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) 1996-2016, 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_sync               - sync heap, for when no WAL has been written
  31  *
  32  * NOTES
  33  *        This file contains the heap_ routines which implement
  34  *        the POSTGRES heap access method used for all POSTGRES
  35  *        relations.
  36  *
  37  *-------------------------------------------------------------------------
  38  */
  39 #include "postgres.h"
  40
  41 #include "access/heapam.h"
  42 #include "access/heapam_xlog.h"
  43 #include "access/hio.h"
  44 #include "access/multixact.h"
  45 #include "access/parallel.h"
  46 #include "access/relscan.h"
  47 #include "access/sysattr.h"
  48 #include "access/transam.h"
  49 #include "access/tuptoaster.h"
  50 #include "access/valid.h"
  51 #include "access/visibilitymap.h"
  52 #include "access/xact.h"
  53 #include "access/xlog.h"
  54 #include "access/xloginsert.h"
  55 #include "access/xlogutils.h"
  56 #include "catalog/catalog.h"
  57 #include "catalog/namespace.h"
  58 #include "miscadmin.h"
  59 #include "pgstat.h"
  60 #include "storage/bufmgr.h"
  61 #include "storage/freespace.h"
  62 #include "storage/lmgr.h"
  63 #include "storage/predicate.h"
  64 #include "storage/procarray.h"
  65 #include "storage/smgr.h"
  66 #include "storage/spin.h"
  67 #include "storage/standby.h"
  68 #include "utils/datum.h"
  69 #include "utils/inval.h"
  70 #include "utils/lsyscache.h"
  71 #include "utils/relcache.h"
  72 #include "utils/snapmgr.h"
  73 #include "utils/syscache.h"
  74 #include "utils/tqual.h"
  75
  76
  77 /* GUC variable */
  78 bool            synchronize_seqscans = true;
  79
  80
  81 static HeapScanDesc heap_beginscan_internal(Relation relation,
  82                                                 Snapshot snapshot,
  83                                                 int nkeys, ScanKey key,
  84                                                 ParallelHeapScanDesc parallel_scan,
  85                                                 bool allow_strat,
  86                                                 bool allow_sync,
  87                                                 bool allow_pagemode,
  88                                                 bool is_bitmapscan,
  89                                                 bool is_samplescan,
  90                                                 bool temp_snap);
  91 static BlockNumber heap_parallelscan_nextpage(HeapScanDesc scan);
  92 static HeapTuple heap_prepare_insert(Relation relation, HeapTuple tup,
  93                                         TransactionId xid, CommandId cid, int options);
  94 static XLogRecPtr log_heap_update(Relation reln, Buffer oldbuf,
  95                                 Buffer newbuf, HeapTuple oldtup,
  96                                 HeapTuple newtup, HeapTuple old_key_tup,
  97                                 bool all_visible_cleared, bool new_all_visible_cleared);
  98 static void HeapSatisfiesHOTandKeyUpdate(Relation relation,
  99                                                          Bitmapset *hot_attrs,
 100                                                          Bitmapset *key_attrs, Bitmapset *id_attrs,
 101                                                          bool *satisfies_hot, bool *satisfies_key,
 102                                                          bool *satisfies_id,
 103                                                          HeapTuple oldtup, HeapTuple newtup);
 104 static bool heap_acquire_tuplock(Relation relation, ItemPointer tid,
 105                                          LockTupleMode mode, LockWaitPolicy wait_policy,
 106                                          bool *have_tuple_lock);
 107 static void compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
 108                                                   uint16 old_infomask2, TransactionId add_to_xmax,
 109                                                   LockTupleMode mode, bool is_update,
 110                                                   TransactionId *result_xmax, uint16 *result_infomask,
 111                                                   uint16 *result_infomask2);
 112 static HTSU_Result heap_lock_updated_tuple(Relation rel, HeapTuple tuple,
 113                                                 ItemPointer ctid, TransactionId xid,
 114                                                 LockTupleMode mode);
 115 static void GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
 116                                            uint16 *new_infomask2);
 117 static TransactionId MultiXactIdGetUpdateXid(TransactionId xmax,
 118                                                 uint16 t_infomask);
 119 static bool DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
 120                                                 LockTupleMode lockmode);
 121 static void MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
 122                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
 123                                 int *remaining);
 124 static bool ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
 125                                                    uint16 infomask, Relation rel, int *remaining);
 126 static XLogRecPtr log_heap_new_cid(Relation relation, HeapTuple tup);
 127 static HeapTuple ExtractReplicaIdentity(Relation rel, HeapTuple tup, bool key_modified,
 128                                            bool *copy);
 129
 130
 131 /*
 132  * Each tuple lock mode has a corresponding heavyweight lock, and one or two
 133  * corresponding MultiXactStatuses (one to merely lock tuples, another one to
 134  * update them).  This table (and the macros below) helps us determine the
 135  * heavyweight lock mode and MultiXactStatus values to use for any particular
 136  * tuple lock strength.
 137  *
 138  * Don't look at lockstatus/updstatus directly!  Use get_mxact_status_for_lock
 139  * instead.
 140  */
 141 static const struct
 142 {
 143         LOCKMODE        hwlock;
 144         int                     lockstatus;
 145         int                     updstatus;
 146 }
 147
 148                         tupleLockExtraInfo[MaxLockTupleMode + 1] =
 149 {
 150         {                                                       /* LockTupleKeyShare */
 151                 AccessShareLock,
 152                 MultiXactStatusForKeyShare,
 153                 -1                                              /* KeyShare does not allow updating tuples */
 154         },
 155         {                                                       /* LockTupleShare */
 156                 RowShareLock,
 157                 MultiXactStatusForShare,
 158                 -1                                              /* Share does not allow updating tuples */
 159         },
 160         {                                                       /* LockTupleNoKeyExclusive */
 161                 ExclusiveLock,
 162                 MultiXactStatusForNoKeyUpdate,
 163                 MultiXactStatusNoKeyUpdate
 164         },
 165         {                                                       /* LockTupleExclusive */
 166                 AccessExclusiveLock,
 167                 MultiXactStatusForUpdate,
 168                 MultiXactStatusUpdate
 169         }
 170 };
 171
 172 /* Get the LOCKMODE for a given MultiXactStatus */
 173 #define LOCKMODE_from_mxstatus(status) \
 174                         (tupleLockExtraInfo[TUPLOCK_from_mxstatus((status))].hwlock)
 175
 176 /*
 177  * Acquire heavyweight locks on tuples, using a LockTupleMode strength value.
 178  * This is more readable than having every caller translate it to lock.h's
 179  * LOCKMODE.
 180  */
 181 #define LockTupleTuplock(rel, tup, mode) \
 182         LockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 183 #define UnlockTupleTuplock(rel, tup, mode) \
 184         UnlockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 185 #define ConditionalLockTupleTuplock(rel, tup, mode) \
 186         ConditionalLockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 187
 188 /*
 189  * This table maps tuple lock strength values for each particular
 190  * MultiXactStatus value.
