5068         else if (TransactionIdDidCommit(xid))
5069         {
5070                 /*
5071                  * The other transaction committed.  If it was only a locker, then the
5072                  * lock is completely gone now and we can return success; but if it
5073                  * was an update, then what we do depends on whether the two lock
5074                  * modes conflict.  If they conflict, then we must report error to
5075                  * caller. But if they don't, we can fall through to allow the current
5076                  * transaction to lock the tuple.
5077                  *
5078                  * Note: the reason we worry about ISUPDATE here is because as soon as
5079                  * a transaction ends, all its locks are gone and meaningless, and
5080                  * thus we can ignore them; whereas its updates persist.  In the
5081                  * TransactionIdIsInProgress case, above, we don't need to check
5082                  * because we know the lock is still "alive" and thus a conflict needs
5083                  * always be checked.
5084                  */
5085                 if (!ISUPDATE_from_mxstatus(status))
5086                         return HeapTupleMayBeUpdated;
5087
5088                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5089                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5090                         /* bummer */
5091                         return HeapTupleUpdated;
5092
5093                 return HeapTupleMayBeUpdated;
5094         }
5095
5096         /* Not in progress, not aborted, not committed -- must have crashed */
5097         return HeapTupleMayBeUpdated;
5098 }
5099
5100
5101 /*
5102  * Recursive part of heap_lock_updated_tuple
5103  *
5104  * Fetch the tuple pointed to by tid in rel, and mark it as locked by the given
5105  * xid with the given mode; if this tuple is updated, recurse to lock the new
5106  * version as well.
5107  */
5108 static HTSU_Result
5109 heap_lock_updated_tuple_rec(Relation rel, ItemPointer tid, TransactionId xid,
5110                                                         LockTupleMode mode)
5111 {
5112         ItemPointerData tupid;
5113         HeapTupleData mytup;
5114         Buffer          buf;
5115         uint16          new_infomask,
5116                                 new_infomask2,
5117                                 old_infomask,
5118                                 old_infomask2;
5119         TransactionId xmax,
5120                                 new_xmax;
5121         TransactionId priorXmax = InvalidTransactionId;
5122
5123         ItemPointerCopy(tid, &tupid);
5124
5125         for (;;)
5126         {
5127                 new_infomask = 0;
5128                 new_xmax = InvalidTransactionId;
5129                 ItemPointerCopy(&tupid, &(mytup.t_self));
5130
5131                 if (!heap_fetch(rel, SnapshotAny, &mytup, &buf, false, NULL))
5132                 {
5133                         /*
5134                          * if we fail to find the updated version of the tuple, it's
5135                          * because it was vacuumed/pruned away after its creator
5136                          * transaction aborted.  So behave as if we got to the end of the
5137                          * chain, and there's no further tuple to lock: return success to
5138                          * caller.
5139                          */
5140                         return HeapTupleMayBeUpdated;
5141                 }
5142
5143 l4:
5144                 CHECK_FOR_INTERRUPTS();
5145                 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5146
5147                 /*
5148                  * Check the tuple XMIN against prior XMAX, if any.  If we reached the
5149                  * end of the chain, we're done, so return success.
5150                  */
5151                 if (TransactionIdIsValid(priorXmax) &&
5152                         !TransactionIdEquals(HeapTupleHeaderGetXmin(mytup.t_data),
5153                                                                  priorXmax))
5154                 {
5155                         UnlockReleaseBuffer(buf);
5156                         return HeapTupleMayBeUpdated;
5157                 }
5158
5159                 old_infomask = mytup.t_data->t_infomask;
5160                 old_infomask2 = mytup.t_data->t_infomask2;
5161                 xmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5162
5163                 /*
5164                  * If this tuple version has been updated or locked by some concurrent
5165                  * transaction(s), what we do depends on whether our lock mode
5166                  * conflicts with what those other transactions hold, and also on the
5167                  * status of them.
5168                  */
5169                 if (!(old_infomask & HEAP_XMAX_INVALID))
5170                 {
5171                         TransactionId rawxmax;
5172                         bool            needwait;
5173
5174                         rawxmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5175                         if (old_infomask & HEAP_XMAX_IS_MULTI)
5176                         {
5177                                 int                     nmembers;
5178                                 int                     i;
5179                                 MultiXactMember *members;
5180
5181                                 nmembers = GetMultiXactIdMembers(rawxmax, &members, false,
5182                                                                          HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5183                                 for (i = 0; i < nmembers; i++)
5184                                 {
5185                                         HTSU_Result res;
5186
5187                                         res = test_lockmode_for_conflict(members[i].status,
5188                                                                                                          members[i].xid,
5189                                                                                                          mode, &needwait);
5190
5191                                         if (needwait)
5192                                         {
5193                                                 LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5194                                                 XactLockTableWait(members[i].xid, rel,
5195                                                                                   &mytup.t_data->t_ctid,
5196                                                                                   XLTW_LockUpdated);
5197                                                 pfree(members);
5198                                                 goto l4;
5199                                         }
5200                                         if (res != HeapTupleMayBeUpdated)
5201                                         {
5202                                                 UnlockReleaseBuffer(buf);
5203                                                 pfree(members);
5204                                                 return res;
5205                                         }
5206                                 }
5207                                 if (members)
5208                                         pfree(members);
5209                         }
5210                         else
5211                         {
5212                                 HTSU_Result res;
5213                                 MultiXactStatus status;
5214
5215                                 /*
5216                                  * For a non-multi Xmax, we first need to compute the
5217                                  * corresponding MultiXactStatus by using the infomask bits.
5218                                  */
5219                                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5220                                 {
5221                                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5222                                                 status = MultiXactStatusForKeyShare;
5223                                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5224                                                 status = MultiXactStatusForShare;
5225                                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5226                                         {
5227                                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
5228                                                         status = MultiXactStatusForUpdate;
5229                                                 else
5230                                                         status = MultiXactStatusForNoKeyUpdate;
5231                                         }
5232                                         else
5233                                         {
5234                                                 /*
5235                                                  * LOCK_ONLY present alone (a pg_upgraded tuple marked
5236                                                  * as share-locked in the old cluster) shouldn't be
5237                                                  * seen in the middle of an update chain.
5238                                                  */
5239                                                 elog(ERROR, "invalid lock status in tuple");
5240                                         }
5241                                 }
5242                                 else
5243                                 {
5244                                         /* it's an update, but which kind? */
5245                                         if (old_infomask2 & HEAP_KEYS_UPDATED)
5246                                                 status = MultiXactStatusUpdate;
5247                                         else
5248                                                 status = MultiXactStatusNoKeyUpdate;
5249                                 }
5250
5251                                 res = test_lockmode_for_conflict(status, rawxmax, mode,
5252                                                                                                  &needwait);
5253                                 if (needwait)
5254                                 {
5255                                         LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5256                                         XactLockTableWait(rawxmax, rel, &mytup.t_data->t_ctid,
5257                                                                           XLTW_LockUpdated);
5258                                         goto l4;
5259                                 }
5260                                 if (res != HeapTupleMayBeUpdated)
5261                                 {
5262                                         UnlockReleaseBuffer(buf);
5263                                         return res;
5264                                 }
5265                         }
5266                 }
5267
5268                 /* compute the new Xmax and infomask values for the tuple ... */
5269                 compute_new_xmax_infomask(xmax, old_infomask, mytup.t_data->t_infomask2,
5270                                                                   xid, mode, false,
5271                                                                   &new_xmax, &new_infomask, &new_infomask2);
5272
5273                 START_CRIT_SECTION();
5274
5275                 /* ... and set them */
5276                 HeapTupleHeaderSetXmax(mytup.t_data, new_xmax);
5277                 mytup.t_data->t_infomask &= ~HEAP_XMAX_BITS;
5278                 mytup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5279                 mytup.t_data->t_infomask |= new_infomask;
5280                 mytup.t_data->t_infomask2 |= new_infomask2;
5281
5282                 MarkBufferDirty(buf);
5283
5284                 /* XLOG stuff */
5285                 if (RelationNeedsWAL(rel))
5286                 {
5287                         xl_heap_lock_updated xlrec;
5288                         XLogRecPtr      recptr;
5289                         XLogRecData rdata[2];
5290                         Page            page = BufferGetPage(buf);
5291
5292                         xlrec.target.node = rel->rd_node;
5293                         xlrec.target.tid = mytup.t_self;
5294                         xlrec.xmax = new_xmax;
5295                         xlrec.infobits_set = compute_infobits(new_infomask, new_infomask2);
5296
5297                         rdata[0].data = (char *) &xlrec;
5298                         rdata[0].len = SizeOfHeapLockUpdated;
5299                         rdata[0].buffer = InvalidBuffer;
5300                         rdata[0].next = &(rdata[1]);
5301
5302                         rdata[1].data = NULL;
5303                         rdata[1].len = 0;
5304                         rdata[1].buffer = buf;
5305                         rdata[1].buffer_std = true;
5306                         rdata[1].next = NULL;
5307
5308                         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_LOCK_UPDATED, rdata);
5309
5310                         PageSetLSN(page, recptr);
5311                 }
5312
5313                 END_CRIT_SECTION();
5314
5315                 /* if we find the end of update chain, we're done. */
5316                 if (mytup.t_data->t_infomask & HEAP_XMAX_INVALID ||
5317                         ItemPointerEquals(&mytup.t_self, &mytup.t_data->t_ctid) ||
5318                         HeapTupleHeaderIsOnlyLocked(mytup.t_data))
5319                 {
5320                         UnlockReleaseBuffer(buf);
5321                         return HeapTupleMayBeUpdated;
5322                 }
5323
5324                 /* tail recursion */
5325                 priorXmax = HeapTupleHeaderGetUpdateXid(mytup.t_data);
5326                 ItemPointerCopy(&(mytup.t_data->t_ctid), &tupid);
5327                 UnlockReleaseBuffer(buf);
5328         }
5329 }
5330
5331 /*
5332  * heap_lock_updated_tuple
5333  *              Follow update chain when locking an updated tuple, acquiring locks (row
5334  *              marks) on the updated versions.
5335  *
5336  * The initial tuple is assumed to be already locked.
5337  *
5338  * This function doesn't check visibility, it just inconditionally marks the
5339  * tuple(s) as locked.  If any tuple in the updated chain is being deleted
5340  * concurrently (or updated with the key being modified), sleep until the
5341  * transaction doing it is finished.
5342  *
5343  * Note that we don't acquire heavyweight tuple locks on the tuples we walk
5344  * when we have to wait for other transactions to release them, as opposed to
5345  * what heap_lock_tuple does.  The reason is that having more than one
5346  * transaction walking the chain is probably uncommon enough that risk of
5347  * starvation is not likely: one of the preconditions for being here is that
5348  * the snapshot in use predates the update that created this tuple (because we
5349  * started at an earlier version of the tuple), but at the same time such a
5350  * transaction cannot be using repeatable read or serializable isolation
5351  * levels, because that would lead to a serializability failure.
5352  */
5353 static HTSU_Result
5354 heap_lock_updated_tuple(Relation rel, HeapTuple tuple, ItemPointer ctid,
5355                                                 TransactionId xid, LockTupleMode mode)
5356 {
5357         if (!ItemPointerEquals(&tuple->t_self, ctid))
5358         {
5359                 /*
5360                  * If this is the first possibly-multixact-able operation in the
5361                  * current transaction, set my per-backend OldestMemberMXactId
5362                  * setting. We can be certain that the transaction will never become a
5363                  * member of any older MultiXactIds than that.  (We have to do this
5364                  * even if we end up just using our own TransactionId below, since
5365                  * some other backend could incorporate our XID into a MultiXact
5366                  * immediately afterwards.)
5367                  */
5368                 MultiXactIdSetOldestMember();
5369
5370                 return heap_lock_updated_tuple_rec(rel, ctid, xid, mode);
5371         }
5372
5373         /* nothing to lock */
5374         return HeapTupleMayBeUpdated;
5375 }
5376
5377
5378 /*
5379  * heap_inplace_update - update a tuple "in place" (ie, overwrite it)
5380  *
5381  * Overwriting violates both MVCC and transactional safety, so the uses
5382  * of this function in Postgres are extremely limited.  Nonetheless we
5383  * find some places to use it.
5384  *
5385  * The tuple cannot change size, and therefore it's reasonable to assume
5386  * that its null bitmap (if any) doesn't change either.  So we just
5387  * overwrite the data portion of the tuple without touching the null
5388  * bitmap or any of the header fields.
5389  *
5390  * tuple is an in-memory tuple structure containing the data to be written
5391  * over the target tuple.  Also, tuple->t_self identifies the target tuple.
5392  */
5393 void
5394 heap_inplace_update(Relation relation, HeapTuple tuple)
5395 {
5396         Buffer          buffer;
5397         Page            page;
5398         OffsetNumber offnum;
5399         ItemId          lp = NULL;
5400         HeapTupleHeader htup;
5401         uint32          oldlen;
5402         uint32          newlen;
5403
5404         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&(tuple->t_self)));
5405         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5406         page = (Page) BufferGetPage(buffer);
5407
5408         offnum = ItemPointerGetOffsetNumber(&(tuple->t_self));
5409         if (PageGetMaxOffsetNumber(page) >= offnum)
5410                 lp = PageGetItemId(page, offnum);
5411
5412         if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
5413                 elog(ERROR, "heap_inplace_update: invalid lp");
5414
5415         htup = (HeapTupleHeader) PageGetItem(page, lp);
5416
5417         oldlen = ItemIdGetLength(lp) - htup->t_hoff;
5418         newlen = tuple->t_len - tuple->t_data->t_hoff;
5419         if (oldlen != newlen || htup->t_hoff != tuple->t_data->t_hoff)
5420                 elog(ERROR, "heap_inplace_update: wrong tuple length");
5421
5422         /* NO EREPORT(ERROR) from here till changes are logged */
5423         START_CRIT_SECTION();
5424
5425         memcpy((char *) htup + htup->t_hoff,
5426                    (char *) tuple->t_data + tuple->t_data->t_hoff,
5427                    newlen);
5428
5429         MarkBufferDirty(buffer);
5430
5431         /* XLOG stuff */
5432         if (RelationNeedsWAL(relation))
5433         {
5434                 xl_heap_inplace xlrec;
5435                 XLogRecPtr      recptr;
5436                 XLogRecData rdata[2];
5437
5438                 xlrec.target.node = relation->rd_node;
5439                 xlrec.target.tid = tuple->t_self;
5440
5441                 rdata[0].data = (char *) &xlrec;
5442                 rdata[0].len = SizeOfHeapInplace;
5443                 rdata[0].buffer = InvalidBuffer;
5444                 rdata[0].next = &(rdata[1]);
5445
5446                 rdata[1].data = (char *) htup + htup->t_hoff;
5447                 rdata[1].len = newlen;
5448                 rdata[1].buffer = buffer;
5449                 rdata[1].buffer_std = true;
5450                 rdata[1].next = NULL;
5451
5452                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_INPLACE, rdata);
5453
5454                 PageSetLSN(page, recptr);
5455         }
5456
5457         END_CRIT_SECTION();
5458
5459         UnlockReleaseBuffer(buffer);
5460
5461         /*
5462          * Send out shared cache inval if necessary.  Note that because we only
5463          * pass the new version of the tuple, this mustn't be used for any
5464          * operations that could change catcache lookup keys.  But we aren't
5465          * bothering with index updates either, so that's true a fortiori.