 191  */
 192 static const int MultiXactStatusLock[MaxMultiXactStatus + 1] =
 193 {
 194         LockTupleKeyShare,                      /* ForKeyShare */
 195         LockTupleShare,                         /* ForShare */
 196         LockTupleNoKeyExclusive,        /* ForNoKeyUpdate */
 197         LockTupleExclusive,                     /* ForUpdate */
 198         LockTupleNoKeyExclusive,        /* NoKeyUpdate */
 199         LockTupleExclusive                      /* Update */
 200 };
 201
 202 /* Get the LockTupleMode for a given MultiXactStatus */
 203 #define TUPLOCK_from_mxstatus(status) \
 204                         (MultiXactStatusLock[(status)])
 205
 206 /* ----------------------------------------------------------------
 207  *                                               heap support routines
 208  * ----------------------------------------------------------------
 209  */
 210
 211 /* ----------------
 212  *              initscan - scan code common to heap_beginscan and heap_rescan
 213  * ----------------
 214  */
 215 static void
 216 initscan(HeapScanDesc scan, ScanKey key, bool keep_startblock)
 217 {
 218         bool            allow_strat;
 219         bool            allow_sync;
 220
 221         /*
 222          * Determine the number of blocks we have to scan.
 223          *
 224          * It is sufficient to do this once at scan start, since any tuples added
 225          * while the scan is in progress will be invisible to my snapshot anyway.
 226          * (That is not true when using a non-MVCC snapshot.  However, we couldn't
 227          * guarantee to return tuples added after scan start anyway, since they
 228          * might go into pages we already scanned.  To guarantee consistent
 229          * results for a non-MVCC snapshot, the caller must hold some higher-level
 230          * lock that ensures the interesting tuple(s) won't change.)
 231          */
 232         if (scan->rs_parallel != NULL)
 233                 scan->rs_nblocks = scan->rs_parallel->phs_nblocks;
 234         else
 235                 scan->rs_nblocks = RelationGetNumberOfBlocks(scan->rs_rd);
 236
 237         /*
 238          * If the table is large relative to NBuffers, use a bulk-read access
 239          * strategy and enable synchronized scanning (see syncscan.c).  Although
 240          * the thresholds for these features could be different, we make them the
 241          * same so that there are only two behaviors to tune rather than four.
 242          * (However, some callers need to be able to disable one or both of these
 243          * behaviors, independently of the size of the table; also there is a GUC
 244          * variable that can disable synchronized scanning.)
 245          *
 246          * Note that heap_parallelscan_initialize has a very similar test; if you
 247          * change this, consider changing that one, too.
 248          */
 249         if (!RelationUsesLocalBuffers(scan->rs_rd) &&
 250                 scan->rs_nblocks > NBuffers / 4)
 251         {
 252                 allow_strat = scan->rs_allow_strat;
 253                 allow_sync = scan->rs_allow_sync;
 254         }
 255         else
 256                 allow_strat = allow_sync = false;
 257
 258         if (allow_strat)
 259         {
 260                 /* During a rescan, keep the previous strategy object. */
 261                 if (scan->rs_strategy == NULL)
 262                         scan->rs_strategy = GetAccessStrategy(BAS_BULKREAD);
 263         }
 264         else
 265         {
 266                 if (scan->rs_strategy != NULL)
 267                         FreeAccessStrategy(scan->rs_strategy);
 268                 scan->rs_strategy = NULL;
 269         }
 270
 271         if (scan->rs_parallel != NULL)
 272         {
 273                 /* For parallel scan, believe whatever ParallelHeapScanDesc says. */
 274                 scan->rs_syncscan = scan->rs_parallel->phs_syncscan;
 275         }
 276         else if (keep_startblock)
 277         {
 278                 /*
 279                  * When rescanning, we want to keep the previous startblock setting,
 280                  * so that rewinding a cursor doesn't generate surprising results.
 281                  * Reset the active syncscan setting, though.
 282                  */
 283                 scan->rs_syncscan = (allow_sync && synchronize_seqscans);
 284         }
 285         else if (allow_sync && synchronize_seqscans)
 286         {
 287                 scan->rs_syncscan = true;
 288                 scan->rs_startblock = ss_get_location(scan->rs_rd, scan->rs_nblocks);
 289         }
 290         else
 291         {
 292                 scan->rs_syncscan = false;
 293                 scan->rs_startblock = 0;
 294         }
 295
 296         scan->rs_numblocks = InvalidBlockNumber;
 297         scan->rs_inited = false;
 298         scan->rs_ctup.t_data = NULL;
 299         ItemPointerSetInvalid(&scan->rs_ctup.t_self);
 300         scan->rs_cbuf = InvalidBuffer;
 301         scan->rs_cblock = InvalidBlockNumber;
 302
 303         /* page-at-a-time fields are always invalid when not rs_inited */
 304
 305         /*
 306          * copy the scan key, if appropriate
 307          */
 308         if (key != NULL)
 309                 memcpy(scan->rs_key, key, scan->rs_nkeys * sizeof(ScanKeyData));
 310
 311         /*
 312          * Currently, we don't have a stats counter for bitmap heap scans (but the
 313          * underlying bitmap index scans will be counted) or sample scans (we only
 314          * update stats for tuple fetches there)
 315          */
 316         if (!scan->rs_bitmapscan && !scan->rs_samplescan)
 317                 pgstat_count_heap_scan(scan->rs_rd);
 318 }
 319
 320 /*
 321  * heap_setscanlimits - restrict range of a heapscan
 322  *
 323  * startBlk is the page to start at
 324  * numBlks is number of pages to scan (InvalidBlockNumber means "all")
 325  */
 326 void
 327 heap_setscanlimits(HeapScanDesc scan, BlockNumber startBlk, BlockNumber numBlks)
 328 {
 329         Assert(!scan->rs_inited);       /* else too late to change */
 330         Assert(!scan->rs_syncscan); /* else rs_startblock is significant */
 331
 332         /* Check startBlk is valid (but allow case of zero blocks...) */
 333         Assert(startBlk == 0 || startBlk < scan->rs_nblocks);
 334
 335         scan->rs_startblock = startBlk;
 336         scan->rs_numblocks = numBlks;
 337 }
 338
 339 /*
 340  * heapgetpage - subroutine for heapgettup()
 341  *
 342  * This routine reads and pins the specified page of the relation.
 343  * In page-at-a-time mode it performs additional work, namely determining
 344  * which tuples on the page are visible.
 345  */
 346 void
 347 heapgetpage(HeapScanDesc scan, BlockNumber page)
 348 {
 349         Buffer          buffer;
 350         Snapshot        snapshot;
 351         Page            dp;
 352         int                     lines;
 353         int                     ntup;
 354         OffsetNumber lineoff;
 355         ItemId          lpp;
 356         bool            all_visible;
 357
 358         Assert(page < scan->rs_nblocks);
 359
 360         /* release previous scan buffer, if any */
 361         if (BufferIsValid(scan->rs_cbuf))
 362         {
 363                 ReleaseBuffer(scan->rs_cbuf);
 364                 scan->rs_cbuf = InvalidBuffer;
 365         }
 366
 367         /*
 368          * Be sure to check for interrupts at least once per page.  Checks at
 369          * higher code levels won't be able to stop a seqscan that encounters many
 370          * pages' worth of consecutive dead tuples.
 371          */
 372         CHECK_FOR_INTERRUPTS();
 373
 374         /* read page using selected strategy */
 375         scan->rs_cbuf = ReadBufferExtended(scan->rs_rd, MAIN_FORKNUM, page,
 376                                                                            RBM_NORMAL, scan->rs_strategy);
 377         scan->rs_cblock = page;
 378
 379         if (!scan->rs_pageatatime)
 380                 return;
 381
 382         buffer = scan->rs_cbuf;
 383         snapshot = scan->rs_snapshot;
 384
 385         /*
 386          * Prune and repair fragmentation for the whole page, if possible.