5466          */
5467         if (!IsBootstrapProcessingMode())
5468                 CacheInvalidateHeapTuple(relation, tuple, NULL);
5469 }
5470
5471 #define         FRM_NOOP                                0x0001
5472 #define         FRM_INVALIDATE_XMAX             0x0002
5473 #define         FRM_RETURN_IS_XID               0x0004
5474 #define         FRM_RETURN_IS_MULTI             0x0008
5475 #define         FRM_MARK_COMMITTED              0x0010
5476
5477 /*
5478  * FreezeMultiXactId
5479  *              Determine what to do during freezing when a tuple is marked by a
5480  *              MultiXactId.
5481  *
5482  * NB -- this might have the side-effect of creating a new MultiXactId!
5483  *
5484  * "flags" is an output value; it's used to tell caller what to do on return.
5485  * Possible flags are:
5486  * FRM_NOOP
5487  *              don't do anything -- keep existing Xmax
5488  * FRM_INVALIDATE_XMAX
5489  *              mark Xmax as InvalidTransactionId and set XMAX_INVALID flag.
5490  * FRM_RETURN_IS_XID
5491  *              The Xid return value is a single update Xid to set as xmax.
5492  * FRM_MARK_COMMITTED
5493  *              Xmax can be marked as HEAP_XMAX_COMMITTED
5494  * FRM_RETURN_IS_MULTI
5495  *              The return value is a new MultiXactId to set as new Xmax.
5496  *              (caller must obtain proper infomask bits using GetMultiXactIdHintBits)
5497  */
5498 static TransactionId
5499 FreezeMultiXactId(MultiXactId multi, uint16 t_infomask,
5500                                   TransactionId cutoff_xid, MultiXactId cutoff_multi,
5501                                   uint16 *flags)
5502 {
5503         TransactionId xid = InvalidTransactionId;
5504         int                     i;
5505         MultiXactMember *members;
5506         int                     nmembers;
5507         bool            need_replace;
5508         int                     nnewmembers;
5509         MultiXactMember *newmembers;
5510         bool            has_lockers;
5511         TransactionId update_xid;
5512         bool            update_committed;
5513         bool            allow_old;
5514
5515         *flags = 0;
5516
5517         /* We should only be called in Multis */
5518         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
5519
5520         if (!MultiXactIdIsValid(multi))
5521         {
5522                 /* Ensure infomask bits are appropriately set/reset */
5523                 *flags |= FRM_INVALIDATE_XMAX;
5524                 return InvalidTransactionId;
5525         }
5526         else if (MultiXactIdPrecedes(multi, cutoff_multi))
5527         {
5528                 /*
5529                  * This old multi cannot possibly have members still running.  If it
5530                  * was a locker only, it can be removed without any further
5531                  * consideration; but if it contained an update, we might need to
5532                  * preserve it.
5533                  *
5534                  * Don't assert MultiXactIdIsRunning if the multi came from a
5535                  * pg_upgrade'd share-locked tuple, though, as doing that causes an
5536                  * error to be raised unnecessarily.
5537                  */
5538                 Assert((!(t_infomask & HEAP_LOCK_MASK) &&
5539                                 HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)) ||
5540                            !MultiXactIdIsRunning(multi,
5541                                                                          HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)));
5542                 if (HEAP_XMAX_IS_LOCKED_ONLY(t_infomask))
5543                 {
5544                         *flags |= FRM_INVALIDATE_XMAX;
5545                         xid = InvalidTransactionId; /* not strictly necessary */
5546                 }
5547                 else
5548                 {
5549                         /* replace multi by update xid */
5550                         xid = MultiXactIdGetUpdateXid(multi, t_infomask);
5551
5552                         /* wasn't only a lock, xid needs to be valid */
5553                         Assert(TransactionIdIsValid(xid));
5554
5555                         /*
5556                          * If the xid is older than the cutoff, it has to have aborted,
5557                          * otherwise the tuple would have gotten pruned away.
5558                          */
5559                         if (TransactionIdPrecedes(xid, cutoff_xid))
5560                         {
5561                                 Assert(!TransactionIdDidCommit(xid));
5562                                 *flags |= FRM_INVALIDATE_XMAX;
5563                                 xid = InvalidTransactionId;             /* not strictly necessary */
5564                         }
5565                         else
5566                         {
5567                                 *flags |= FRM_RETURN_IS_XID;
5568                         }
5569                 }
5570
5571                 return xid;
5572         }
5573
5574         /*
5575          * This multixact might have or might not have members still running, but
5576          * we know it's valid and is newer than the cutoff point for multis.
5577          * However, some member(s) of it may be below the cutoff for Xids, so we
5578          * need to walk the whole members array to figure out what to do, if
5579          * anything.
5580          */
5581
5582         allow_old = !(t_infomask & HEAP_LOCK_MASK) &&
5583                 HEAP_XMAX_IS_LOCKED_ONLY(t_infomask);
5584         nmembers =
5585                 GetMultiXactIdMembers(multi, &members, allow_old,
5586                                                           HEAP_XMAX_IS_LOCKED_ONLY(t_infomask));
5587         if (nmembers <= 0)
5588         {
5589                 /* Nothing worth keeping */
5590                 *flags |= FRM_INVALIDATE_XMAX;
5591                 return InvalidTransactionId;
5592         }
5593
5594         /* is there anything older than the cutoff? */
5595         need_replace = false;
5596         for (i = 0; i < nmembers; i++)
5597         {
5598                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
5599                 {
5600                         need_replace = true;
5601                         break;
5602                 }
5603         }
5604
5605         /*
5606          * In the simplest case, there is no member older than the cutoff; we can
5607          * keep the existing MultiXactId as is.
5608          */
5609         if (!need_replace)
5610         {
5611                 *flags |= FRM_NOOP;
5612                 pfree(members);
5613                 return InvalidTransactionId;
5614         }
5615
5616         /*
5617          * If the multi needs to be updated, figure out which members do we need
5618          * to keep.
5619          */
5620         nnewmembers = 0;
5621         newmembers = palloc(sizeof(MultiXactMember) * nmembers);
5622         has_lockers = false;
5623         update_xid = InvalidTransactionId;
5624         update_committed = false;
5625
5626         for (i = 0; i < nmembers; i++)
5627         {
5628                 /*
5629                  * Determine whether to keep this member or ignore it.
5630                  */
5631                 if (ISUPDATE_from_mxstatus(members[i].status))
5632                 {
5633                         TransactionId xid = members[i].xid;
5634
5635                         /*
5636                          * It's an update; should we keep it?  If the transaction is known
5637                          * aborted or crashed then it's okay to ignore it, otherwise not.
5638                          * Note that an updater older than cutoff_xid cannot possibly be
5639                          * committed, because HeapTupleSatisfiesVacuum would have returned
5640                          * HEAPTUPLE_DEAD and we would not be trying to freeze the tuple.
5641                          *
5642                          * As with all tuple visibility routines, it's critical to test
5643                          * TransactionIdIsInProgress before TransactionIdDidCommit,
5644                          * because of race conditions explained in detail in tqual.c.
5645                          */
5646                         if (TransactionIdIsCurrentTransactionId(xid) ||
5647                                 TransactionIdIsInProgress(xid))
5648                         {
5649                                 Assert(!TransactionIdIsValid(update_xid));
5650                                 update_xid = xid;
5651                         }
5652                         else if (TransactionIdDidCommit(xid))
5653                         {
5654                                 /*
5655                                  * The transaction committed, so we can tell caller to set
5656                                  * HEAP_XMAX_COMMITTED.  (We can only do this because we know
5657                                  * the transaction is not running.)
5658                                  */
5659                                 Assert(!TransactionIdIsValid(update_xid));
5660                                 update_committed = true;
5661                                 update_xid = xid;
5662                         }
5663
5664                         /*
5665                          * Not in progress, not committed -- must be aborted or crashed;
5666                          * we can ignore it.
5667                          */
5668
5669                         /*
5670                          * Since the tuple wasn't marked HEAPTUPLE_DEAD by vacuum, the
5671                          * update Xid cannot possibly be older than the xid cutoff.
5672                          */
5673                         Assert(!TransactionIdIsValid(update_xid) ||
5674                                    !TransactionIdPrecedes(update_xid, cutoff_xid));
5675
5676                         /*
5677                          * If we determined that it's an Xid corresponding to an update
5678                          * that must be retained, additionally add it to the list of
5679                          * members of the new Multi, in case we end up using that.  (We
5680                          * might still decide to use only an update Xid and not a multi,
5681                          * but it's easier to maintain the list as we walk the old members
5682                          * list.)
5683                          */
5684                         if (TransactionIdIsValid(update_xid))
5685                                 newmembers[nnewmembers++] = members[i];
5686                 }
5687                 else
5688                 {
5689                         /* We only keep lockers if they are still running */
5690                         if (TransactionIdIsCurrentTransactionId(members[i].xid) ||
5691                                 TransactionIdIsInProgress(members[i].xid))
5692                         {
5693                                 /* running locker cannot possibly be older than the cutoff */
5694                                 Assert(!TransactionIdPrecedes(members[i].xid, cutoff_xid));
5695                                 newmembers[nnewmembers++] = members[i];
5696                                 has_lockers = true;
5697                         }
5698                 }
5699         }
5700
5701         pfree(members);
5702
5703         if (nnewmembers == 0)
5704         {
5705                 /* nothing worth keeping!? Tell caller to remove the whole thing */
5706                 *flags |= FRM_INVALIDATE_XMAX;
5707                 xid = InvalidTransactionId;
5708         }
5709         else if (TransactionIdIsValid(update_xid) && !has_lockers)
5710         {
5711                 /*
5712                  * If there's a single member and it's an update, pass it back alone
5713                  * without creating a new Multi.  (XXX we could do this when there's a
5714                  * single remaining locker, too, but that would complicate the API too
5715                  * much; moreover, the case with the single updater is more
5716                  * interesting, because those are longer-lived.)
5717                  */
5718                 Assert(nnewmembers == 1);
5719                 *flags |= FRM_RETURN_IS_XID;
5720                 if (update_committed)
5721                         *flags |= FRM_MARK_COMMITTED;
5722                 xid = update_xid;
5723         }
5724         else
5725         {
5726                 /*
5727                  * Create a new multixact with the surviving members of the previous
5728                  * one, to set as new Xmax in the tuple.
5729                  */
5730                 xid = MultiXactIdCreateFromMembers(nnewmembers, newmembers);
5731                 *flags |= FRM_RETURN_IS_MULTI;
5732         }
5733
5734         pfree(newmembers);
5735
5736         return xid;
5737 }
5738
5739 /*
5740  * heap_prepare_freeze_tuple
5741  *
5742  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
5743  * are older than the specified cutoff XID and cutoff MultiXactId.  If so,
5744  * setup enough state (in the *frz output argument) to later execute and
5745  * WAL-log what we would need to do, and return TRUE.  Return FALSE if nothing
5746  * is to be changed.
5747  *
5748  * Caller is responsible for setting the offset field, if appropriate.
5749  *
5750  * It is assumed that the caller has checked the tuple with
5751  * HeapTupleSatisfiesVacuum() and determined that it is not HEAPTUPLE_DEAD
5752  * (else we should be removing the tuple, not freezing it).
5753  *
5754  * NB: cutoff_xid *must* be <= the current global xmin, to ensure that any
5755  * XID older than it could neither be running nor seen as running by any
5756  * open transaction.  This ensures that the replacement will not change
5757  * anyone's idea of the tuple state.
5758  * Similarly, cutoff_multi must be less than or equal to the smallest
5759  * MultiXactId used by any transaction currently open.
5760  *
5761  * If the tuple is in a shared buffer, caller must hold an exclusive lock on
5762  * that buffer.
5763  *
5764  * NB: It is not enough to set hint bits to indicate something is
5765  * committed/invalid -- they might not be set on a standby, or after crash
5766  * recovery.  We really need to remove old xids.
5767  */
5768 bool
5769 heap_prepare_freeze_tuple(HeapTupleHeader tuple, TransactionId cutoff_xid,
5770                                                   TransactionId cutoff_multi,
5771                                                   xl_heap_freeze_tuple *frz)
5772
5773 {
5774         bool            changed = false;
5775         bool            freeze_xmax = false;
5776         TransactionId xid;
5777
5778         frz->frzflags = 0;
5779         frz->t_infomask2 = tuple->t_infomask2;
5780         frz->t_infomask = tuple->t_infomask;
5781         frz->xmax = HeapTupleHeaderGetRawXmax(tuple);
5782
5783         /* Process xmin */
5784         xid = HeapTupleHeaderGetXmin(tuple);
5785         if (TransactionIdIsNormal(xid) &&
5786                 TransactionIdPrecedes(xid, cutoff_xid))
5787         {
5788                 frz->t_infomask |= HEAP_XMIN_FROZEN;
5789                 changed = true;
5790         }
5791
5792         /*
5793          * Process xmax.  To thoroughly examine the current Xmax value we need to
5794          * resolve a MultiXactId to its member Xids, in case some of them are
5795          * below the given cutoff for Xids.  In that case, those values might need
5796          * freezing, too.  Also, if a multi needs freezing, we cannot simply take
5797          * it out --- if there's a live updater Xid, it needs to be kept.
5798          *
5799          * Make sure to keep heap_tuple_needs_freeze in sync with this.
5800          */
5801         xid = HeapTupleHeaderGetRawXmax(tuple);
5802
5803         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
5804         {
5805                 TransactionId newxmax;
5806                 uint16          flags;
5807
5808                 newxmax = FreezeMultiXactId(xid, tuple->t_infomask,
5809                                                                         cutoff_xid, cutoff_multi, &flags);
5810
5811                 if (flags & FRM_INVALIDATE_XMAX)
5812                         freeze_xmax = true;
5813                 else if (flags & FRM_RETURN_IS_XID)
5814                 {
5815                         /*
5816                          * NB -- some of these transformations are only valid because we
5817                          * know the return Xid is a tuple updater (i.e. not merely a
5818                          * locker.) Also note that the only reason we don't explicitely
5819                          * worry about HEAP_KEYS_UPDATED is because it lives in
5820                          * t_infomask2 rather than t_infomask.
5821                          */
5822                         frz->t_infomask &= ~HEAP_XMAX_BITS;
5823                         frz->xmax = newxmax;
5824                         if (flags & FRM_MARK_COMMITTED)
5825                                 frz->t_infomask &= HEAP_XMAX_COMMITTED;
5826                         changed = true;
5827                 }
5828                 else if (flags & FRM_RETURN_IS_MULTI)
5829                 {
5830                         uint16          newbits;
5831                         uint16          newbits2;
5832
5833                         /*
5834                          * We can't use GetMultiXactIdHintBits directly on the new multi
5835                          * here; that routine initializes the masks to all zeroes, which
5836                          * would lose other bits we need.  Doing it this way ensures all
5837                          * unrelated bits remain untouched.
5838                          */
5839                         frz->t_infomask &= ~HEAP_XMAX_BITS;
5840                         frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5841                         GetMultiXactIdHintBits(newxmax, &newbits, &newbits2);
5842                         frz->t_infomask |= newbits;
5843                         frz->t_infomask2 |= newbits2;
5844
5845                         frz->xmax = newxmax;
5846
5847                         changed = true;
5848                 }
5849                 else
5850                 {
5851                         Assert(flags & FRM_NOOP);
5852                 }
5853         }
5854         else if (TransactionIdIsNormal(xid) &&
5855                          TransactionIdPrecedes(xid, cutoff_xid))
5856         {
5857                 freeze_xmax = true;
5858         }
5859
5860         if (freeze_xmax)
5861         {
5862                 frz->xmax = InvalidTransactionId;
5863
5864                 /*
5865                  * The tuple might be marked either XMAX_INVALID or XMAX_COMMITTED +
5866                  * LOCKED.  Normalize to INVALID just to be sure no one gets confused.
5867                  * Also get rid of the HEAP_KEYS_UPDATED bit.
5868                  */
5869                 frz->t_infomask &= ~HEAP_XMAX_BITS;
5870                 frz->t_infomask |= HEAP_XMAX_INVALID;
5871                 frz->t_infomask2 &= ~HEAP_HOT_UPDATED;
5872                 frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5873                 changed = true;
5874         }
5875
5876         /*
5877          * Old-style VACUUM FULL is gone, but we have to keep this code as long as
5878          * we support having MOVED_OFF/MOVED_IN tuples in the database.
5879          */
5880         if (tuple->t_infomask & HEAP_MOVED)
5881         {
5882                 xid = HeapTupleHeaderGetXvac(tuple);
5883                 if (TransactionIdIsNormal(xid) &&
5884                         TransactionIdPrecedes(xid, cutoff_xid))
5885                 {
5886                         /*
5887                          * If a MOVED_OFF tuple is not dead, the xvac transaction must
5888                          * have failed; whereas a non-dead MOVED_IN tuple must mean the
5889                          * xvac transaction succeeded.
5890                          */
5891                         if (tuple->t_infomask & HEAP_MOVED_OFF)
5892                                 frz->frzflags |= XLH_INVALID_XVAC;
5893                         else
5894                                 frz->frzflags |= XLH_FREEZE_XVAC;
5895
5896                         /*
5897                          * Might as well fix the hint bits too; usually XMIN_COMMITTED
5898                          * will already be set here, but there's a small chance not.
5899                          */
5900                         Assert(!(tuple->t_infomask & HEAP_XMIN_INVALID));
5901                         frz->t_infomask |= HEAP_XMIN_COMMITTED;
5902                         changed = true;
5903                 }
5904         }
5905
5906         return changed;
5907 }
5908
5909 /*
5910  * heap_execute_freeze_tuple
5911  *              Execute the prepared freezing of a tuple.
5912  *
5913  * Caller is responsible for ensuring that no other backend can access the
5914  * storage underlying this tuple, either by holding an exclusive lock on the
5915  * buffer containing it (which is what lazy VACUUM does), or by having it by
5916  * in private storage (which is what CLUSTER and friends do).
5917  *
5918  * Note: it might seem we could make the changes without exclusive lock, since
5919  * TransactionId read/write is assumed atomic anyway.  However there is a race
5920  * condition: someone who just fetched an old XID that we overwrite here could
5921  * conceivably not finish checking the XID against pg_clog before we finish
5922  * the VACUUM and perhaps truncate off the part of pg_clog he needs.  Getting
5923  * exclusive lock ensures no other backend is in process of checking the
5924  * tuple status.  Also, getting exclusive lock makes it safe to adjust the
5925  * infomask bits.
5926  *
5927  * NB: All code in here must be safe to execute during crash recovery!
5928  */
5929 void
5930 heap_execute_freeze_tuple(HeapTupleHeader tuple, xl_heap_freeze_tuple *frz)
5931 {
5932         HeapTupleHeaderSetXmax(tuple, frz->xmax);
5933
5934         if (frz->frzflags & XLH_FREEZE_XVAC)
5935                 HeapTupleHeaderSetXvac(tuple, FrozenTransactionId);
5936
5937         if (frz->frzflags & XLH_INVALID_XVAC)
5938                 HeapTupleHeaderSetXvac(tuple, InvalidTransactionId);
5939
5940         tuple->t_infomask = frz->t_infomask;
5941         tuple->t_infomask2 = frz->t_infomask2;
5942 }
5943
5944 /*
5945  * heap_freeze_tuple
5946  *              Freeze tuple in place, without WAL logging.
5947  *
5948  * Useful for callers like CLUSTER that perform their own WAL logging.
5949  */
5950 bool
5951 heap_freeze_tuple(HeapTupleHeader tuple, TransactionId cutoff_xid,
5952                                   TransactionId cutoff_multi)
5953 {
5954         xl_heap_freeze_tuple frz;
5955         bool            do_freeze;
5956
5957         do_freeze = heap_prepare_freeze_tuple(tuple, cutoff_xid, cutoff_multi,
5958                                                                                   &frz);
5959
5960         /*
5961          * Note that because this is not a WAL-logged operation, we don't need to
5962          * fill in the offset in the freeze record.
5963          */
5964
5965         if (do_freeze)
5966                 heap_execute_freeze_tuple(tuple, &frz);
5967         return do_freeze;
5968 }
5969
5970 /*
5971  * For a given MultiXactId, return the hint bits that should be set in the
5972  * tuple's infomask.
5973  *
5974  * Normally this should be called for a multixact that was just created, and
5975  * so is on our local cache, so the GetMembers call is fast.
5976  */
5977 static void
5978 GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
5979                                            uint16 *new_infomask2)
5980 {
5981         int                     nmembers;
5982         MultiXactMember *members;
5983         int                     i;
5984         uint16          bits = HEAP_XMAX_IS_MULTI;
5985         uint16          bits2 = 0;
5986         bool            has_update = false;
5987         LockTupleMode strongest = LockTupleKeyShare;
5988
5989         /*
5990          * We only use this in multis we just created, so they cannot be values
5991          * pre-pg_upgrade.
5992          */
5993         nmembers = GetMultiXactIdMembers(multi, &members, false, false);
5994
5995         for (i = 0; i < nmembers; i++)
5996         {
5997                 LockTupleMode mode;
5998
5999                 /*
6000                  * Remember the strongest lock mode held by any member of the
6001                  * multixact.
6002                  */
6003                 mode = TUPLOCK_from_mxstatus(members[i].status);
6004                 if (mode > strongest)
6005                         strongest = mode;
6006
6007                 /* See what other bits we need */
6008                 switch (members[i].status)
6009                 {
6010                         case MultiXactStatusForKeyShare:
6011                         case MultiXactStatusForShare:
6012                         case MultiXactStatusForNoKeyUpdate:
6013                                 break;
6014
6015                         case MultiXactStatusForUpdate:
6016                                 bits2 |= HEAP_KEYS_UPDATED;
6017                                 break;
6018
6019                         case MultiXactStatusNoKeyUpdate:
6020                                 has_update = true;
6021                                 break;
6022
6023                         case MultiXactStatusUpdate:
6024                                 bits2 |= HEAP_KEYS_UPDATED;
6025                                 has_update = true;
6026                                 break;
6027                 }
6028         }
6029
6030         if (strongest == LockTupleExclusive ||
6031                 strongest == LockTupleNoKeyExclusive)
6032                 bits |= HEAP_XMAX_EXCL_LOCK;
6033         else if (strongest == LockTupleShare)
6034                 bits |= HEAP_XMAX_SHR_LOCK;
6035         else if (strongest == LockTupleKeyShare)
6036                 bits |= HEAP_XMAX_KEYSHR_LOCK;
6037
6038         if (!has_update)
6039                 bits |= HEAP_XMAX_LOCK_ONLY;
6040
6041         if (nmembers > 0)
6042                 pfree(members);
6043
6044         *new_infomask = bits;
6045         *new_infomask2 = bits2;
6046 }
6047
6048 /*
6049  * MultiXactIdGetUpdateXid
6050  *
6051  * Given a multixact Xmax and corresponding infomask, which does not have the
6052  * HEAP_XMAX_LOCK_ONLY bit set, obtain and return the Xid of the updating
6053  * transaction.
6054  *
6055  * Caller is expected to check the status of the updating transaction, if
6056  * necessary.
6057  */
6058 static TransactionId
6059 MultiXactIdGetUpdateXid(TransactionId xmax, uint16 t_infomask)
6060 {
6061         TransactionId update_xact = InvalidTransactionId;
6062         MultiXactMember *members;
6063         int                     nmembers;
6064
6065         Assert(!(t_infomask & HEAP_XMAX_LOCK_ONLY));
6066         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6067
6068         /*
6069          * Since we know the LOCK_ONLY bit is not set, this cannot be a multi from
6070          * pre-pg_upgrade.
6071          */
6072         nmembers = GetMultiXactIdMembers(xmax, &members, false, false);
6073
6074         if (nmembers > 0)
6075         {
6076                 int                     i;
6077
6078                 for (i = 0; i < nmembers; i++)
6079                 {
6080                         /* Ignore lockers */
6081                         if (!ISUPDATE_from_mxstatus(members[i].status))
6082                                 continue;
6083
6084                         /* there can be at most one updater */
6085                         Assert(update_xact == InvalidTransactionId);
6086                         update_xact = members[i].xid;
6087 #ifndef USE_ASSERT_CHECKING
6088
6089                         /*
6090                          * in an assert-enabled build, walk the whole array to ensure
6091                          * there's no other updater.
6092                          */
6093                         break;
6094 #endif
6095                 }
6096
6097                 pfree(members);
6098         }
6099
6100         return update_xact;
6101 }
6102
6103 /*
6104  * HeapTupleGetUpdateXid
6105  *              As above, but use a HeapTupleHeader
6106  *
6107  * See also HeapTupleHeaderGetUpdateXid, which can be used without previously
6108  * checking the hint bits.
6109  */
6110 TransactionId
6111 HeapTupleGetUpdateXid(HeapTupleHeader tuple)
6112 {
6113         return MultiXactIdGetUpdateXid(HeapTupleHeaderGetRawXmax(tuple),
6114                                                                    tuple->t_infomask);
6115 }
6116
6117 /*
6118  * Do_MultiXactIdWait
6119  *              Actual implementation for the two functions below.
6120  *
6121  * 'multi', 'status' and 'infomask' indicate what to sleep on (the status is
6122  * needed to ensure we only sleep on conflicting members, and the infomask is
6123  * used to optimize multixact access in case it's a lock-only multi); 'nowait'
6124  * indicates whether to use conditional lock acquisition, to allow callers to
6125  * fail if lock is unavailable.  'rel', 'ctid' and 'oper' are used to set up
6126  * context information for error messages.  'remaining', if not NULL, receives
6127  * the number of members that are still running, including any (non-aborted)
6128  * subtransactions of our own transaction.
6129  *
6130  * We do this by sleeping on each member using XactLockTableWait.  Any
6131  * members that belong to the current backend are *not* waited for, however;
6132  * this would not merely be useless but would lead to Assert failure inside
6133  * XactLockTableWait.  By the time this returns, it is certain that all
6134  * transactions *of other backends* that were members of the MultiXactId
6135  * that conflict with the requested status are dead (and no new ones can have
6136  * been added, since it is not legal to add members to an existing
6137  * MultiXactId).
6138  *
6139  * But by the time we finish sleeping, someone else may have changed the Xmax
6140  * of the containing tuple, so the caller needs to iterate on us somehow.
6141  *
6142  * Note that in case we return false, the number of remaining members is
6143  * not to be trusted.
6144  */
6145 static bool
6146 Do_MultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6147                                    uint16 infomask, bool nowait,
6148                                    Relation rel, ItemPointer ctid, XLTW_Oper oper,
6149                                    int *remaining)
6150 {
6151         bool            allow_old;
6152         bool            result = true;
6153         MultiXactMember *members;
6154         int                     nmembers;
6155         int                     remain = 0;
6156
6157         allow_old = !(infomask & HEAP_LOCK_MASK) && HEAP_XMAX_IS_LOCKED_ONLY(infomask);
6158         nmembers = GetMultiXactIdMembers(multi, &members, allow_old,
6159                                                                          HEAP_XMAX_IS_LOCKED_ONLY(infomask));
6160
6161         if (nmembers >= 0)
6162         {
6163                 int                     i;
6164
6165                 for (i = 0; i < nmembers; i++)
6166                 {
6167                         TransactionId memxid = members[i].xid;
6168                         MultiXactStatus memstatus = members[i].status;
6169
6170                         if (TransactionIdIsCurrentTransactionId(memxid))
6171                         {
6172                                 remain++;
6173                                 continue;
6174                         }
6175
6176                         if (!DoLockModesConflict(LOCKMODE_from_mxstatus(memstatus),
6177                                                                          LOCKMODE_from_mxstatus(status)))
6178                         {
6179                                 if (remaining && TransactionIdIsInProgress(memxid))
6180                                         remain++;
6181                                 continue;
6182                         }
6183
6184                         /*
6185                          * This member conflicts with our multi, so we have to sleep (or
6186                          * return failure, if asked to avoid waiting.)
6187                          *
6188                          * Note that we don't set up an error context callback ourselves,
6189                          * but instead we pass the info down to XactLockTableWait.  This
6190                          * might seem a bit wasteful because the context is set up and
6191                          * tore down for each member of the multixact, but in reality it
6192                          * should be barely noticeable, and it avoids duplicate code.
6193                          */
6194                         if (nowait)
6195                         {
6196                                 result = ConditionalXactLockTableWait(memxid);
6197                                 if (!result)
6198                                         break;
6199                         }
6200                         else
6201                                 XactLockTableWait(memxid, rel, ctid, oper);
6202                 }
6203
6204                 pfree(members);
6205         }
6206
6207         if (remaining)
6208                 *remaining = remain;
6209
6210         return result;
6211 }
6212
6213 /*
6214  * MultiXactIdWait
6215  *              Sleep on a MultiXactId.
6216  *
6217  * By the time we finish sleeping, someone else may have changed the Xmax
6218  * of the containing tuple, so the caller needs to iterate on us somehow.
6219  *
6220  * We return (in *remaining, if not NULL) the number of members that are still
6221  * running, including any (non-aborted) subtransactions of our own transaction.
6222  */
6223 static void
6224 MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
6225                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
6226                                 int *remaining)
6227 {
6228         (void) Do_MultiXactIdWait(multi, status, infomask, false,
6229                                                           rel, ctid, oper, remaining);
6230 }
6231
6232 /*
6233  * ConditionalMultiXactIdWait
6234  *              As above, but only lock if we can get the lock without blocking.
6235  *
6236  * By the time we finish sleeping, someone else may have changed the Xmax
6237  * of the containing tuple, so the caller needs to iterate on us somehow.
6238  *
6239  * If the multixact is now all gone, return true.  Returns false if some
6240  * transactions might still be running.
6241  *
6242  * We return (in *remaining, if not NULL) the number of members that are still
6243  * running, including any (non-aborted) subtransactions of our own transaction.
6244  */
6245 static bool
6246 ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6247                                                    uint16 infomask, Relation rel, int *remaining)
6248 {
6249         return Do_MultiXactIdWait(multi, status, infomask, true,
6250                                                           rel, NULL, XLTW_None, remaining);
6251 }
6252
6253 /*
6254  * heap_tuple_needs_freeze
6255  *
6256  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6257  * are older than the specified cutoff XID or MultiXactId.  If so, return TRUE.
6258  *
6259  * It doesn't matter whether the tuple is alive or dead, we are checking
6260  * to see if a tuple needs to be removed or frozen to avoid wraparound.
6261  *
6262  * NB: Cannot rely on hint bits here, they might not be set after a crash or
6263  * on a standby.
6264  */
6265 bool
6266 heap_tuple_needs_freeze(HeapTupleHeader tuple, TransactionId cutoff_xid,
6267                                                 MultiXactId cutoff_multi, Buffer buf)
6268 {
6269         TransactionId xid;
6270
6271         xid = HeapTupleHeaderGetXmin(tuple);
6272         if (TransactionIdIsNormal(xid) &&
6273                 TransactionIdPrecedes(xid, cutoff_xid))
6274                 return true;
6275
6276         /*
6277          * The considerations for multixacts are complicated; look at
6278          * heap_freeze_tuple for justifications.  This routine had better be in
6279          * sync with that one!