 387          */
 388         heap_page_prune_opt(scan->rs_rd, buffer);
 389
 390         /*
 391          * We must hold share lock on the buffer content while examining tuple
 392          * visibility.  Afterwards, however, the tuples we have found to be
 393          * visible are guaranteed good as long as we hold the buffer pin.
 394          */
 395         LockBuffer(buffer, BUFFER_LOCK_SHARE);
 396
 397         dp = (Page) BufferGetPage(buffer);
 398         lines = PageGetMaxOffsetNumber(dp);
 399         ntup = 0;
 400
 401         /*
 402          * If the all-visible flag indicates that all tuples on the page are
 403          * visible to everyone, we can skip the per-tuple visibility tests.
 404          *
 405          * Note: In hot standby, a tuple that's already visible to all
 406          * transactions in the master might still be invisible to a read-only
 407          * transaction in the standby. We partly handle this problem by tracking
 408          * the minimum xmin of visible tuples as the cut-off XID while marking a
 409          * page all-visible on master and WAL log that along with the visibility
 410          * map SET operation. In hot standby, we wait for (or abort) all
 411          * transactions that can potentially may not see one or more tuples on the
 412          * page. That's how index-only scans work fine in hot standby. A crucial
 413          * difference between index-only scans and heap scans is that the
 414          * index-only scan completely relies on the visibility map where as heap
 415          * scan looks at the page-level PD_ALL_VISIBLE flag. We are not sure if
 416          * the page-level flag can be trusted in the same way, because it might
 417          * get propagated somehow without being explicitly WAL-logged, e.g. via a
 418          * full page write. Until we can prove that beyond doubt, let's check each
 419          * tuple for visibility the hard way.
 420          */
 421         all_visible = PageIsAllVisible(dp) && !snapshot->takenDuringRecovery;
 422
 423         for (lineoff = FirstOffsetNumber, lpp = PageGetItemId(dp, lineoff);
 424                  lineoff <= lines;
 425                  lineoff++, lpp++)
 426         {
 427                 if (ItemIdIsNormal(lpp))
 428                 {
 429                         HeapTupleData loctup;
 430                         bool            valid;
 431
 432                         loctup.t_tableOid = RelationGetRelid(scan->rs_rd);
 433                         loctup.t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 434                         loctup.t_len = ItemIdGetLength(lpp);
 435                         ItemPointerSet(&(loctup.t_self), page, lineoff);
 436
 437                         if (all_visible)
 438                                 valid = true;
 439                         else
 440                                 valid = HeapTupleSatisfiesVisibility(&loctup, snapshot, buffer);
 441
 442                         CheckForSerializableConflictOut(valid, scan->rs_rd, &loctup,
 443                                                                                         buffer, snapshot);
 444
 445                         if (valid)
 446                                 scan->rs_vistuples[ntup++] = lineoff;
 447                 }
 448         }
 449
 450         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
 451
 452         Assert(ntup <= MaxHeapTuplesPerPage);
 453         scan->rs_ntuples = ntup;
 454 }
 455
 456 /* ----------------
 457  *              heapgettup - fetch next heap tuple
 458  *
 459  *              Initialize the scan if not already done; then advance to the next
 460  *              tuple as indicated by "dir"; return the next tuple in scan->rs_ctup,
 461  *              or set scan->rs_ctup.t_data = NULL if no more tuples.
 462  *
 463  * dir == NoMovementScanDirection means "re-fetch the tuple indicated
 464  * by scan->rs_ctup".
 465  *
 466  * Note: the reason nkeys/key are passed separately, even though they are
 467  * kept in the scan descriptor, is that the caller may not want us to check
 468  * the scankeys.
 469  *
 470  * Note: when we fall off the end of the scan in either direction, we
 471  * reset rs_inited.  This means that a further request with the same
 472  * scan direction will restart the scan, which is a bit odd, but a
 473  * request with the opposite scan direction will start a fresh scan
 474  * in the proper direction.  The latter is required behavior for cursors,
 475  * while the former case is generally undefined behavior in Postgres
 476  * so we don't care too much.
 477  * ----------------
 478  */
 479 static void
 480 heapgettup(HeapScanDesc scan,
 481                    ScanDirection dir,
 482                    int nkeys,
 483                    ScanKey key)
 484 {
 485         HeapTuple       tuple = &(scan->rs_ctup);
 486         Snapshot        snapshot = scan->rs_snapshot;
 487         bool            backward = ScanDirectionIsBackward(dir);
 488         BlockNumber page;
 489         bool            finished;
 490         Page            dp;
 491         int                     lines;
 492         OffsetNumber lineoff;
 493         int                     linesleft;
 494         ItemId          lpp;
 495
 496         /*
 497          * calculate next starting lineoff, given scan direction
 498          */
 499         if (ScanDirectionIsForward(dir))
 500         {
 501                 if (!scan->rs_inited)
 502                 {
 503                         /*
 504                          * return null immediately if relation is empty
 505                          */
 506                         if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
 507                         {
 508                                 Assert(!BufferIsValid(scan->rs_cbuf));
 509                                 tuple->t_data = NULL;
 510                                 return;
 511                         }
 512                         if (scan->rs_parallel != NULL)
 513                         {
 514                                 page = heap_parallelscan_nextpage(scan);
 515
 516                                 /* Other processes might have already finished the scan. */
 517                                 if (page == InvalidBlockNumber)
 518                                 {
 519                                         Assert(!BufferIsValid(scan->rs_cbuf));
 520                                         tuple->t_data = NULL;
 521                                         return;
 522                                 }
 523                         }
 524                         else
 525                                 page = scan->rs_startblock;             /* first page */
 526                         heapgetpage(scan, page);
 527                         lineoff = FirstOffsetNumber;            /* first offnum */
 528                         scan->rs_inited = true;
 529                 }
 530                 else
 531                 {
 532                         /* continue from previously returned page/tuple */
 533                         page = scan->rs_cblock;         /* current page */
 534                         lineoff =                       /* next offnum */
 535                                 OffsetNumberNext(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 536                 }
 537
 538                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 539
 540                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 541                 lines = PageGetMaxOffsetNumber(dp);
 542                 /* page and lineoff now reference the physically next tid */
 543
 544                 linesleft = lines - lineoff + 1;
 545         }
 546         else if (backward)
 547         {
 548                 /* backward parallel scan not supported */
 549                 Assert(scan->rs_parallel == NULL);
 550
 551                 if (!scan->rs_inited)
 552                 {
 553                         /*
 554                          * return null immediately if relation is empty
 555                          */
 556                         if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
 557                         {
 558                                 Assert(!BufferIsValid(scan->rs_cbuf));
 559                                 tuple->t_data = NULL;
 560                                 return;
 561                         }
 562
 563                         /*
 564                          * Disable reporting to syncscan logic in a backwards scan; it's
 565                          * not very likely anyone else is doing the same thing at the same
 566                          * time, and much more likely that we'll just bollix things for
 567                          * forward scanners.