6280          */
6281         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
6282         {
6283                 MultiXactId multi;
6284
6285                 multi = HeapTupleHeaderGetRawXmax(tuple);
6286                 if (!MultiXactIdIsValid(multi))
6287                 {
6288                         /* no xmax set, ignore */
6289                         ;
6290                 }
6291                 else if (MultiXactIdPrecedes(multi, cutoff_multi))
6292                         return true;
6293                 else
6294                 {
6295                         MultiXactMember *members;
6296                         int                     nmembers;
6297                         int                     i;
6298                         bool            allow_old;
6299
6300                         /* need to check whether any member of the mxact is too old */
6301
6302                         allow_old = !(tuple->t_infomask & HEAP_LOCK_MASK) &&
6303                                 HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask);
6304                         nmembers = GetMultiXactIdMembers(multi, &members, allow_old,
6305                                                                 HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask));
6306
6307                         for (i = 0; i < nmembers; i++)
6308                         {
6309                                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
6310                                 {
6311                                         pfree(members);
6312                                         return true;
6313                                 }
6314                         }
6315                         if (nmembers > 0)
6316                                 pfree(members);
6317                 }
6318         }
6319         else
6320         {
6321                 xid = HeapTupleHeaderGetRawXmax(tuple);
6322                 if (TransactionIdIsNormal(xid) &&
6323                         TransactionIdPrecedes(xid, cutoff_xid))
6324                         return true;
6325         }
6326
6327         if (tuple->t_infomask & HEAP_MOVED)
6328         {
6329                 xid = HeapTupleHeaderGetXvac(tuple);
6330                 if (TransactionIdIsNormal(xid) &&
6331                         TransactionIdPrecedes(xid, cutoff_xid))
6332                         return true;
6333         }
6334
6335         return false;
6336 }
6337
6338 /* ----------------
6339  *              heap_markpos    - mark scan position
6340  * ----------------
6341  */
6342 void
6343 heap_markpos(HeapScanDesc scan)
6344 {
6345         /* Note: no locking manipulations needed */
6346
6347         if (scan->rs_ctup.t_data != NULL)
6348         {
6349                 scan->rs_mctid = scan->rs_ctup.t_self;
6350                 if (scan->rs_pageatatime)
6351                         scan->rs_mindex = scan->rs_cindex;
6352         }
6353         else
6354                 ItemPointerSetInvalid(&scan->rs_mctid);
6355 }
6356
6357 /* ----------------
6358  *              heap_restrpos   - restore position to marked location
6359  * ----------------
6360  */
6361 void
6362 heap_restrpos(HeapScanDesc scan)
6363 {
6364         /* XXX no amrestrpos checking that ammarkpos called */
6365
6366         if (!ItemPointerIsValid(&scan->rs_mctid))
6367         {
6368                 scan->rs_ctup.t_data = NULL;
6369
6370                 /*
6371                  * unpin scan buffers
6372                  */
6373                 if (BufferIsValid(scan->rs_cbuf))
6374                         ReleaseBuffer(scan->rs_cbuf);
6375                 scan->rs_cbuf = InvalidBuffer;
6376                 scan->rs_cblock = InvalidBlockNumber;
6377                 scan->rs_inited = false;
6378         }
6379         else
6380         {
6381                 /*
6382                  * If we reached end of scan, rs_inited will now be false.  We must
6383                  * reset it to true to keep heapgettup from doing the wrong thing.
6384                  */
6385                 scan->rs_inited = true;
6386                 scan->rs_ctup.t_self = scan->rs_mctid;
6387                 if (scan->rs_pageatatime)
6388                 {
6389                         scan->rs_cindex = scan->rs_mindex;
6390                         heapgettup_pagemode(scan,
6391                                                                 NoMovementScanDirection,
6392                                                                 0,              /* needn't recheck scan keys */
6393                                                                 NULL);
6394                 }
6395                 else
6396                         heapgettup(scan,
6397                                            NoMovementScanDirection,
6398                                            0,           /* needn't recheck scan keys */
6399                                            NULL);
6400         }
6401 }
6402
6403 /*
6404  * If 'tuple' contains any visible XID greater than latestRemovedXid,
6405  * ratchet forwards latestRemovedXid to the greatest one found.
6406  * This is used as the basis for generating Hot Standby conflicts, so
6407  * if a tuple was never visible then removing it should not conflict
6408  * with queries.
6409  */
6410 void
6411 HeapTupleHeaderAdvanceLatestRemovedXid(HeapTupleHeader tuple,
6412                                                                            TransactionId *latestRemovedXid)
6413 {
6414         TransactionId xmin = HeapTupleHeaderGetXmin(tuple);
6415         TransactionId xmax = HeapTupleHeaderGetUpdateXid(tuple);
6416         TransactionId xvac = HeapTupleHeaderGetXvac(tuple);
6417
6418         if (tuple->t_infomask & HEAP_MOVED)
6419         {
6420                 if (TransactionIdPrecedes(*latestRemovedXid, xvac))
6421                         *latestRemovedXid = xvac;
6422         }
6423
6424         /*
6425          * Ignore tuples inserted by an aborted transaction or if the tuple was
6426          * updated/deleted by the inserting transaction.
6427          *
6428          * Look for a committed hint bit, or if no xmin bit is set, check clog.
6429          * This needs to work on both master and standby, where it is used to
6430          * assess btree delete records.
6431          */
6432         if (HeapTupleHeaderXminCommitted(tuple) ||
6433                 (!HeapTupleHeaderXminInvalid(tuple) && TransactionIdDidCommit(xmin)))
6434         {
6435                 if (xmax != xmin &&
6436                         TransactionIdFollows(xmax, *latestRemovedXid))
6437                         *latestRemovedXid = xmax;
6438         }
6439
6440         /* *latestRemovedXid may still be invalid at end */
6441 }
6442
6443 /*
6444  * Perform XLogInsert to register a heap cleanup info message. These
6445  * messages are sent once per VACUUM and are required because
6446  * of the phasing of removal operations during a lazy VACUUM.
6447  * see comments for vacuum_log_cleanup_info().
6448  */
6449 XLogRecPtr
6450 log_heap_cleanup_info(RelFileNode rnode, TransactionId latestRemovedXid)
6451 {
6452         xl_heap_cleanup_info xlrec;
6453         XLogRecPtr      recptr;
6454         XLogRecData rdata;
6455
6456         xlrec.node = rnode;
6457         xlrec.latestRemovedXid = latestRemovedXid;
6458
6459         rdata.data = (char *) &xlrec;
6460         rdata.len = SizeOfHeapCleanupInfo;
6461         rdata.buffer = InvalidBuffer;
6462         rdata.next = NULL;
6463
6464         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_CLEANUP_INFO, &rdata);
6465
6466         return recptr;
6467 }
6468
6469 /*
6470  * Perform XLogInsert for a heap-clean operation.  Caller must already
6471  * have modified the buffer and marked it dirty.
6472  *
6473  * Note: prior to Postgres 8.3, the entries in the nowunused[] array were
6474  * zero-based tuple indexes.  Now they are one-based like other uses
6475  * of OffsetNumber.
6476  *
6477  * We also include latestRemovedXid, which is the greatest XID present in
6478  * the removed tuples. That allows recovery processing to cancel or wait
6479  * for long standby queries that can still see these tuples.
6480  */
6481 XLogRecPtr
6482 log_heap_clean(Relation reln, Buffer buffer,
6483                            OffsetNumber *redirected, int nredirected,
6484                            OffsetNumber *nowdead, int ndead,
6485                            OffsetNumber *nowunused, int nunused,
6486                            TransactionId latestRemovedXid)
6487 {
6488         xl_heap_clean xlrec;
6489         uint8           info;
6490         XLogRecPtr      recptr;
6491         XLogRecData rdata[4];
6492
6493         /* Caller should not call me on a non-WAL-logged relation */
6494         Assert(RelationNeedsWAL(reln));
6495
6496         xlrec.node = reln->rd_node;
6497         xlrec.block = BufferGetBlockNumber(buffer);
6498         xlrec.latestRemovedXid = latestRemovedXid;
6499         xlrec.nredirected = nredirected;
6500         xlrec.ndead = ndead;
6501
6502         rdata[0].data = (char *) &xlrec;
6503         rdata[0].len = SizeOfHeapClean;
6504         rdata[0].buffer = InvalidBuffer;
6505         rdata[0].next = &(rdata[1]);
6506
6507         /*
6508          * The OffsetNumber arrays are not actually in the buffer, but we pretend
6509          * that they are.  When XLogInsert stores the whole buffer, the offset
6510          * arrays need not be stored too.  Note that even if all three arrays are
6511          * empty, we want to expose the buffer as a candidate for whole-page
6512          * storage, since this record type implies a defragmentation operation
6513          * even if no item pointers changed state.
6514          */
6515         if (nredirected > 0)
6516         {
6517                 rdata[1].data = (char *) redirected;
6518                 rdata[1].len = nredirected * sizeof(OffsetNumber) * 2;
6519         }
6520         else
6521         {
6522                 rdata[1].data = NULL;
6523                 rdata[1].len = 0;
6524         }
6525         rdata[1].buffer = buffer;
6526         rdata[1].buffer_std = true;
6527         rdata[1].next = &(rdata[2]);
6528
6529         if (ndead > 0)
6530         {
6531                 rdata[2].data = (char *) nowdead;
6532                 rdata[2].len = ndead * sizeof(OffsetNumber);
6533         }
6534         else
6535         {
6536                 rdata[2].data = NULL;
6537                 rdata[2].len = 0;
6538         }
6539         rdata[2].buffer = buffer;
6540         rdata[2].buffer_std = true;
6541         rdata[2].next = &(rdata[3]);
6542
6543         if (nunused > 0)
6544         {
6545                 rdata[3].data = (char *) nowunused;
6546                 rdata[3].len = nunused * sizeof(OffsetNumber);
6547         }
6548         else
6549         {
6550                 rdata[3].data = NULL;
6551                 rdata[3].len = 0;
6552         }
6553         rdata[3].buffer = buffer;
6554         rdata[3].buffer_std = true;
6555         rdata[3].next = NULL;
6556
6557         info = XLOG_HEAP2_CLEAN;
6558         recptr = XLogInsert(RM_HEAP2_ID, info, rdata);
6559
6560         return recptr;
6561 }
6562
6563 /*
6564  * Perform XLogInsert for a heap-freeze operation.  Caller must have already
6565  * modified the buffer and marked it dirty.
6566  */
6567 XLogRecPtr
6568 log_heap_freeze(Relation reln, Buffer buffer, TransactionId cutoff_xid,
6569                                 xl_heap_freeze_tuple *tuples, int ntuples)
6570 {
6571         xl_heap_freeze_page xlrec;
6572         XLogRecPtr      recptr;
6573         XLogRecData rdata[2];
6574
6575         /* Caller should not call me on a non-WAL-logged relation */
6576         Assert(RelationNeedsWAL(reln));
6577         /* nor when there are no tuples to freeze */
6578         Assert(ntuples > 0);
6579
6580         xlrec.node = reln->rd_node;
6581         xlrec.block = BufferGetBlockNumber(buffer);
6582         xlrec.cutoff_xid = cutoff_xid;
6583         xlrec.ntuples = ntuples;
6584
6585         rdata[0].data = (char *) &xlrec;
6586         rdata[0].len = SizeOfHeapFreezePage;
6587         rdata[0].buffer = InvalidBuffer;
6588         rdata[0].next = &(rdata[1]);
6589
6590         /*
6591          * The freeze plan array is not actually in the buffer, but pretend that
6592          * it is.  When XLogInsert stores the whole buffer, the freeze plan need
6593          * not be stored too.
6594          */
6595         rdata[1].data = (char *) tuples;
6596         rdata[1].len = ntuples * sizeof(xl_heap_freeze_tuple);
6597         rdata[1].buffer = buffer;
6598         rdata[1].buffer_std = true;
6599         rdata[1].next = NULL;
6600
6601         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_FREEZE_PAGE, rdata);
6602
6603         return recptr;
6604 }
6605
6606 /*
6607  * Perform XLogInsert for a heap-visible operation.  'block' is the block
6608  * being marked all-visible, and vm_buffer is the buffer containing the
6609  * corresponding visibility map block.  Both should have already been modified
6610  * and dirtied.
6611  *
6612  * If checksums are enabled, we also add the heap_buffer to the chain to
6613  * protect it from being torn.
6614  */
6615 XLogRecPtr
6616 log_heap_visible(RelFileNode rnode, Buffer heap_buffer, Buffer vm_buffer,
6617                                  TransactionId cutoff_xid)
6618 {
6619         xl_heap_visible xlrec;
6620         XLogRecPtr      recptr;
6621         XLogRecData rdata[3];
6622
6623         Assert(BufferIsValid(heap_buffer));
6624         Assert(BufferIsValid(vm_buffer));
6625
6626         xlrec.node = rnode;
6627         xlrec.block = BufferGetBlockNumber(heap_buffer);
6628         xlrec.cutoff_xid = cutoff_xid;
6629
6630         rdata[0].data = (char *) &xlrec;
6631         rdata[0].len = SizeOfHeapVisible;
6632         rdata[0].buffer = InvalidBuffer;
6633         rdata[0].next = &(rdata[1]);
6634
6635         rdata[1].data = NULL;
6636         rdata[1].len = 0;
6637         rdata[1].buffer = vm_buffer;
6638         rdata[1].buffer_std = false;
6639         rdata[1].next = NULL;
6640
6641         if (XLogHintBitIsNeeded())
6642         {
6643                 rdata[1].next = &(rdata[2]);
6644
6645                 rdata[2].data = NULL;
6646                 rdata[2].len = 0;
6647                 rdata[2].buffer = heap_buffer;
6648                 rdata[2].buffer_std = true;
6649                 rdata[2].next = NULL;
6650         }
6651
6652         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_VISIBLE, rdata);
6653
6654         return recptr;
6655 }
6656
6657 /*
6658  * Perform XLogInsert for a heap-update operation.  Caller must already
6659  * have modified the buffer(s) and marked them dirty.
6660  */
6661 static XLogRecPtr
6662 log_heap_update(Relation reln, Buffer oldbuf,
6663                                 Buffer newbuf, HeapTuple oldtup, HeapTuple newtup,
6664                                 HeapTuple old_key_tuple,
6665                                 bool all_visible_cleared, bool new_all_visible_cleared)
6666 {
6667         xl_heap_update xlrec;
6668         xl_heap_header_len xlhdr;
6669         xl_heap_header_len xlhdr_idx;
6670         uint8           info;
6671         uint16          prefix_suffix[2];
6672         uint16          prefixlen = 0,
6673                                 suffixlen = 0;
6674         XLogRecPtr      recptr;
6675         XLogRecData rdata[9];
6676         Page            page = BufferGetPage(newbuf);
6677         bool            need_tuple_data = RelationIsLogicallyLogged(reln);
6678         int                     nr;
6679         Buffer          newbufref;
6680
6681         /* Caller should not call me on a non-WAL-logged relation */
6682         Assert(RelationNeedsWAL(reln));
6683
6684         if (HeapTupleIsHeapOnly(newtup))
6685                 info = XLOG_HEAP_HOT_UPDATE;
6686         else
6687                 info = XLOG_HEAP_UPDATE;
6688
6689         /*
6690          * If the old and new tuple are on the same page, we only need to log the
6691          * parts of the new tuple that were changed.  That saves on the amount of
6692          * WAL we need to write.  Currently, we just count any unchanged bytes in
6693          * the beginning and end of the tuple.  That's quick to check, and
6694          * perfectly covers the common case that only one field is updated.
6695          *
6696          * We could do this even if the old and new tuple are on different pages,
6697          * but only if we don't make a full-page image of the old page, which is
6698          * difficult to know in advance.  Also, if the old tuple is corrupt for
6699          * some reason, it would allow the corruption to propagate the new page,
6700          * so it seems best to avoid.  Under the general assumption that most
6701          * updates tend to create the new tuple version on the same page, there
6702          * isn't much to be gained by doing this across pages anyway.
6703          *
6704          * Skip this if we're taking a full-page image of the new page, as we
6705          * don't include the new tuple in the WAL record in that case.  Also
6706          * disable if wal_level='logical', as logical decoding needs to be able to
6707          * read the new tuple in whole from the WAL record alone.
6708          */
6709         if (oldbuf == newbuf && !need_tuple_data &&
6710                 !XLogCheckBufferNeedsBackup(newbuf))
6711         {
6712                 char       *oldp = (char *) oldtup->t_data + oldtup->t_data->t_hoff;
6713                 char       *newp = (char *) newtup->t_data + newtup->t_data->t_hoff;
6714                 int                     oldlen = oldtup->t_len - oldtup->t_data->t_hoff;
6715                 int                     newlen = newtup->t_len - newtup->t_data->t_hoff;
6716
6717                 /* Check for common prefix between old and new tuple */
6718                 for (prefixlen = 0; prefixlen < Min(oldlen, newlen); prefixlen++)
6719                 {
6720                         if (newp[prefixlen] != oldp[prefixlen])
6721                                 break;
6722                 }
6723
6724                 /*
6725                  * Storing the length of the prefix takes 2 bytes, so we need to save
6726                  * at least 3 bytes or there's no point.
6727                  */
6728                 if (prefixlen < 3)
6729                         prefixlen = 0;
6730
6731                 /* Same for suffix */
6732                 for (suffixlen = 0; suffixlen < Min(oldlen, newlen) - prefixlen; suffixlen++)
6733                 {
6734                         if (newp[newlen - suffixlen - 1] != oldp[oldlen - suffixlen - 1])
6735                                 break;
6736                 }
6737                 if (suffixlen < 3)
6738                         suffixlen = 0;
6739         }
6740
6741         xlrec.target.node = reln->rd_node;
6742         xlrec.target.tid = oldtup->t_self;
6743         xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data);
6744         xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask,
6745                                                                                           oldtup->t_data->t_infomask2);
6746         xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data);
6747         xlrec.flags = 0;
6748         if (all_visible_cleared)
6749                 xlrec.flags |= XLOG_HEAP_ALL_VISIBLE_CLEARED;
6750         xlrec.newtid = newtup->t_self;
6751         if (new_all_visible_cleared)
6752                 xlrec.flags |= XLOG_HEAP_NEW_ALL_VISIBLE_CLEARED;
6753         if (prefixlen > 0)
6754                 xlrec.flags |= XLOG_HEAP_PREFIX_FROM_OLD;
6755         if (suffixlen > 0)
6756                 xlrec.flags |= XLOG_HEAP_SUFFIX_FROM_OLD;
6757
6758         /* If new tuple is the single and first tuple on page... */
6759         if (ItemPointerGetOffsetNumber(&(newtup->t_self)) == FirstOffsetNumber &&
6760                 PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
6761         {
6762                 info |= XLOG_HEAP_INIT_PAGE;
6763                 newbufref = InvalidBuffer;
6764         }
6765         else
6766                 newbufref = newbuf;
6767
6768         rdata[0].data = NULL;
6769         rdata[0].len = 0;
6770         rdata[0].buffer = oldbuf;
6771         rdata[0].buffer_std = true;
6772         rdata[0].next = &(rdata[1]);
6773
6774         rdata[1].data = (char *) &xlrec;
6775         rdata[1].len = SizeOfHeapUpdate;
6776         rdata[1].buffer = InvalidBuffer;
6777         rdata[1].next = &(rdata[2]);
6778
6779         /* prefix and/or suffix length fields */
6780         if (prefixlen > 0 || suffixlen > 0)
6781         {
6782                 if (prefixlen > 0 && suffixlen > 0)
6783                 {
6784                         prefix_suffix[0] = prefixlen;
6785                         prefix_suffix[1] = suffixlen;
6786                         rdata[2].data = (char *) &prefix_suffix;
6787                         rdata[2].len = 2 * sizeof(uint16);
6788                 }
6789                 else if (prefixlen > 0)
6790                 {
6791                         rdata[2].data = (char *) &prefixlen;
6792                         rdata[2].len = sizeof(uint16);
6793                 }
6794                 else
6795                 {
6796                         rdata[2].data = (char *) &suffixlen;
6797                         rdata[2].len = sizeof(uint16);
6798                 }
6799                 rdata[2].buffer = newbufref;
6800                 rdata[2].buffer_std = true;
6801                 rdata[2].next = &(rdata[3]);
6802                 nr = 3;
6803         }
6804         else
6805                 nr = 2;
6806
6807         xlhdr.header.t_infomask2 = newtup->t_data->t_infomask2;
6808         xlhdr.header.t_infomask = newtup->t_data->t_infomask;
6809         xlhdr.header.t_hoff = newtup->t_data->t_hoff;
6810         Assert(offsetof(HeapTupleHeaderData, t_bits) +prefixlen + suffixlen <= newtup->t_len);
6811         xlhdr.t_len = newtup->t_len - offsetof(HeapTupleHeaderData, t_bits) -prefixlen - suffixlen;
6812
6813         /*
6814          * As with insert records, we need not store this rdata segment if we
6815          * decide to store the whole buffer instead, unless we're doing logical
6816          * decoding.
6817          */
6818         rdata[nr].data = (char *) &xlhdr;
6819         rdata[nr].len = SizeOfHeapHeaderLen;
6820         rdata[nr].buffer = need_tuple_data ? InvalidBuffer : newbufref;
6821         rdata[nr].buffer_std = true;
6822         rdata[nr].next = &(rdata[nr + 1]);
6823         nr++;
6824
6825         /*
6826          * PG73FORMAT: write bitmap [+ padding] [+ oid] + data
6827          *
6828          * The 'data' doesn't include the common prefix or suffix.
6829          */
6830         if (prefixlen == 0)
6831         {
6832                 rdata[nr].data = ((char *) newtup->t_data) + offsetof(HeapTupleHeaderData, t_bits);
6833                 rdata[nr].len = newtup->t_len - offsetof(HeapTupleHeaderData, t_bits) -suffixlen;
6834                 rdata[nr].buffer = need_tuple_data ? InvalidBuffer : newbufref;
6835                 rdata[nr].buffer_std = true;
6836                 rdata[nr].next = NULL;
6837                 nr++;
6838         }
6839         else
6840         {
6841                 /*
6842                  * Have to write the null bitmap and data after the common prefix as
6843                  * two separate rdata entries.
6844                  */
6845                 /* bitmap [+ padding] [+ oid] */
6846                 if (newtup->t_data->t_hoff - offsetof(HeapTupleHeaderData, t_bits) >0)
6847                 {
6848                         rdata[nr - 1].next = &(rdata[nr]);
6849                         rdata[nr].data = ((char *) newtup->t_data) + offsetof(HeapTupleHeaderData, t_bits);
6850                         rdata[nr].len = newtup->t_data->t_hoff - offsetof(HeapTupleHeaderData, t_bits);
6851                         rdata[nr].buffer = need_tuple_data ? InvalidBuffer : newbufref;
6852                         rdata[nr].buffer_std = true;
6853                         rdata[nr].next = NULL;
6854                         nr++;
6855                 }
6856
6857                 /* data after common prefix */
6858                 rdata[nr - 1].next = &(rdata[nr]);
6859                 rdata[nr].data = ((char *) newtup->t_data) + newtup->t_data->t_hoff + prefixlen;
6860                 rdata[nr].len = newtup->t_len - newtup->t_data->t_hoff - prefixlen - suffixlen;
6861                 rdata[nr].buffer = need_tuple_data ? InvalidBuffer : newbufref;
6862                 rdata[nr].buffer_std = true;
6863                 rdata[nr].next = NULL;
6864                 nr++;
6865         }
6866
6867         /*
6868          * Separate storage for the FPW buffer reference of the new page in the
6869          * wal_level >= logical case.
6870          */
6871         if (need_tuple_data)
6872         {
6873                 rdata[nr - 1].next = &(rdata[nr]);
6874
6875                 rdata[nr].data = NULL,
6876                         rdata[nr].len = 0;
6877                 rdata[nr].buffer = newbufref;
6878                 rdata[nr].buffer_std = true;
6879                 rdata[nr].next = NULL;
6880                 nr++;
6881
6882                 xlrec.flags |= XLOG_HEAP_CONTAINS_NEW_TUPLE;
6883
6884                 /* We need to log a tuple identity */
6885                 if (old_key_tuple)
6886                 {
6887                         /* don't really need this, but its more comfy to decode */
6888                         xlhdr_idx.header.t_infomask2 = old_key_tuple->t_data->t_infomask2;
6889                         xlhdr_idx.header.t_infomask = old_key_tuple->t_data->t_infomask;
6890                         xlhdr_idx.header.t_hoff = old_key_tuple->t_data->t_hoff;
6891                         xlhdr_idx.t_len = old_key_tuple->t_len;
6892
6893                         rdata[nr - 1].next = &(rdata[nr]);
6894                         rdata[nr].data = (char *) &xlhdr_idx;
6895                         rdata[nr].len = SizeOfHeapHeaderLen;
6896                         rdata[nr].buffer = InvalidBuffer;
6897                         rdata[nr].next = &(rdata[nr + 1]);
6898                         nr++;
6899
6900                         /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
6901                         rdata[nr].data = (char *) old_key_tuple->t_data
6902                                 + offsetof(HeapTupleHeaderData, t_bits);
6903                         rdata[nr].len = old_key_tuple->t_len
6904                                 - offsetof(HeapTupleHeaderData, t_bits);
6905                         rdata[nr].buffer = InvalidBuffer;
6906                         rdata[nr].next = NULL;
6907                         nr++;
6908
6909                         if (reln->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
6910                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_TUPLE;
6911                         else
6912                                 xlrec.flags |= XLOG_HEAP_CONTAINS_OLD_KEY;
6913                 }
6914         }
6915
6916         recptr = XLogInsert(RM_HEAP_ID, info, rdata);
6917
6918         return recptr;
6919 }
6920
6921 /*
6922  * Perform XLogInsert of a XLOG_HEAP2_NEW_CID record
6923  *
6924  * This is only used in wal_level >= WAL_LEVEL_LOGICAL, and only for catalog
6925  * tuples.
6926  */
6927 static XLogRecPtr
6928 log_heap_new_cid(Relation relation, HeapTuple tup)
6929 {
6930         xl_heap_new_cid xlrec;
6931
6932         XLogRecPtr      recptr;
6933         XLogRecData rdata[1];
6934         HeapTupleHeader hdr = tup->t_data;
6935
6936         Assert(ItemPointerIsValid(&tup->t_self));
6937         Assert(tup->t_tableOid != InvalidOid);
6938
6939         xlrec.top_xid = GetTopTransactionId();
6940         xlrec.target.node = relation->rd_node;
6941         xlrec.target.tid = tup->t_self;
6942
6943         /*
6944          * If the tuple got inserted & deleted in the same TX we definitely have a
6945          * combocid, set cmin and cmax.
6946          */
6947         if (hdr->t_infomask & HEAP_COMBOCID)
6948         {
6949                 Assert(!(hdr->t_infomask & HEAP_XMAX_INVALID));
6950                 Assert(!HeapTupleHeaderXminInvalid(hdr));
6951                 xlrec.cmin = HeapTupleHeaderGetCmin(hdr);
6952                 xlrec.cmax = HeapTupleHeaderGetCmax(hdr);
6953                 xlrec.combocid = HeapTupleHeaderGetRawCommandId(hdr);
6954         }
6955         /* No combocid, so only cmin or cmax can be set by this TX */
6956         else
6957         {
6958                 /*
6959                  * Tuple inserted.
6960                  *
6961                  * We need to check for LOCK ONLY because multixacts might be
6962                  * transferred to the new tuple in case of FOR KEY SHARE updates in
6963                  * which case there will be a xmax, although the tuple just got
6964                  * inserted.
6965                  */
6966                 if (hdr->t_infomask & HEAP_XMAX_INVALID ||
6967                         HEAP_XMAX_IS_LOCKED_ONLY(hdr->t_infomask))
6968                 {
6969                         xlrec.cmin = HeapTupleHeaderGetRawCommandId(hdr);
6970                         xlrec.cmax = InvalidCommandId;
6971                 }
6972                 /* Tuple from a different tx updated or deleted. */
6973                 else
6974                 {
6975                         xlrec.cmin = InvalidCommandId;
6976                         xlrec.cmax = HeapTupleHeaderGetRawCommandId(hdr);
6977
6978                 }
6979                 xlrec.combocid = InvalidCommandId;
6980         }
6981
6982         rdata[0].data = (char *) &xlrec;
6983         rdata[0].len = SizeOfHeapNewCid;
6984         rdata[0].buffer = InvalidBuffer;
6985         rdata[0].next = NULL;
6986
6987         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_NEW_CID, rdata);
6988
6989         return recptr;
6990 }
6991
6992 /*
6993  * Build a heap tuple representing the configured REPLICA IDENTITY to represent
6994  * the old tuple in a UPDATE or DELETE.
6995  *
6996  * Returns NULL if there's no need to log a identity or if there's no suitable
6997  * key in the Relation relation.
6998  */
6999 static HeapTuple
7000 ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_changed, bool *copy)
7001 {
7002         TupleDesc       desc = RelationGetDescr(relation);
7003         Oid                     replidindex;
7004         Relation        idx_rel;
7005         TupleDesc       idx_desc;
7006         char            replident = relation->rd_rel->relreplident;
7007         HeapTuple       key_tuple = NULL;
7008         bool            nulls[MaxHeapAttributeNumber];
7009         Datum           values[MaxHeapAttributeNumber];
7010         int                     natt;
7011
7012         *copy = false;
7013
7014         if (!RelationIsLogicallyLogged(relation))
7015                 return NULL;
7016
7017         if (replident == REPLICA_IDENTITY_NOTHING)
7018                 return NULL;
7019
7020         if (replident == REPLICA_IDENTITY_FULL)
7021         {
7022                 /*
7023                  * When logging the entire old tuple, it very well could contain
7024                  * toasted columns. If so, force them to be inlined.
7025                  */
7026                 if (HeapTupleHasExternal(tp))
7027                 {
7028                         *copy = true;
7029                         tp = toast_flatten_tuple(tp, RelationGetDescr(relation));
7030                 }
7031                 return tp;
7032         }
7033
7034         /* if the key hasn't changed and we're only logging the key, we're done */
7035         if (!key_changed)
7036                 return NULL;
7037
7038         /* find the replica identity index */
7039         replidindex = RelationGetReplicaIndex(relation);
7040         if (!OidIsValid(replidindex))
7041         {
7042                 elog(DEBUG4, "could not find configured replica identity for table \"%s\"",
7043                          RelationGetRelationName(relation));
7044                 return NULL;
7045         }
7046
7047         idx_rel = RelationIdGetRelation(replidindex);
7048         idx_desc = RelationGetDescr(idx_rel);
7049
7050         /* deform tuple, so we have fast access to columns */
7051         heap_deform_tuple(tp, desc, values, nulls);
7052
7053         /* set all columns to NULL, regardless of whether they actually are */
7054         memset(nulls, 1, sizeof(nulls));
7055
7056         /*
7057          * Now set all columns contained in the index to NOT NULL, they cannot
7058          * currently be NULL.