 568                          */
 569                         scan->rs_syncscan = false;
 570                         /* start from last page of the scan */
 571                         if (scan->rs_startblock > 0)
 572                                 page = scan->rs_startblock - 1;
 573                         else
 574                                 page = scan->rs_nblocks - 1;
 575                         heapgetpage(scan, page);
 576                 }
 577                 else
 578                 {
 579                         /* continue from previously returned page/tuple */
 580                         page = scan->rs_cblock;         /* current page */
 581                 }
 582
 583                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 584
 585                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 586                 lines = PageGetMaxOffsetNumber(dp);
 587
 588                 if (!scan->rs_inited)
 589                 {
 590                         lineoff = lines;        /* final offnum */
 591                         scan->rs_inited = true;
 592                 }
 593                 else
 594                 {
 595                         lineoff =                       /* previous offnum */
 596                                 OffsetNumberPrev(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 597                 }
 598                 /* page and lineoff now reference the physically previous tid */
 599
 600                 linesleft = lineoff;
 601         }
 602         else
 603         {
 604                 /*
 605                  * ``no movement'' scan direction: refetch prior tuple
 606                  */
 607                 if (!scan->rs_inited)
 608                 {
 609                         Assert(!BufferIsValid(scan->rs_cbuf));
 610                         tuple->t_data = NULL;
 611                         return;
 612                 }
 613
 614                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 615                 if (page != scan->rs_cblock)
 616                         heapgetpage(scan, page);
 617
 618                 /* Since the tuple was previously fetched, needn't lock page here */
 619                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 620                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 621                 lpp = PageGetItemId(dp, lineoff);
 622                 Assert(ItemIdIsNormal(lpp));
 623
 624                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 625                 tuple->t_len = ItemIdGetLength(lpp);
 626
 627                 return;
 628         }
 629
 630         /*
 631          * advance the scan until we find a qualifying tuple or run out of stuff
 632          * to scan
 633          */
 634         lpp = PageGetItemId(dp, lineoff);
 635         for (;;)
 636         {
 637                 while (linesleft > 0)
 638                 {
 639                         if (ItemIdIsNormal(lpp))
 640                         {
 641                                 bool            valid;
 642
 643                                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 644                                 tuple->t_len = ItemIdGetLength(lpp);
 645                                 ItemPointerSet(&(tuple->t_self), page, lineoff);
 646
 647                                 /*
 648                                  * if current tuple qualifies, return it.
 649                                  */
 650                                 valid = HeapTupleSatisfiesVisibility(tuple,
 651                                                                                                          snapshot,
 652                                                                                                          scan->rs_cbuf);
 653
 654                                 CheckForSerializableConflictOut(valid, scan->rs_rd, tuple,
 655                                                                                                 scan->rs_cbuf, snapshot);
 656
 657                                 if (valid && key != NULL)
 658                                         HeapKeyTest(tuple, RelationGetDescr(scan->rs_rd),
 659                                                                 nkeys, key, valid);
 660
 661                                 if (valid)
 662                                 {
 663                                         LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 664                                         return;
 665                                 }
 666                         }
 667
 668                         /*
 669                          * otherwise move to the next item on the page
 670                          */
 671                         --linesleft;
 672                         if (backward)
 673                         {
 674                                 --lpp;                  /* move back in this page's ItemId array */
 675                                 --lineoff;
 676                         }
 677                         else
 678                         {
 679                                 ++lpp;                  /* move forward in this page's ItemId array */
 680                                 ++lineoff;
 681                         }
 682                 }
 683
 684                 /*
 685                  * if we get here, it means we've exhausted the items on this page and
 686                  * it's time to move to the next.
 687                  */
 688                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 689
 690                 /*
 691                  * advance to next/prior page and detect end of scan
 692                  */
 693                 if (backward)
 694                 {
 695                         finished = (page == scan->rs_startblock) ||
 696                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
 697                         if (page == 0)
 698                                 page = scan->rs_nblocks;
 699                         page--;
 700                 }
 701                 else if (scan->rs_parallel != NULL)
 702                 {
 703                         page = heap_parallelscan_nextpage(scan);
 704                         finished = (page == InvalidBlockNumber);
 705                 }
 706                 else
 707                 {
 708                         page++;
 709                         if (page >= scan->rs_nblocks)
 710                                 page = 0;
 711                         finished = (page == scan->rs_startblock) ||
 712                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
 713
 714                         /*
 715                          * Report our new scan position for synchronization purposes. We
 716                          * don't do that when moving backwards, however. That would just
 717                          * mess up any other forward-moving scanners.
 718                          *
 719                          * Note: we do this before checking for end of scan so that the
 720                          * final state of the position hint is back at the start of the
 721                          * rel.  That's not strictly necessary, but otherwise when you run
 722                          * the same query multiple times the starting position would shift
 723                          * a little bit backwards on every invocation, which is confusing.
 724                          * We don't guarantee any specific ordering in general, though.
 725                          */
 726                         if (scan->rs_syncscan)
 727                                 ss_report_location(scan->rs_rd, page);
 728                 }
 729
 730                 /*
 731                  * return NULL if we've exhausted all the pages
 732                  */
 733                 if (finished)
 734                 {
 735                         if (BufferIsValid(scan->rs_cbuf))
 736                                 ReleaseBuffer(scan->rs_cbuf);
 737                         scan->rs_cbuf = InvalidBuffer;
 738                         scan->rs_cblock = InvalidBlockNumber;
 739                         tuple->t_data = NULL;
 740                         scan->rs_inited = false;
 741                         return;
 742                 }
 743
 744                 heapgetpage(scan, page);
 745
 746                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 747
 748                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 749                 lines = PageGetMaxOffsetNumber((Page) dp);
 750                 linesleft = lines;
 751                 if (backward)
 752                 {
 753                         lineoff = lines;
 754                         lpp = PageGetItemId(dp, lines);
 755                 }
 756                 else
 757                 {
 758                         lineoff = FirstOffsetNumber;
 759                         lpp = PageGetItemId(dp, FirstOffsetNumber);
 760                 }
 761         }
 762 }
 763
 764 /* ----------------
 765  *              heapgettup_pagemode - fetch next heap tuple in page-at-a-time mode
 766  *
 767  *              Same API as heapgettup, but used in page-at-a-time mode
 768  *
 769  * The internal logic is much the same as heapgettup's too, but there are some
 770  * differences: we do not take the buffer content lock (that only needs to
 771  * happen inside heapgetpage), and we iterate through just the tuples listed
 772  * in rs_vistuples[] rather than all tuples on the page.  Notice that
 773  * lineindex is 0-based, where the corresponding loop variable lineoff in
 774  * heapgettup is 1-based.