7059          */
7060         for (natt = 0; natt < idx_desc->natts; natt++)
7061         {
7062                 int                     attno = idx_rel->rd_index->indkey.values[natt];
7063
7064                 if (attno < 0)
7065                 {
7066                         /*
7067                          * The OID column can appear in an index definition, but that's
7068                          * OK, becuse we always copy the OID if present (see below). Other
7069                          * system columns may not.
7070                          */
7071                         if (attno == ObjectIdAttributeNumber)
7072                                 continue;
7073                         elog(ERROR, "system column in index");
7074                 }
7075                 nulls[attno - 1] = false;
7076         }
7077
7078         key_tuple = heap_form_tuple(desc, values, nulls);
7079         *copy = true;
7080         RelationClose(idx_rel);
7081
7082         /*
7083          * Always copy oids if the table has them, even if not included in the
7084          * index. The space in the logged tuple is used anyway, so there's little
7085          * point in not including the information.
7086          */
7087         if (relation->rd_rel->relhasoids)
7088                 HeapTupleSetOid(key_tuple, HeapTupleGetOid(tp));
7089
7090         /*
7091          * If the tuple, which by here only contains indexed columns, still has
7092          * toasted columns, force them to be inlined. This is somewhat unlikely
7093          * since there's limits on the size of indexed columns, so we don't
7094          * duplicate toast_flatten_tuple()s functionality in the above loop over
7095          * the indexed columns, even if it would be more efficient.
7096          */
7097         if (HeapTupleHasExternal(key_tuple))
7098         {
7099                 HeapTuple       oldtup = key_tuple;
7100
7101                 key_tuple = toast_flatten_tuple(oldtup, RelationGetDescr(relation));
7102                 heap_freetuple(oldtup);
7103         }
7104
7105         return key_tuple;
7106 }
7107
7108 /*
7109  * Handles CLEANUP_INFO
7110  */
7111 static void
7112 heap_xlog_cleanup_info(XLogRecPtr lsn, XLogRecord *record)
7113 {
7114         xl_heap_cleanup_info *xlrec = (xl_heap_cleanup_info *) XLogRecGetData(record);
7115
7116         if (InHotStandby)
7117                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, xlrec->node);
7118
7119         /*
7120          * Actual operation is a no-op. Record type exists to provide a means for
7121          * conflict processing to occur before we begin index vacuum actions. see
7122          * vacuumlazy.c and also comments in btvacuumpage()
7123          */
7124
7125         /* Backup blocks are not used in cleanup_info records */
7126         Assert(!(record->xl_info & XLR_BKP_BLOCK_MASK));
7127 }
7128
7129 /*
7130  * Handles HEAP2_CLEAN record type
7131  */
7132 static void
7133 heap_xlog_clean(XLogRecPtr lsn, XLogRecord *record)
7134 {
7135         xl_heap_clean *xlrec = (xl_heap_clean *) XLogRecGetData(record);
7136         Buffer          buffer;
7137         Size            freespace = 0;
7138         RelFileNode rnode;
7139         BlockNumber blkno;
7140         XLogRedoAction action;
7141
7142         rnode = xlrec->node;
7143         blkno = xlrec->block;
7144
7145         /*
7146          * We're about to remove tuples. In Hot Standby mode, ensure that there's
7147          * no queries running for which the removed tuples are still visible.
7148          *
7149          * Not all HEAP2_CLEAN records remove tuples with xids, so we only want to
7150          * conflict on the records that cause MVCC failures for user queries. If
7151          * latestRemovedXid is invalid, skip conflict processing.
7152          */
7153         if (InHotStandby && TransactionIdIsValid(xlrec->latestRemovedXid))
7154                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, rnode);
7155
7156         /*
7157          * If we have a full-page image, restore it (using a cleanup lock) and
7158          * we're done.
7159          */
7160         action = XLogReadBufferForRedoExtended(lsn, record, 0,
7161                                                                                    rnode, MAIN_FORKNUM, blkno,
7162                                                                                    RBM_NORMAL, true, &buffer);
7163         if (action == BLK_NEEDS_REDO)
7164         {
7165                 Page            page = (Page) BufferGetPage(buffer);
7166                 OffsetNumber *end;
7167                 OffsetNumber *redirected;
7168                 OffsetNumber *nowdead;
7169                 OffsetNumber *nowunused;
7170                 int                     nredirected;
7171                 int                     ndead;
7172                 int                     nunused;
7173
7174                 nredirected = xlrec->nredirected;
7175                 ndead = xlrec->ndead;
7176                 end = (OffsetNumber *) ((char *) xlrec + record->xl_len);
7177                 redirected = (OffsetNumber *) ((char *) xlrec + SizeOfHeapClean);
7178                 nowdead = redirected + (nredirected * 2);
7179                 nowunused = nowdead + ndead;
7180                 nunused = (end - nowunused);
7181                 Assert(nunused >= 0);
7182
7183                 /* Update all item pointers per the record, and repair fragmentation */
7184                 heap_page_prune_execute(buffer,
7185                                                                 redirected, nredirected,
7186                                                                 nowdead, ndead,
7187                                                                 nowunused, nunused);
7188
7189                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7190
7191                 /*
7192                  * Note: we don't worry about updating the page's prunability hints.
7193                  * At worst this will cause an extra prune cycle to occur soon.
7194                  */
7195
7196                 PageSetLSN(page, lsn);
7197                 MarkBufferDirty(buffer);
7198         }
7199         if (BufferIsValid(buffer))
7200                 UnlockReleaseBuffer(buffer);
7201
7202         /*
7203          * Update the FSM as well.
7204          *
7205          * XXX: Don't do this if the page was restored from full page image. We
7206          * don't bother to update the FSM in that case, it doesn't need to be
7207          * totally accurate anyway.
7208          */
7209         if (action == BLK_NEEDS_REDO)
7210                 XLogRecordPageWithFreeSpace(xlrec->node, xlrec->block, freespace);
7211 }
7212
7213 /*
7214  * Replay XLOG_HEAP2_VISIBLE record.
7215  *
7216  * The critical integrity requirement here is that we must never end up with
7217  * a situation where the visibility map bit is set, and the page-level
7218  * PD_ALL_VISIBLE bit is clear.  If that were to occur, then a subsequent
7219  * page modification would fail to clear the visibility map bit.
7220  */
7221 static void
7222 heap_xlog_visible(XLogRecPtr lsn, XLogRecord *record)
7223 {
7224         xl_heap_visible *xlrec = (xl_heap_visible *) XLogRecGetData(record);
7225         Buffer          buffer;
7226         Page            page;
7227         RelFileNode rnode;
7228         BlockNumber blkno;
7229         XLogRedoAction action;
7230
7231         rnode = xlrec->node;
7232         blkno = xlrec->block;
7233
7234         /*
7235          * If there are any Hot Standby transactions running that have an xmin
7236          * horizon old enough that this page isn't all-visible for them, they
7237          * might incorrectly decide that an index-only scan can skip a heap fetch.
7238          *
7239          * NB: It might be better to throw some kind of "soft" conflict here that
7240          * forces any index-only scan that is in flight to perform heap fetches,
7241          * rather than killing the transaction outright.
7242          */
7243         if (InHotStandby)
7244                 ResolveRecoveryConflictWithSnapshot(xlrec->cutoff_xid, rnode);
7245
7246         /*
7247          * Read the heap page, if it still exists. If the heap file has dropped or
7248          * truncated later in recovery, we don't need to update the page, but we'd
7249          * better still update the visibility map.
7250          */
7251         action = XLogReadBufferForRedo(lsn, record, 1, rnode, blkno, &buffer);
7252         if (action == BLK_NEEDS_REDO)
7253         {
7254                 /*
7255                  * We don't bump the LSN of the heap page when setting the visibility
7256                  * map bit (unless checksums are enabled, in which case we must),
7257                  * because that would generate an unworkable volume of full-page
7258                  * writes.  This exposes us to torn page hazards, but since we're not
7259                  * inspecting the existing page contents in any way, we don't care.
7260                  *
7261                  * However, all operations that clear the visibility map bit *do* bump
7262                  * the LSN, and those operations will only be replayed if the XLOG LSN
7263                  * follows the page LSN.  Thus, if the page LSN has advanced past our
7264                  * XLOG record's LSN, we mustn't mark the page all-visible, because
7265                  * the subsequent update won't be replayed to clear the flag.
7266                  */
7267                 page = BufferGetPage(buffer);
7268                 PageSetAllVisible(page);
7269                 MarkBufferDirty(buffer);
7270         }
7271         else if (action == BLK_RESTORED)
7272         {
7273                 /*
7274                  * If heap block was backed up, restore it. This can only happen with
7275                  * checksums enabled.
7276                  */
7277                 Assert(DataChecksumsEnabled());
7278         }
7279         if (BufferIsValid(buffer))
7280                 UnlockReleaseBuffer(buffer);
7281
7282         /*
7283          * Even if we skipped the heap page update due to the LSN interlock, it's
7284          * still safe to update the visibility map.  Any WAL record that clears
7285          * the visibility map bit does so before checking the page LSN, so any
7286          * bits that need to be cleared will still be cleared.
7287          */
7288         if (record->xl_info & XLR_BKP_BLOCK(0))
7289                 (void) RestoreBackupBlock(lsn, record, 0, false, false);
7290         else
7291         {
7292                 Relation        reln;
7293                 Buffer          vmbuffer = InvalidBuffer;
7294
7295                 reln = CreateFakeRelcacheEntry(rnode);
7296                 visibilitymap_pin(reln, blkno, &vmbuffer);
7297
7298                 /*
7299                  * Don't set the bit if replay has already passed this point.
7300                  *
7301                  * It might be safe to do this unconditionally; if replay has passed
7302                  * this point, we'll replay at least as far this time as we did
7303                  * before, and if this bit needs to be cleared, the record responsible
7304                  * for doing so should be again replayed, and clear it.  For right
7305                  * now, out of an abundance of conservatism, we use the same test here
7306                  * we did for the heap page.  If this results in a dropped bit, no
7307                  * real harm is done; and the next VACUUM will fix it.
7308                  */
7309                 if (lsn > PageGetLSN(BufferGetPage(vmbuffer)))
7310                         visibilitymap_set(reln, blkno, InvalidBuffer, lsn, vmbuffer,
7311                                                           xlrec->cutoff_xid);
7312
7313                 ReleaseBuffer(vmbuffer);
7314                 FreeFakeRelcacheEntry(reln);
7315         }
7316 }
7317
7318 /*
7319  * Replay XLOG_HEAP2_FREEZE_PAGE records
7320  */
7321 static void
7322 heap_xlog_freeze_page(XLogRecPtr lsn, XLogRecord *record)
7323 {
7324         xl_heap_freeze_page *xlrec = (xl_heap_freeze_page *) XLogRecGetData(record);
7325         TransactionId cutoff_xid = xlrec->cutoff_xid;
7326         Buffer          buffer;
7327         Page            page;
7328         int                     ntup;
7329
7330         /*
7331          * In Hot Standby mode, ensure that there's no queries running which still
7332          * consider the frozen xids as running.
7333          */
7334         if (InHotStandby)
7335                 ResolveRecoveryConflictWithSnapshot(cutoff_xid, xlrec->node);
7336
7337         if (XLogReadBufferForRedo(lsn, record, 0, xlrec->node, xlrec->block,
7338                                                           &buffer) == BLK_NEEDS_REDO)
7339         {
7340                 page = BufferGetPage(buffer);
7341
7342                 /* now execute freeze plan for each frozen tuple */
7343                 for (ntup = 0; ntup < xlrec->ntuples; ntup++)
7344                 {
7345                         xl_heap_freeze_tuple *xlrec_tp;
7346                         ItemId          lp;
7347                         HeapTupleHeader tuple;
7348
7349                         xlrec_tp = &xlrec->tuples[ntup];
7350                         lp = PageGetItemId(page, xlrec_tp->offset); /* offsets are one-based */
7351                         tuple = (HeapTupleHeader) PageGetItem(page, lp);
7352
7353                         heap_execute_freeze_tuple(tuple, xlrec_tp);
7354                 }
7355
7356                 PageSetLSN(page, lsn);
7357                 MarkBufferDirty(buffer);
7358         }
7359         if (BufferIsValid(buffer))
7360                 UnlockReleaseBuffer(buffer);
7361 }
7362
7363 /*
7364  * Given an "infobits" field from an XLog record, set the correct bits in the
7365  * given infomask and infomask2 for the tuple touched by the record.
7366  *
7367  * (This is the reverse of compute_infobits).
7368  */
7369 static void
7370 fix_infomask_from_infobits(uint8 infobits, uint16 *infomask, uint16 *infomask2)
7371 {
7372         *infomask &= ~(HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY |
7373                                    HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_EXCL_LOCK);
7374         *infomask2 &= ~HEAP_KEYS_UPDATED;
7375
7376         if (infobits & XLHL_XMAX_IS_MULTI)
7377                 *infomask |= HEAP_XMAX_IS_MULTI;
7378         if (infobits & XLHL_XMAX_LOCK_ONLY)
7379                 *infomask |= HEAP_XMAX_LOCK_ONLY;
7380         if (infobits & XLHL_XMAX_EXCL_LOCK)
7381                 *infomask |= HEAP_XMAX_EXCL_LOCK;
7382         /* note HEAP_XMAX_SHR_LOCK isn't considered here */
7383         if (infobits & XLHL_XMAX_KEYSHR_LOCK)
7384                 *infomask |= HEAP_XMAX_KEYSHR_LOCK;
7385
7386         if (infobits & XLHL_KEYS_UPDATED)
7387                 *infomask2 |= HEAP_KEYS_UPDATED;
7388 }
7389
7390 static void
7391 heap_xlog_delete(XLogRecPtr lsn, XLogRecord *record)
7392 {
7393         xl_heap_delete *xlrec = (xl_heap_delete *) XLogRecGetData(record);
7394         Buffer          buffer;
7395         Page            page;
7396         OffsetNumber offnum;
7397         ItemId          lp = NULL;
7398         HeapTupleHeader htup;
7399         BlockNumber blkno;
7400         RelFileNode target_node;
7401
7402         blkno = ItemPointerGetBlockNumber(&(xlrec->target.tid));
7403         target_node = xlrec->target.node;
7404
7405         /*
7406          * The visibility map may need to be fixed even if the heap page is
7407          * already up-to-date.