 775  * ----------------
 776  */
 777 static void
 778 heapgettup_pagemode(HeapScanDesc scan,
 779                                         ScanDirection dir,
 780                                         int nkeys,
 781                                         ScanKey key)
 782 {
 783         HeapTuple       tuple = &(scan->rs_ctup);
 784         bool            backward = ScanDirectionIsBackward(dir);
 785         BlockNumber page;
 786         bool            finished;
 787         Page            dp;
 788         int                     lines;
 789         int                     lineindex;
 790         OffsetNumber lineoff;
 791         int                     linesleft;
 792         ItemId          lpp;
 793
 794         /*
 795          * calculate next starting lineindex, given scan direction
 796          */
 797         if (ScanDirectionIsForward(dir))
 798         {
 799                 if (!scan->rs_inited)
 800                 {
 801                         /*
 802                          * return null immediately if relation is empty
 803                          */
 804                         if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
 805                         {
 806                                 Assert(!BufferIsValid(scan->rs_cbuf));
 807                                 tuple->t_data = NULL;
 808                                 return;
 809                         }
 810                         if (scan->rs_parallel != NULL)
 811                         {
 812                                 page = heap_parallelscan_nextpage(scan);
 813
 814                                 /* Other processes might have already finished the scan. */
 815                                 if (page == InvalidBlockNumber)
 816                                 {
 817                                         Assert(!BufferIsValid(scan->rs_cbuf));
 818                                         tuple->t_data = NULL;
 819                                         return;
 820                                 }
 821                         }
 822                         else
 823                                 page = scan->rs_startblock;             /* first page */
 824                         heapgetpage(scan, page);
 825                         lineindex = 0;
 826                         scan->rs_inited = true;
 827                 }
 828                 else
 829                 {
 830                         /* continue from previously returned page/tuple */
 831                         page = scan->rs_cblock;         /* current page */
 832                         lineindex = scan->rs_cindex + 1;
 833                 }
 834
 835                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 836                 lines = scan->rs_ntuples;
 837                 /* page and lineindex now reference the next visible tid */
 838
 839                 linesleft = lines - lineindex;
 840         }
 841         else if (backward)
 842         {
 843                 /* backward parallel scan not supported */
 844                 Assert(scan->rs_parallel == NULL);
 845
 846                 if (!scan->rs_inited)
 847                 {
 848                         /*
 849                          * return null immediately if relation is empty
 850                          */
 851                         if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
 852                         {
 853                                 Assert(!BufferIsValid(scan->rs_cbuf));
 854                                 tuple->t_data = NULL;
 855                                 return;
 856                         }
 857
 858                         /*
 859                          * Disable reporting to syncscan logic in a backwards scan; it's
 860                          * not very likely anyone else is doing the same thing at the same
 861                          * time, and much more likely that we'll just bollix things for
 862                          * forward scanners.
 863                          */
 864                         scan->rs_syncscan = false;
 865                         /* start from last page of the scan */
 866                         if (scan->rs_startblock > 0)
 867                                 page = scan->rs_startblock - 1;
 868                         else
 869                                 page = scan->rs_nblocks - 1;
 870                         heapgetpage(scan, page);
 871                 }
 872                 else
 873                 {
 874                         /* continue from previously returned page/tuple */
 875                         page = scan->rs_cblock;         /* current page */
 876                 }
 877
 878                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 879                 lines = scan->rs_ntuples;
 880
 881                 if (!scan->rs_inited)
 882                 {
 883                         lineindex = lines - 1;
 884                         scan->rs_inited = true;
 885                 }
 886                 else
 887                 {
 888                         lineindex = scan->rs_cindex - 1;
 889                 }
 890                 /* page and lineindex now reference the previous visible tid */
 891
 892                 linesleft = lineindex + 1;
 893         }
 894         else
 895         {
 896                 /*
 897                  * ``no movement'' scan direction: refetch prior tuple
 898                  */
 899                 if (!scan->rs_inited)
 900                 {
 901                         Assert(!BufferIsValid(scan->rs_cbuf));
 902                         tuple->t_data = NULL;
 903                         return;
 904                 }
 905
 906                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 907                 if (page != scan->rs_cblock)
 908                         heapgetpage(scan, page);
 909
 910                 /* Since the tuple was previously fetched, needn't lock page here */
 911                 dp = (Page) BufferGetPage(scan->rs_cbuf);
 912                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 913                 lpp = PageGetItemId(dp, lineoff);
 914                 Assert(ItemIdIsNormal(lpp));
 915
 916                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 917                 tuple->t_len = ItemIdGetLength(lpp);
 918
 919                 /* check that rs_cindex is in sync */
 920                 Assert(scan->rs_cindex < scan->rs_ntuples);
 921                 Assert(lineoff == scan->rs_vistuples[scan->rs_cindex]);
 922
 923                 return;
 924         }
 925
 926         /*
 927          * advance the scan until we find a qualifying tuple or run out of stuff
 928          * to scan
 929          */
 930         for (;;)
 931         {
 932                 while (linesleft > 0)
 933                 {
 934                         lineoff = scan->rs_vistuples[lineindex];
 935                         lpp = PageGetItemId(dp, lineoff);
 936                         Assert(ItemIdIsNormal(lpp));
 937
 938                         tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 939                         tuple->t_len = ItemIdGetLength(lpp);
 940                         ItemPointerSet(&(tuple->t_self), page, lineoff);
 941
 942                         /*
 943                          * if current tuple qualifies, return it.
 944                          */
 945                         if (key != NULL)
 946                         {
 947                                 bool            valid;
 948
 949                                 HeapKeyTest(tuple, RelationGetDescr(scan->rs_rd),
 950                                                         nkeys, key, valid);
 951                                 if (valid)
 952                                 {
 953                                         scan->rs_cindex = lineindex;
 954                                         return;
 955                                 }
 956                         }
 957                         else
 958                         {
 959                                 scan->rs_cindex = lineindex;
 960                                 return;
 961                         }
 962
 963                         /*
 964                          * otherwise move to the next item on the page
 965                          */
 966                         --linesleft;
 967                         if (backward)
 968                                 --lineindex;
 969                         else
 970                                 ++lineindex;
 971                 }
 972
 973                 /*
 974                  * if we get here, it means we've exhausted the items on this page and
 975                  * it's time to move to the next.
 976                  */
 977                 if (backward)
 978                 {
 979                         finished = (page == scan->rs_startblock) ||
 980                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
 981                         if (page == 0)
 982                                 page = scan->rs_nblocks;
 983                         page--;
 984                 }
 985                 else if (scan->rs_parallel != NULL)
 986                 {
 987                         page = heap_parallelscan_nextpage(scan);
 988                         finished = (page == InvalidBlockNumber);
 989                 }
 990                 else
 991                 {
 992                         page++;
 993                         if (page >= scan->rs_nblocks)
 994                                 page = 0;
 995                         finished = (page == scan->rs_startblock) ||
 996                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
 997
 998                         /*
 999                          * Report our new scan position for synchronization purposes. We
1000                          * don't do that when moving backwards, however. That would just
1001                          * mess up any other forward-moving scanners.
1002                          *
1003                          * Note: we do this before checking for end of scan so that the
1004                          * final state of the position hint is back at the start of the
1005                          * rel.  That's not strictly necessary, but otherwise when you run
1006                          * the same query multiple times the starting position would shift
1007                          * a little bit backwards on every invocation, which is confusing.
1008                          * We don't guarantee any specific ordering in general, though.
1009                          */
1010                         if (scan->rs_syncscan)
1011                                 ss_report_location(scan->rs_rd, page);
1012                 }
1013
1014                 /*
1015                  * return NULL if we've exhausted all the pages
1016                  */
1017                 if (finished)
1018                 {
1019                         if (BufferIsValid(scan->rs_cbuf))
1020                                 ReleaseBuffer(scan->rs_cbuf);
1021                         scan->rs_cbuf = InvalidBuffer;
1022                         scan->rs_cblock = InvalidBlockNumber;
1023                         tuple->t_data = NULL;
1024                         scan->rs_inited = false;
1025                         return;
1026                 }
1027
1028                 heapgetpage(scan, page);
1029
1030                 dp = (Page) BufferGetPage(scan->rs_cbuf);
1031                 lines = scan->rs_ntuples;
1032                 linesleft = lines;
1033                 if (backward)
1034                         lineindex = lines - 1;
1035                 else
1036                         lineindex = 0;
1037         }
1038 }
1039
1040
1041 #if defined(DISABLE_COMPLEX_MACRO)
1042 /*
1043  * This is formatted so oddly so that the correspondence to the macro
1044  * definition in access/htup_details.h is maintained.
1045  */
1046 Datum
1047 fastgetattr(HeapTuple tup, int attnum, TupleDesc tupleDesc,
1048                         bool *isnull)
1049 {
1050         return (
1051                         (attnum) > 0 ?