7408          */
7409         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7410         {
7411                 Relation        reln = CreateFakeRelcacheEntry(target_node);
7412                 Buffer          vmbuffer = InvalidBuffer;
7413
7414                 visibilitymap_pin(reln, blkno, &vmbuffer);
7415                 visibilitymap_clear(reln, blkno, vmbuffer);
7416                 ReleaseBuffer(vmbuffer);
7417                 FreeFakeRelcacheEntry(reln);
7418         }
7419
7420         if (XLogReadBufferForRedo(lsn, record, 0, target_node, blkno, &buffer)
7421                 == BLK_NEEDS_REDO)
7422         {
7423                 page = (Page) BufferGetPage(buffer);
7424
7425                 offnum = ItemPointerGetOffsetNumber(&(xlrec->target.tid));
7426                 if (PageGetMaxOffsetNumber(page) >= offnum)
7427                         lp = PageGetItemId(page, offnum);
7428
7429                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
7430                         elog(PANIC, "heap_delete_redo: invalid lp");
7431
7432                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7433
7434                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
7435                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7436                 HeapTupleHeaderClearHotUpdated(htup);
7437                 fix_infomask_from_infobits(xlrec->infobits_set,
7438                                                                    &htup->t_infomask, &htup->t_infomask2);
7439                 HeapTupleHeaderSetXmax(htup, xlrec->xmax);
7440                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
7441
7442                 /* Mark the page as a candidate for pruning */
7443                 PageSetPrunable(page, record->xl_xid);
7444
7445                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7446                         PageClearAllVisible(page);
7447
7448                 /* Make sure there is no forward chain link in t_ctid */
7449                 htup->t_ctid = xlrec->target.tid;
7450                 PageSetLSN(page, lsn);
7451                 MarkBufferDirty(buffer);
7452         }
7453         if (BufferIsValid(buffer))
7454                 UnlockReleaseBuffer(buffer);
7455 }
7456
7457 static void
7458 heap_xlog_insert(XLogRecPtr lsn, XLogRecord *record)
7459 {
7460         xl_heap_insert *xlrec = (xl_heap_insert *) XLogRecGetData(record);
7461         Buffer          buffer;
7462         Page            page;
7463         OffsetNumber offnum;
7464         struct
7465         {
7466                 HeapTupleHeaderData hdr;
7467                 char            data[MaxHeapTupleSize];
7468         }                       tbuf;
7469         HeapTupleHeader htup;
7470         xl_heap_header xlhdr;
7471         uint32          newlen;
7472         Size            freespace = 0;
7473         RelFileNode target_node;
7474         BlockNumber blkno;
7475         XLogRedoAction action;
7476
7477         target_node = xlrec->target.node;
7478         blkno = ItemPointerGetBlockNumber(&(xlrec->target.tid));
7479
7480         /*
7481          * The visibility map may need to be fixed even if the heap page is
7482          * already up-to-date.
7483          */
7484         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7485         {
7486                 Relation        reln = CreateFakeRelcacheEntry(target_node);
7487                 Buffer          vmbuffer = InvalidBuffer;
7488
7489                 visibilitymap_pin(reln, blkno, &vmbuffer);
7490                 visibilitymap_clear(reln, blkno, vmbuffer);
7491                 ReleaseBuffer(vmbuffer);
7492                 FreeFakeRelcacheEntry(reln);
7493         }
7494
7495         /*
7496          * If we inserted the first and only tuple on the page, re-initialize
7497          * the page from scratch.
7498          */
7499         if (record->xl_info & XLOG_HEAP_INIT_PAGE)
7500         {
7501                 XLogReadBufferForRedoExtended(lsn, record, 0,
7502                                                                           target_node, MAIN_FORKNUM, blkno,
7503                                                                           RBM_ZERO, false, &buffer);
7504                 page = BufferGetPage(buffer);
7505                 PageInit(page, BufferGetPageSize(buffer), 0);
7506                 action = BLK_NEEDS_REDO;
7507         }
7508         else
7509                 action = XLogReadBufferForRedo(lsn, record, 0, target_node, blkno,
7510                                                                            &buffer);
7511
7512         if (action == BLK_NEEDS_REDO)
7513         {
7514                 page = BufferGetPage(buffer);
7515
7516                 offnum = ItemPointerGetOffsetNumber(&(xlrec->target.tid));
7517                 if (PageGetMaxOffsetNumber(page) + 1 < offnum)
7518                         elog(PANIC, "heap_insert_redo: invalid max offset number");
7519
7520                 newlen = record->xl_len - SizeOfHeapInsert - SizeOfHeapHeader;
7521                 Assert(newlen <= MaxHeapTupleSize);
7522                 memcpy((char *) &xlhdr,
7523                            (char *) xlrec + SizeOfHeapInsert,
7524                            SizeOfHeapHeader);
7525                 htup = &tbuf.hdr;
7526                 MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData));
7527                 /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
7528                 memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits),
7529                            (char *) xlrec + SizeOfHeapInsert + SizeOfHeapHeader,
7530                            newlen);
7531                 newlen += offsetof(HeapTupleHeaderData, t_bits);
7532                 htup->t_infomask2 = xlhdr.t_infomask2;
7533                 htup->t_infomask = xlhdr.t_infomask;
7534                 htup->t_hoff = xlhdr.t_hoff;
7535                 HeapTupleHeaderSetXmin(htup, record->xl_xid);
7536                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
7537                 htup->t_ctid = xlrec->target.tid;
7538
7539                 offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
7540                 if (offnum == InvalidOffsetNumber)
7541                         elog(PANIC, "heap_insert_redo: failed to add tuple");
7542
7543                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7544
7545                 PageSetLSN(page, lsn);
7546
7547                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7548                         PageClearAllVisible(page);
7549
7550                 MarkBufferDirty(buffer);
7551         }
7552         if (BufferIsValid(buffer))
7553                 UnlockReleaseBuffer(buffer);
7554
7555         /*
7556          * If the page is running low on free space, update the FSM as well.
7557          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
7558          * better than that without knowing the fill-factor for the table.
7559          *
7560          * XXX: Don't do this if the page was restored from full page image. We
7561          * don't bother to update the FSM in that case, it doesn't need to be
7562          * totally accurate anyway.
7563          */
7564         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
7565                 XLogRecordPageWithFreeSpace(xlrec->target.node, blkno, freespace);
7566 }
7567
7568 /*
7569  * Handles MULTI_INSERT record type.
7570  */
7571 static void
7572 heap_xlog_multi_insert(XLogRecPtr lsn, XLogRecord *record)
7573 {
7574         char       *recdata = XLogRecGetData(record);
7575         xl_heap_multi_insert *xlrec;
7576         RelFileNode rnode;
7577         BlockNumber blkno;
7578         Buffer          buffer;
7579         Page            page;
7580         struct
7581         {
7582                 HeapTupleHeaderData hdr;
7583                 char            data[MaxHeapTupleSize];
7584         }                       tbuf;
7585         HeapTupleHeader htup;
7586         uint32          newlen;
7587         Size            freespace = 0;
7588         int                     i;
7589         bool            isinit = (record->xl_info & XLOG_HEAP_INIT_PAGE) != 0;
7590         XLogRedoAction action;
7591
7592         /*
7593          * Insertion doesn't overwrite MVCC data, so no conflict processing is
7594          * required.
7595          */
7596
7597         xlrec = (xl_heap_multi_insert *) recdata;
7598         recdata += SizeOfHeapMultiInsert;
7599
7600         rnode = xlrec->node;
7601         blkno = xlrec->blkno;
7602
7603         /*
7604          * If we're reinitializing the page, the tuples are stored in order from
7605          * FirstOffsetNumber. Otherwise there's an array of offsets in the WAL
7606          * record.
7607          */
7608         if (!isinit)
7609                 recdata += sizeof(OffsetNumber) * xlrec->ntuples;
7610
7611         /*
7612          * The visibility map may need to be fixed even if the heap page is
7613          * already up-to-date.
7614          */
7615         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7616         {
7617                 Relation        reln = CreateFakeRelcacheEntry(rnode);
7618                 Buffer          vmbuffer = InvalidBuffer;
7619
7620                 visibilitymap_pin(reln, blkno, &vmbuffer);
7621                 visibilitymap_clear(reln, blkno, vmbuffer);
7622                 ReleaseBuffer(vmbuffer);
7623                 FreeFakeRelcacheEntry(reln);
7624         }
7625
7626         if (isinit)
7627         {
7628                 XLogReadBufferForRedoExtended(lsn, record, 0,
7629                                                                           rnode, MAIN_FORKNUM, blkno,
7630                                                                           RBM_ZERO, false, &buffer);
7631                 page = BufferGetPage(buffer);
7632                 PageInit(page, BufferGetPageSize(buffer), 0);
7633                 action = BLK_NEEDS_REDO;
7634         }
7635         else
7636                 action = XLogReadBufferForRedo(lsn, record, 0, rnode, blkno, &buffer);
7637
7638         if (action == BLK_NEEDS_REDO)
7639         {
7640                 page = BufferGetPage(buffer);
7641                 for (i = 0; i < xlrec->ntuples; i++)
7642                 {
7643                         OffsetNumber offnum;
7644                         xl_multi_insert_tuple *xlhdr;
7645
7646                         if (isinit)
7647                                 offnum = FirstOffsetNumber + i;
7648                         else
7649                                 offnum = xlrec->offsets[i];
7650                         if (PageGetMaxOffsetNumber(page) + 1 < offnum)
7651                                 elog(PANIC, "heap_multi_insert_redo: invalid max offset number");
7652
7653                         xlhdr = (xl_multi_insert_tuple *) SHORTALIGN(recdata);
7654                         recdata = ((char *) xlhdr) + SizeOfMultiInsertTuple;
7655
7656                         newlen = xlhdr->datalen;
7657                         Assert(newlen <= MaxHeapTupleSize);
7658                         htup = &tbuf.hdr;
7659                         MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData));
7660                         /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
7661                         memcpy((char *) htup + offsetof(HeapTupleHeaderData, t_bits),
7662                                    (char *) recdata,
7663                                    newlen);
7664                         recdata += newlen;
7665
7666                         newlen += offsetof(HeapTupleHeaderData, t_bits);
7667                         htup->t_infomask2 = xlhdr->t_infomask2;
7668                         htup->t_infomask = xlhdr->t_infomask;
7669                         htup->t_hoff = xlhdr->t_hoff;
7670                         HeapTupleHeaderSetXmin(htup, record->xl_xid);
7671                         HeapTupleHeaderSetCmin(htup, FirstCommandId);
7672                         ItemPointerSetBlockNumber(&htup->t_ctid, blkno);
7673                         ItemPointerSetOffsetNumber(&htup->t_ctid, offnum);
7674
7675                         offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
7676                         if (offnum == InvalidOffsetNumber)
7677                                 elog(PANIC, "heap_multi_insert_redo: failed to add tuple");
7678                 }
7679
7680                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7681
7682                 PageSetLSN(page, lsn);
7683
7684                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7685                         PageClearAllVisible(page);
7686
7687                 MarkBufferDirty(buffer);
7688         }
7689         if (BufferIsValid(buffer))
7690                 UnlockReleaseBuffer(buffer);
7691
7692         /*
7693          * If the page is running low on free space, update the FSM as well.
7694          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
7695          * better than that without knowing the fill-factor for the table.
7696          *
7697          * XXX: Don't do this if the page was restored from full page image. We
7698          * don't bother to update the FSM in that case, it doesn't need to be
7699          * totally accurate anyway.
7700          */
7701         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
7702                 XLogRecordPageWithFreeSpace(xlrec->node, blkno, freespace);
7703 }
7704
7705 /*
7706  * Handles UPDATE and HOT_UPDATE
7707  */
7708 static void
7709 heap_xlog_update(XLogRecPtr lsn, XLogRecord *record, bool hot_update)
7710 {
7711         xl_heap_update *xlrec = (xl_heap_update *) XLogRecGetData(record);
7712         RelFileNode rnode;
7713         BlockNumber oldblk;
7714         BlockNumber newblk;
7715         Buffer          obuffer,
7716                                 nbuffer;
7717         Page            page;
7718         OffsetNumber offnum;
7719         ItemId          lp = NULL;
7720         HeapTupleData oldtup;
7721         HeapTupleHeader htup;
7722         char       *recdata;
7723         uint16          prefixlen = 0,
7724                                 suffixlen = 0;
7725         char       *newp;
7726         struct
7727         {
7728                 HeapTupleHeaderData hdr;
7729                 char            data[MaxHeapTupleSize];
7730         }                       tbuf;
7731         xl_heap_header_len xlhdr;
7732         uint32          newlen;
7733         Size            freespace = 0;
7734         XLogRedoAction oldaction;
7735         XLogRedoAction newaction;
7736
7737         /* initialize to keep the compiler quiet */
7738         oldtup.t_data = NULL;
7739         oldtup.t_len = 0;
7740
7741         rnode = xlrec->target.node;
7742         newblk = ItemPointerGetBlockNumber(&xlrec->newtid);
7743         oldblk = ItemPointerGetBlockNumber(&xlrec->target.tid);
7744
7745         /*
7746          * The visibility map may need to be fixed even if the heap page is
7747          * already up-to-date.
7748          */
7749         if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7750         {
7751                 Relation        reln = CreateFakeRelcacheEntry(rnode);
7752                 Buffer          vmbuffer = InvalidBuffer;
7753
7754                 visibilitymap_pin(reln, oldblk, &vmbuffer);
7755                 visibilitymap_clear(reln, oldblk, vmbuffer);
7756                 ReleaseBuffer(vmbuffer);
7757                 FreeFakeRelcacheEntry(reln);
7758         }
7759
7760         /*
7761          * In normal operation, it is important to lock the two pages in
7762          * page-number order, to avoid possible deadlocks against other update
7763          * operations going the other way.  However, during WAL replay there can
7764          * be no other update happening, so we don't need to worry about that. But
7765          * we *do* need to worry that we don't expose an inconsistent state to Hot
7766          * Standby queries --- so the original page can't be unlocked before we've
7767          * added the new tuple to the new page.
7768          */
7769
7770         /* Deal with old tuple version */
7771         oldaction = XLogReadBufferForRedo(lsn, record, 0, rnode, oldblk, &obuffer);
7772         if (oldaction == BLK_NEEDS_REDO)
7773         {
7774                 page = (Page) BufferGetPage(obuffer);
7775
7776                 offnum = ItemPointerGetOffsetNumber(&(xlrec->target.tid));
7777                 if (PageGetMaxOffsetNumber(page) >= offnum)
7778                         lp = PageGetItemId(page, offnum);
7779
7780                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
7781                         elog(PANIC, "heap_update_redo: invalid lp");
7782
7783                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7784
7785                 oldtup.t_data = htup;
7786                 oldtup.t_len = ItemIdGetLength(lp);
7787
7788                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
7789                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
7790                 if (hot_update)
7791                         HeapTupleHeaderSetHotUpdated(htup);
7792                 else
7793                         HeapTupleHeaderClearHotUpdated(htup);
7794                 fix_infomask_from_infobits(xlrec->old_infobits_set, &htup->t_infomask,
7795                                                                    &htup->t_infomask2);
7796                 HeapTupleHeaderSetXmax(htup, xlrec->old_xmax);
7797                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
7798                 /* Set forward chain link in t_ctid */
7799                 htup->t_ctid = xlrec->newtid;
7800
7801                 /* Mark the page as a candidate for pruning */
7802                 PageSetPrunable(page, record->xl_xid);
7803
7804                 if (xlrec->flags & XLOG_HEAP_ALL_VISIBLE_CLEARED)
7805                         PageClearAllVisible(page);
7806
7807                 PageSetLSN(page, lsn);
7808                 MarkBufferDirty(obuffer);
7809         }
7810
7811         /*
7812          * Read the page the new tuple goes into, if different from old.
7813          */
7814         if (oldblk == newblk)
7815         {
7816                 nbuffer = obuffer;
7817                 newaction = oldaction;
7818         }
7819         else if (record->xl_info & XLOG_HEAP_INIT_PAGE)
7820         {
7821                 XLogReadBufferForRedoExtended(lsn, record, 1,
7822                                                                           rnode, MAIN_FORKNUM, newblk,
7823                                                                           RBM_ZERO, false, &nbuffer);
7824                 page = (Page) BufferGetPage(nbuffer);
7825                 PageInit(page, BufferGetPageSize(nbuffer), 0);
7826                 newaction = BLK_NEEDS_REDO;
7827         }
7828         else
7829                 newaction = XLogReadBufferForRedo(lsn, record, 1, rnode, newblk,
7830                                                                                   &nbuffer);
7831
7832         /*
7833          * The visibility map may need to be fixed even if the heap page is
7834          * already up-to-date.
7835          */
7836         if (xlrec->flags & XLOG_HEAP_NEW_ALL_VISIBLE_CLEARED)
7837         {
7838                 Relation        reln = CreateFakeRelcacheEntry(xlrec->target.node);
7839                 Buffer          vmbuffer = InvalidBuffer;
7840
7841                 visibilitymap_pin(reln, newblk, &vmbuffer);
7842                 visibilitymap_clear(reln, newblk, vmbuffer);
7843                 ReleaseBuffer(vmbuffer);
7844                 FreeFakeRelcacheEntry(reln);
7845         }
7846
7847         /* Deal with new tuple */
7848         if (newaction == BLK_NEEDS_REDO)
7849         {
7850                 page = (Page) BufferGetPage(nbuffer);
7851
7852                 offnum = ItemPointerGetOffsetNumber(&(xlrec->newtid));
7853                 if (PageGetMaxOffsetNumber(page) + 1 < offnum)
7854                         elog(PANIC, "heap_update_redo: invalid max offset number");
7855
7856                 recdata = (char *) xlrec + SizeOfHeapUpdate;
7857
7858                 if (xlrec->flags & XLOG_HEAP_PREFIX_FROM_OLD)
7859                 {
7860                         Assert(newblk == oldblk);
7861                         memcpy(&prefixlen, recdata, sizeof(uint16));
7862                         recdata += sizeof(uint16);
7863                 }
7864                 if (xlrec->flags & XLOG_HEAP_SUFFIX_FROM_OLD)
7865                 {
7866                         Assert(newblk == oldblk);
7867                         memcpy(&suffixlen, recdata, sizeof(uint16));
7868                         recdata += sizeof(uint16);
7869                 }
7870
7871                 memcpy((char *) &xlhdr, recdata, SizeOfHeapHeaderLen);
7872                 recdata += SizeOfHeapHeaderLen;
7873
7874                 Assert(xlhdr.t_len + prefixlen + suffixlen <= MaxHeapTupleSize);
7875                 htup = &tbuf.hdr;
7876                 MemSet((char *) htup, 0, sizeof(HeapTupleHeaderData));
7877
7878                 /*
7879                  * Reconstruct the new tuple using the prefix and/or suffix from the
7880                  * old tuple, and the data stored in the WAL record.
7881                  */
7882                 newp = (char *) htup + offsetof(HeapTupleHeaderData, t_bits);
7883                 if (prefixlen > 0)
7884                 {
7885                         int                     len;
7886
7887                         /* copy bitmap [+ padding] [+ oid] from WAL record */
7888                         len = xlhdr.header.t_hoff - offsetof(HeapTupleHeaderData, t_bits);
7889                         memcpy(newp, recdata, len);
7890                         recdata += len;
7891                         newp += len;
7892
7893                         /* copy prefix from old tuple */
7894                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_data->t_hoff, prefixlen);
7895                         newp += prefixlen;
7896
7897                         /* copy new tuple data from WAL record */
7898                         len = xlhdr.t_len - (xlhdr.header.t_hoff - offsetof(HeapTupleHeaderData, t_bits));
7899                         memcpy(newp, recdata, len);
7900                         recdata += len;
7901                         newp += len;
7902                 }
7903                 else
7904                 {
7905                         /*
7906                          * copy bitmap [+ padding] [+ oid] + data from record, all in one
7907                          * go
7908                          */
7909                         memcpy(newp, recdata, xlhdr.t_len);
7910                         recdata += xlhdr.t_len;
7911                         newp += xlhdr.t_len;
7912                 }
7913                 /* copy suffix from old tuple */
7914                 if (suffixlen > 0)
7915                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_len - suffixlen, suffixlen);
7916
7917                 newlen = offsetof(HeapTupleHeaderData, t_bits) + xlhdr.t_len + prefixlen + suffixlen;
7918                 htup->t_infomask2 = xlhdr.header.t_infomask2;
7919                 htup->t_infomask = xlhdr.header.t_infomask;
7920                 htup->t_hoff = xlhdr.header.t_hoff;
7921
7922                 HeapTupleHeaderSetXmin(htup, record->xl_xid);
7923                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
7924                 HeapTupleHeaderSetXmax(htup, xlrec->new_xmax);
7925                 /* Make sure there is no forward chain link in t_ctid */
7926                 htup->t_ctid = xlrec->newtid;
7927
7928                 offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
7929                 if (offnum == InvalidOffsetNumber)
7930                         elog(PANIC, "heap_update_redo: failed to add tuple");
7931
7932                 if (xlrec->flags & XLOG_HEAP_NEW_ALL_VISIBLE_CLEARED)
7933                         PageClearAllVisible(page);
7934
7935                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7936
7937                 PageSetLSN(page, lsn);
7938                 MarkBufferDirty(nbuffer);
7939         }
7940         if (BufferIsValid(nbuffer) && nbuffer != obuffer)
7941                 UnlockReleaseBuffer(nbuffer);
7942         if (BufferIsValid(obuffer))
7943                 UnlockReleaseBuffer(obuffer);
7944
7945         /*
7946          * If the new page is running low on free space, update the FSM as well.
7947          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
7948          * better than that without knowing the fill-factor for the table.
7949          *
7950          * However, don't update the FSM on HOT updates, because after crash
7951          * recovery, either the old or the new tuple will certainly be dead and
7952          * prunable. After pruning, the page will have roughly as much free space
7953          * as it did before the update, assuming the new tuple is about the same
7954          * size as the old one.
7955          *
7956          * XXX: Don't do this if the page was restored from full page image. We
7957          * don't bother to update the FSM in that case, it doesn't need to be
7958          * totally accurate anyway.
7959          */
7960         if (newaction == BLK_NEEDS_REDO && !hot_update && freespace < BLCKSZ / 5)
7961                 XLogRecordPageWithFreeSpace(xlrec->target.node,
7962                                                                  ItemPointerGetBlockNumber(&(xlrec->newtid)),
7963                                                                         freespace);
7964 }
7965
7966 static void
7967 heap_xlog_lock(XLogRecPtr lsn, XLogRecord *record)
7968 {
7969         xl_heap_lock *xlrec = (xl_heap_lock *) XLogRecGetData(record);
7970         Buffer          buffer;
7971         Page            page;
7972         OffsetNumber offnum;
7973         ItemId          lp = NULL;
7974         HeapTupleHeader htup;
7975
7976         if (XLogReadBufferForRedo(lsn, record, 0, xlrec->target.node,
7977                                                           ItemPointerGetBlockNumber(&xlrec->target.tid),
7978                                                           &buffer) == BLK_NEEDS_REDO)
7979         {
7980                 page = (Page) BufferGetPage(buffer);
7981
7982                 offnum = ItemPointerGetOffsetNumber(&(xlrec->target.tid));
7983                 if (PageGetMaxOffsetNumber(page) >= offnum)
7984                         lp = PageGetItemId(page, offnum);
7985
7986                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
7987                         elog(PANIC, "heap_lock_redo: invalid lp");
7988
7989                 htup = (HeapTupleHeader) PageGetItem(page, lp);
7990
7991                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
7992                                                                    &htup->t_infomask2);
7993
7994                 /*
7995                  * Clear relevant update flags, but only if the modified infomask says
7996                  * there's no update.
7997                  */
7998                 if (HEAP_XMAX_IS_LOCKED_ONLY(htup->t_infomask))
7999                 {
8000                         HeapTupleHeaderClearHotUpdated(htup);
8001                         /* Make sure there is no forward chain link in t_ctid */
8002                         htup->t_ctid = xlrec->target.tid;
8003                 }
8004                 HeapTupleHeaderSetXmax(htup, xlrec->locking_xid);
8005                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
8006                 PageSetLSN(page, lsn);
8007                 MarkBufferDirty(buffer);
8008         }
8009         if (BufferIsValid(buffer))
8010                 UnlockReleaseBuffer(buffer);
8011 }
8012
8013 static void
8014 heap_xlog_lock_updated(XLogRecPtr lsn, XLogRecord *record)
8015 {
8016         xl_heap_lock_updated *xlrec =
8017         (xl_heap_lock_updated *) XLogRecGetData(record);
8018         Buffer          buffer;
8019         Page            page;
8020         OffsetNumber offnum;
8021         ItemId          lp = NULL;
8022         HeapTupleHeader htup;
8023
8024         if (XLogReadBufferForRedo(lsn, record, 0, xlrec->target.node,
8025                                                           ItemPointerGetBlockNumber(&(xlrec->target.tid)),
8026                                                           &buffer) == BLK_NEEDS_REDO)
8027         {
8028                 page = BufferGetPage(buffer);
8029                 offnum = ItemPointerGetOffsetNumber(&(xlrec->target.tid));
8030                 if (PageGetMaxOffsetNumber(page) >= offnum)
8031                         lp = PageGetItemId(page, offnum);
8032
8033                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8034                         elog(PANIC, "heap_xlog_lock_updated: invalid lp");
8035
8036                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8037
8038                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
8039                                                                    &htup->t_infomask2);
8040                 HeapTupleHeaderSetXmax(htup, xlrec->xmax);
8041
8042                 PageSetLSN(page, lsn);
8043                 MarkBufferDirty(buffer);
8044         }
8045         if (BufferIsValid(buffer))
8046                 UnlockReleaseBuffer(buffer);
8047 }
8048
8049 static void
8050 heap_xlog_inplace(XLogRecPtr lsn, XLogRecord *record)
8051 {
8052         xl_heap_inplace *xlrec = (xl_heap_inplace *) XLogRecGetData(record);
8053         Buffer          buffer;
8054         Page            page;
8055         OffsetNumber offnum;
8056         ItemId          lp = NULL;
8057         HeapTupleHeader htup;
8058         uint32          oldlen;
8059         uint32          newlen;
8060
8061         if (XLogReadBufferForRedo(lsn, record, 0, xlrec->target.node,
8062                                                           ItemPointerGetBlockNumber(&(xlrec->target.tid)),
8063                                                           &buffer) == BLK_NEEDS_REDO)
8064         {
8065                 page = BufferGetPage(buffer);
8066
8067                 offnum = ItemPointerGetOffsetNumber(&(xlrec->target.tid));
8068                 if (PageGetMaxOffsetNumber(page) >= offnum)
8069                         lp = PageGetItemId(page, offnum);
8070
8071                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8072                         elog(PANIC, "heap_inplace_redo: invalid lp");
8073
8074                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8075
8076                 oldlen = ItemIdGetLength(lp) - htup->t_hoff;
8077                 newlen = record->xl_len - SizeOfHeapInplace;
8078                 if (oldlen != newlen)
8079                         elog(PANIC, "heap_inplace_redo: wrong tuple length");
8080
8081                 memcpy((char *) htup + htup->t_hoff,
8082                            (char *) xlrec + SizeOfHeapInplace,
8083                            newlen);
8084
8085                 PageSetLSN(page, lsn);
8086                 MarkBufferDirty(buffer);
8087         }
8088         if (BufferIsValid(buffer))
8089                 UnlockReleaseBuffer(buffer);
8090 }
8091
8092 void
8093 heap_redo(XLogRecPtr lsn, XLogRecord *record)
8094 {
8095         uint8           info = record->xl_info & ~XLR_INFO_MASK;
8096
8097         /*
8098          * These operations don't overwrite MVCC data so no conflict processing is
8099          * required. The ones in heap2 rmgr do.
8100          */
8101
8102         switch (info & XLOG_HEAP_OPMASK)
8103         {
8104                 case XLOG_HEAP_INSERT:
8105                         heap_xlog_insert(lsn, record);
8106                         break;
8107                 case XLOG_HEAP_DELETE:
8108                         heap_xlog_delete(lsn, record);
8109                         break;
8110                 case XLOG_HEAP_UPDATE:
8111                         heap_xlog_update(lsn, record, false);
8112                         break;
8113                 case XLOG_HEAP_HOT_UPDATE:
8114                         heap_xlog_update(lsn, record, true);
8115                         break;
8116                 case XLOG_HEAP_LOCK:
8117                         heap_xlog_lock(lsn, record);
8118                         break;
8119                 case XLOG_HEAP_INPLACE:
8120                         heap_xlog_inplace(lsn, record);
8121                         break;
8122                 default:
8123                         elog(PANIC, "heap_redo: unknown op code %u", info);
8124         }
8125 }
8126
8127 void
8128 heap2_redo(XLogRecPtr lsn, XLogRecord *record)
8129 {
8130         uint8           info = record->xl_info & ~XLR_INFO_MASK;
8131
8132         switch (info & XLOG_HEAP_OPMASK)
8133         {
8134                 case XLOG_HEAP2_CLEAN:
8135                         heap_xlog_clean(lsn, record);
8136                         break;
8137                 case XLOG_HEAP2_FREEZE_PAGE:
8138                         heap_xlog_freeze_page(lsn, record);
8139                         break;
8140                 case XLOG_HEAP2_CLEANUP_INFO:
8141                         heap_xlog_cleanup_info(lsn, record);
8142                         break;
8143                 case XLOG_HEAP2_VISIBLE:
8144                         heap_xlog_visible(lsn, record);
8145                         break;
8146                 case XLOG_HEAP2_MULTI_INSERT:
8147                         heap_xlog_multi_insert(lsn, record);
8148                         break;
8149                 case XLOG_HEAP2_LOCK_UPDATED:
8150                         heap_xlog_lock_updated(lsn, record);
8151                         break;
8152                 case XLOG_HEAP2_NEW_CID:
8153
8154                         /*
8155                          * Nothing to do on a real replay, only used during logical
8156                          * decoding.
8157                          */
8158                         break;
8159                 case XLOG_HEAP2_REWRITE:
8160                         heap_xlog_logical_rewrite(lsn, record);
8161                         break;
8162                 default:
8163                         elog(PANIC, "heap2_redo: unknown op code %u", info);
8164         }
8165 }
8166
8167 /*
8168  *      heap_sync               - sync a heap, for use when no WAL has been written
8169  *
8170  * This forces the heap contents (including TOAST heap if any) down to disk.
8171  * If we skipped using WAL, and WAL is otherwise needed, we must force the
8172  * relation down to disk before it's safe to commit the transaction.  This
8173  * requires writing out any dirty buffers and then doing a forced fsync.
8174  *
8175  * Indexes are not touched.  (Currently, index operations associated with
8176  * the commands that use this are WAL-logged and so do not need fsync.
8177  * That behavior might change someday, but in any case it's likely that
8178  * any fsync decisions required would be per-index and hence not appropriate
8179  * to be done here.)
8180  */
8181 void
8182 heap_sync(Relation rel)
8183 {
8184         /* non-WAL-logged tables never need fsync */
8185         if (!RelationNeedsWAL(rel))
8186                 return;
8187
8188         /* main heap */
8189         FlushRelationBuffers(rel);
8190         /* FlushRelationBuffers will have opened rd_smgr */
8191         smgrimmedsync(rel->rd_smgr, MAIN_FORKNUM);
8192
8193         /* FSM is not critical, don't bother syncing it */
8194
8195         /* toast heap, if any */
8196         if (OidIsValid(rel->rd_rel->reltoastrelid))
8197         {
8198                 Relation        toastrel;
8199
8200                 toastrel = heap_open(rel->rd_rel->reltoastrelid, AccessShareLock);
8201                 FlushRelationBuffers(toastrel);
8202                 smgrimmedsync(toastrel->rd_smgr, MAIN_FORKNUM);
8203                 heap_close(toastrel, AccessShareLock);
8204         }
8205 }