1052                         (
1053                          (*(isnull) = false),
1054                          HeapTupleNoNulls(tup) ?
1055                          (
1056                           (tupleDesc)->attrs[(attnum) - 1]->attcacheoff >= 0 ?
1057                           (
1058                            fetchatt((tupleDesc)->attrs[(attnum) - 1],
1059                                                 (char *) (tup)->t_data + (tup)->t_data->t_hoff +
1060                                                 (tupleDesc)->attrs[(attnum) - 1]->attcacheoff)
1061                            )
1062                           :
1063                           nocachegetattr((tup), (attnum), (tupleDesc))
1064                           )
1065                          :
1066                          (
1067                           att_isnull((attnum) - 1, (tup)->t_data->t_bits) ?
1068                           (
1069                            (*(isnull) = true),
1070                            (Datum) NULL
1071                            )
1072                           :
1073                           (
1074                            nocachegetattr((tup), (attnum), (tupleDesc))
1075                            )
1076                           )
1077                          )
1078                         :
1079                         (
1080                          (Datum) NULL
1081                          )
1082                 );
1083 }
1084 #endif   /* defined(DISABLE_COMPLEX_MACRO) */
1085
1086
1087 /* ----------------------------------------------------------------
1088  *                                       heap access method interface
1089  * ----------------------------------------------------------------
1090  */
1091
1092 /* ----------------
1093  *              relation_open - open any relation by relation OID
1094  *
1095  *              If lockmode is not "NoLock", the specified kind of lock is
1096  *              obtained on the relation.  (Generally, NoLock should only be
1097  *              used if the caller knows it has some appropriate lock on the
1098  *              relation already.)
1099  *
1100  *              An error is raised if the relation does not exist.
1101  *
1102  *              NB: a "relation" is anything with a pg_class entry.  The caller is
1103  *              expected to check whether the relkind is something it can handle.
1104  * ----------------
1105  */
1106 Relation
1107 relation_open(Oid relationId, LOCKMODE lockmode)
1108 {
1109         Relation        r;
1110
1111         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1112
1113         /* Get the lock before trying to open the relcache entry */
1114         if (lockmode != NoLock)
1115                 LockRelationOid(relationId, lockmode);
1116
1117         /* The relcache does all the real work... */
1118         r = RelationIdGetRelation(relationId);
1119
1120         if (!RelationIsValid(r))
1121                 elog(ERROR, "could not open relation with OID %u", relationId);
1122
1123         /* Make note that we've accessed a temporary relation */
1124         if (RelationUsesLocalBuffers(r))
1125         {
1126                 if (IsParallelWorker())
1127                         ereport(ERROR,
1128                                         (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
1129                                          errmsg("cannot access temporary tables during a parallel operation")));
1130                 MyXactAccessedTempRel = true;
1131         }
1132
1133         pgstat_initstats(r);
1134
1135         return r;
1136 }
1137
1138 /* ----------------
1139  *              try_relation_open - open any relation by relation OID
1140  *
1141  *              Same as relation_open, except return NULL instead of failing
1142  *              if the relation does not exist.
1143  * ----------------
1144  */
1145 Relation
1146 try_relation_open(Oid relationId, LOCKMODE lockmode)
1147 {
1148         Relation        r;
1149
1150         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1151
1152         /* Get the lock first */
1153         if (lockmode != NoLock)
1154                 LockRelationOid(relationId, lockmode);
1155
1156         /*
1157          * Now that we have the lock, probe to see if the relation really exists
1158          * or not.
1159          */
1160         if (!SearchSysCacheExists1(RELOID, ObjectIdGetDatum(relationId)))
1161         {
1162                 /* Release useless lock */
1163                 if (lockmode != NoLock)
1164                         UnlockRelationOid(relationId, lockmode);
1165
1166                 return NULL;
1167         }
1168
1169         /* Should be safe to do a relcache load */
1170         r = RelationIdGetRelation(relationId);
1171
1172         if (!RelationIsValid(r))
1173                 elog(ERROR, "could not open relation with OID %u", relationId);
1174
1175         /* Make note that we've accessed a temporary relation */
1176         if (RelationUsesLocalBuffers(r))
1177         {
1178                 if (IsParallelWorker())
1179                         ereport(ERROR,
1180                                         (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
1181                                          errmsg("cannot access temporary tables during a parallel operation")));
1182                 MyXactAccessedTempRel = true;
1183         }
1184
1185         pgstat_initstats(r);
1186
1187         return r;
1188 }
1189
1190 /* ----------------
1191  *              relation_openrv - open any relation specified by a RangeVar
1192  *
1193  *              Same as relation_open, but the relation is specified by a RangeVar.
1194  * ----------------
1195  */
1196 Relation
1197 relation_openrv(const RangeVar *relation, LOCKMODE lockmode)
1198 {
1199         Oid                     relOid;
1200
1201         /*
1202          * Check for shared-cache-inval messages before trying to open the
1203          * relation.  This is needed even if we already hold a lock on the
1204          * relation, because GRANT/REVOKE are executed without taking any lock on
1205          * the target relation, and we want to be sure we see current ACL
1206          * information.  We can skip this if asked for NoLock, on the assumption
1207          * that such a call is not the first one in the current command, and so we
1208          * should be reasonably up-to-date already.  (XXX this all could stand to
1209          * be redesigned, but for the moment we'll keep doing this like it's been
1210          * done historically.)
1211          */
1212         if (lockmode != NoLock)
1213                 AcceptInvalidationMessages();
1214
1215         /* Look up and lock the appropriate relation using namespace search */
1216         relOid = RangeVarGetRelid(relation, lockmode, false);
1217
1218         /* Let relation_open do the rest */
1219         return relation_open(relOid, NoLock);
1220 }
1221
1222 /* ----------------
1223  *              relation_openrv_extended - open any relation specified by a RangeVar
1224  *
1225  *              Same as relation_openrv, but with an additional missing_ok argument
1226  *              allowing a NULL return rather than an error if the relation is not
1227  *              found.  (Note that some other causes, such as permissions problems,
1228  *              will still result in an ereport.)
1229  * ----------------
1230  */
1231 Relation
1232 relation_openrv_extended(const RangeVar *relation, LOCKMODE lockmode,
1233                                                  bool missing_ok)
1234 {
1235         Oid                     relOid;
1236
1237         /*
1238          * Check for shared-cache-inval messages before trying to open the
1239          * relation.  See comments in relation_openrv().
1240          */
1241         if (lockmode != NoLock)
1242                 AcceptInvalidationMessages();
1243
1244         /* Look up and lock the appropriate relation using namespace search */
1245         relOid = RangeVarGetRelid(relation, lockmode, missing_ok);
1246
1247         /* Return NULL on not-found */
1248         if (!OidIsValid(relOid))
1249                 return NULL;
1250
1251         /* Let relation_open do the rest */
1252         return relation_open(relOid, NoLock);
1253 }
1254
1255 /* ----------------
1256  *              relation_close - close any relation
1257  *
1258  *              If lockmode is not "NoLock", we then release the specified lock.
1259  *
1260  *              Note that it is often sensible to hold a lock beyond relation_close;
1261  *              in that case, the lock is released automatically at xact end.
1262  * ----------------
1263  */
1264 void
1265 relation_close(Relation relation, LOCKMODE lockmode)
1266 {
1267         LockRelId       relid = relation->rd_lockInfo.lockRelId;
1268
1269         Assert(lockmode >= NoLock && lockmode < MAX_LOCKMODES);
1270
1271         /* The relcache does the real work... */
1272         RelationClose(relation);
1273
1274         if (lockmode != NoLock)
1275                 UnlockRelationId(&relid, lockmode);
1276 }
1277
1278
1279 /* ----------------
1280  *              heap_open - open a heap relation by relation OID
1281  *
1282  *              This is essentially relation_open plus check that the relation
1283  *              is not an index nor a composite type.  (The caller should also
1284  *              check that it's not a view or foreign table before assuming it has
1285  *              storage.)
1286  * ----------------
1287  */
1288 Relation
1289 heap_open(Oid relationId, LOCKMODE lockmode)
1290 {
1291         Relation        r;
1292
1293         r = relation_open(relationId, lockmode);
1294
1295         if (r->rd_rel->relkind == RELKIND_INDEX)
1296                 ereport(ERROR,
1297                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1298                                  errmsg("\"%s\" is an index",
1299                                                 RelationGetRelationName(r))));
1300         else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1301                 ereport(ERROR,
1302                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1303                                  errmsg("\"%s\" is a composite type",
1304                                                 RelationGetRelationName(r))));
1305
1306         return r;
1307 }
1308
1309 /* ----------------
1310  *              heap_openrv - open a heap relation specified
1311  *              by a RangeVar node
1312  *
1313  *              As above, but relation is specified by a RangeVar.
1314  * ----------------
1315  */
1316 Relation
1317 heap_openrv(const RangeVar *relation, LOCKMODE lockmode)
1318 {
1319         Relation        r;
1320
1321         r = relation_openrv(relation, lockmode);
1322
1323         if (r->rd_rel->relkind == RELKIND_INDEX)
1324                 ereport(ERROR,
1325                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1326                                  errmsg("\"%s\" is an index",
1327                                                 RelationGetRelationName(r))));
1328         else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1329                 ereport(ERROR,
1330                                 (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1331                                  errmsg("\"%s\" is a composite type",
1332                                                 RelationGetRelationName(r))));
1333
1334         return r;
1335 }
1336
1337 /* ----------------
1338  *              heap_openrv_extended - open a heap relation specified
1339  *              by a RangeVar node
1340  *
1341  *              As above, but optionally return NULL instead of failing for
1342  *              relation-not-found.
1343  * ----------------
1344  */
1345 Relation
1346 heap_openrv_extended(const RangeVar *relation, LOCKMODE lockmode,
1347                                          bool missing_ok)
1348 {
1349         Relation        r;
1350
1351         r = relation_openrv_extended(relation, lockmode, missing_ok);
1352
1353         if (r)
1354         {
1355                 if (r->rd_rel->relkind == RELKIND_INDEX)
1356                         ereport(ERROR,
1357                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1358                                          errmsg("\"%s\" is an index",
1359                                                         RelationGetRelationName(r))));
1360                 else if (r->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
1361                         ereport(ERROR,
1362                                         (errcode(ERRCODE_WRONG_OBJECT_TYPE),
1363                                          errmsg("\"%s\" is a composite type",
1364                                                         RelationGetRelationName(r))));
1365         }
1366
1367         return r;
1368 }
1369
1370
1371 /* ----------------
1372  *              heap_beginscan  - begin relation scan
1373  *
1374  * heap_beginscan is the "standard" case.
1375  *
1376  * heap_beginscan_catalog differs in setting up its own temporary snapshot.
1377  *
1378  * heap_beginscan_strat offers an extended API that lets the caller control
1379  * whether a nondefault buffer access strategy can be used, and whether
1380  * syncscan can be chosen (possibly resulting in the scan not starting from
1381  * block zero).  Both of these default to TRUE with plain heap_beginscan.
1382  *
1383  * heap_beginscan_bm is an alternative entry point for setting up a
1384  * HeapScanDesc for a bitmap heap scan.  Although that scan technology is
1385  * really quite unlike a standard seqscan, there is just enough commonality
1386  * to make it worth using the same data structure.
1387  *
1388  * heap_beginscan_sampling is an alternative entry point for setting up a
1389  * HeapScanDesc for a TABLESAMPLE scan.  As with bitmap scans, it's worth
1390  * using the same data structure although the behavior is rather different.
1391  * In addition to the options offered by heap_beginscan_strat, this call
1392  * also allows control of whether page-mode visibility checking is used.
1393  * ----------------
1394  */
1395 HeapScanDesc
1396 heap_beginscan(Relation relation, Snapshot snapshot,
1397                            int nkeys, ScanKey key)
1398 {
1399         return heap_beginscan_internal(relation, snapshot, nkeys, key, NULL,
1400                                                                    true, true, true, false, false, false);
1401 }
1402
1403 HeapScanDesc
1404 heap_beginscan_catalog(Relation relation, int nkeys, ScanKey key)
1405 {
1406         Oid                     relid = RelationGetRelid(relation);
1407         Snapshot        snapshot = RegisterSnapshot(GetCatalogSnapshot(relid));
1408
1409         return heap_beginscan_internal(relation, snapshot, nkeys, key, NULL,
1410                                                                    true, true, true, false, false, true);
1411 }
1412
1413 HeapScanDesc
1414 heap_beginscan_strat(Relation relation, Snapshot snapshot,
1415                                          int nkeys, ScanKey key,
1416                                          bool allow_strat, bool allow_sync)
1417 {
1418         return heap_beginscan_internal(relation, snapshot, nkeys, key, NULL,
1419                                                                    allow_strat, allow_sync, true,
1420                                                                    false, false, false);
1421 }
1422
1423 HeapScanDesc
1424 heap_beginscan_bm(Relation relation, Snapshot snapshot,
1425                                   int nkeys, ScanKey key)
1426 {
1427         return heap_beginscan_internal(relation, snapshot, nkeys, key, NULL,
1428                                                                    false, false, true, true, false, false);
1429 }
1430
1431 HeapScanDesc
1432 heap_beginscan_sampling(Relation relation, Snapshot snapshot,
1433                                                 int nkeys, ScanKey key,
1434                                           bool allow_strat, bool allow_sync, bool allow_pagemode)
1435 {
1436         return heap_beginscan_internal(relation, snapshot, nkeys, key, NULL,
1437                                                                    allow_strat, allow_sync, allow_pagemode,
1438                                                                    false, true, false);
1439 }
1440
1441 static HeapScanDesc
1442 heap_beginscan_internal(Relation relation, Snapshot snapshot,
1443                                                 int nkeys, ScanKey key,
1444                                                 ParallelHeapScanDesc parallel_scan,
1445                                                 bool allow_strat,
1446                                                 bool allow_sync,
1447                                                 bool allow_pagemode,
1448                                                 bool is_bitmapscan,
1449                                                 bool is_samplescan,
1450                                                 bool temp_snap)
1451 {
1452         HeapScanDesc scan;
1453
1454         /*
1455          * increment relation ref count while scanning relation
1456          *
1457          * This is just to make really sure the relcache entry won't go away while
1458          * the scan has a pointer to it.  Caller should be holding the rel open
1459          * anyway, so this is redundant in all normal scenarios...
1460          */
1461         RelationIncrementReferenceCount(relation);
1462
1463         /*
1464          * allocate and initialize scan descriptor
1465          */
1466         scan = (HeapScanDesc) palloc(sizeof(HeapScanDescData));
1467
1468         scan->rs_rd = relation;
1469         scan->rs_snapshot = snapshot;
1470         scan->rs_nkeys = nkeys;
1471         scan->rs_bitmapscan = is_bitmapscan;
1472         scan->rs_samplescan = is_samplescan;
1473         scan->rs_strategy = NULL;       /* set in initscan */
1474         scan->rs_allow_strat = allow_strat;
1475         scan->rs_allow_sync = allow_sync;
1476         scan->rs_temp_snap = temp_snap;
1477         scan->rs_parallel = parallel_scan;
1478
1479         /*
1480          * we can use page-at-a-time mode if it's an MVCC-safe snapshot
1481          */
1482         scan->rs_pageatatime = allow_pagemode && IsMVCCSnapshot(snapshot);
1483
1484         /*
1485          * For a seqscan in a serializable transaction, acquire a predicate lock
1486          * on the entire relation. This is required not only to lock all the
1487          * matching tuples, but also to conflict with new insertions into the
1488          * table. In an indexscan, we take page locks on the index pages covering
1489          * the range specified in the scan qual, but in a heap scan there is
1490          * nothing more fine-grained to lock. A bitmap scan is a different story,
1491          * there we have already scanned the index and locked the index pages
1492          * covering the predicate. But in that case we still have to lock any
1493          * matching heap tuples.
1494          */
1495         if (!is_bitmapscan)
1496                 PredicateLockRelation(relation, snapshot);
1497
1498         /* we only need to set this up once */
1499         scan->rs_ctup.t_tableOid = RelationGetRelid(relation);
1500
1501         /*
1502          * we do this here instead of in initscan() because heap_rescan also calls
1503          * initscan() and we don't want to allocate memory again
1504          */
1505         if (nkeys > 0)
1506                 scan->rs_key = (ScanKey) palloc(sizeof(ScanKeyData) * nkeys);
1507         else
1508                 scan->rs_key = NULL;
1509
1510         initscan(scan, key, false);
1511
1512         return scan;
1513 }
1514
1515 /* ----------------
1516  *              heap_rescan             - restart a relation scan
1517  * ----------------
1518  */
1519 void
1520 heap_rescan(HeapScanDesc scan,
1521                         ScanKey key)
1522 {
1523         /*
1524          * unpin scan buffers
1525          */
1526         if (BufferIsValid(scan->rs_cbuf))
1527                 ReleaseBuffer(scan->rs_cbuf);
1528
1529         /*
1530          * reinitialize scan descriptor
1531          */
1532         initscan(scan, key, true);
1533
1534         /*
1535          * reset parallel scan, if present
1536          */
1537         if (scan->rs_parallel != NULL)
1538         {
1539                 ParallelHeapScanDesc parallel_scan;
1540
1541                 /*
1542                  * Caller is responsible for making sure that all workers have
1543                  * finished the scan before calling this, so it really shouldn't be
1544                  * necessary to acquire the mutex at all.  We acquire it anyway, just
1545                  * to be tidy.
1546                  */
1547                 parallel_scan = scan->rs_parallel;
1548                 SpinLockAcquire(&parallel_scan->phs_mutex);
1549                 parallel_scan->phs_cblock = parallel_scan->phs_startblock;
1550                 SpinLockRelease(&parallel_scan->phs_mutex);
1551         }
1552 }
1553
1554 /* ----------------
1555  *              heap_rescan_set_params  - restart a relation scan after changing params
1556  *
1557  * This call allows changing the buffer strategy, syncscan, and pagemode
1558  * options before starting a fresh scan.  Note that although the actual use
1559  * of syncscan might change (effectively, enabling or disabling reporting),
1560  * the previously selected startblock will be kept.
1561  * ----------------
1562  */
1563 void
1564 heap_rescan_set_params(HeapScanDesc scan, ScanKey key,
1565                                            bool allow_strat, bool allow_sync, bool allow_pagemode)
1566 {
1567         /* adjust parameters */
1568         scan->rs_allow_strat = allow_strat;
1569         scan->rs_allow_sync = allow_sync;
1570         scan->rs_pageatatime = allow_pagemode && IsMVCCSnapshot(scan->rs_snapshot);
1571         /* ... and rescan */
1572         heap_rescan(scan, key);
1573 }
1574
1575 /* ----------------
1576  *              heap_endscan    - end relation scan
1577  *
1578  *              See how to integrate with index scans.
1579  *              Check handling if reldesc caching.
1580  * ----------------
1581  */
1582 void
1583 heap_endscan(HeapScanDesc scan)
1584 {
1585         /* Note: no locking manipulations needed */
1586
1587         /*
1588          * unpin scan buffers
1589          */
1590         if (BufferIsValid(scan->rs_cbuf))
1591                 ReleaseBuffer(scan->rs_cbuf);
1592
1593         /*
1594          * decrement relation reference count and free scan descriptor storage
1595          */
1596         RelationDecrementReferenceCount(scan->rs_rd);
1597
1598         if (scan->rs_key)
1599                 pfree(scan->rs_key);
1600
1601         if (scan->rs_strategy != NULL)
1602                 FreeAccessStrategy(scan->rs_strategy);
1603
1604         if (scan->rs_temp_snap)
1605                 UnregisterSnapshot(scan->rs_snapshot);
1606
1607         pfree(scan);
1608 }
1609
1610 /* ----------------
1611  *              heap_parallelscan_estimate - estimate storage for ParallelHeapScanDesc
1612  *
1613  *              Sadly, this doesn't reduce to a constant, because the size required
1614  *              to serialize the snapshot can vary.
1615  * ----------------
1616  */
1617 Size
1618 heap_parallelscan_estimate(Snapshot snapshot)
1619 {
1620         return add_size(offsetof(ParallelHeapScanDescData, phs_snapshot_data),
1621                                         EstimateSnapshotSpace(snapshot));
1622 }
1623
1624 /* ----------------
1625  *              heap_parallelscan_initialize - initialize ParallelHeapScanDesc
1626  *
1627  *              Must allow as many bytes of shared memory as returned by
1628  *              heap_parallelscan_estimate.  Call this just once in the leader
1629  *              process; then, individual workers attach via heap_beginscan_parallel.
1630  * ----------------
1631  */
1632 void
1633 heap_parallelscan_initialize(ParallelHeapScanDesc target, Relation relation,
1634                                                          Snapshot snapshot)
1635 {
1636         target->phs_relid = RelationGetRelid(relation);
1637         target->phs_nblocks = RelationGetNumberOfBlocks(relation);
1638         /* compare phs_syncscan initialization to similar logic in initscan */
1639         target->phs_syncscan = synchronize_seqscans &&
1640                 !RelationUsesLocalBuffers(relation) &&
1641                 target->phs_nblocks > NBuffers / 4;
1642         SpinLockInit(&target->phs_mutex);
1643         target->phs_cblock = InvalidBlockNumber;
1644         target->phs_startblock = InvalidBlockNumber;
1645         SerializeSnapshot(snapshot, target->phs_snapshot_data);
1646 }
1647
1648 /* ----------------
1649  *              heap_beginscan_parallel - join a parallel scan
1650  *
1651  *              Caller must hold a suitable lock on the correct relation.
1652  * ----------------
1653  */
1654 HeapScanDesc
1655 heap_beginscan_parallel(Relation relation, ParallelHeapScanDesc parallel_scan)
1656 {
1657         Snapshot        snapshot;
1658
1659         Assert(RelationGetRelid(relation) == parallel_scan->phs_relid);
1660         snapshot = RestoreSnapshot(parallel_scan->phs_snapshot_data);
1661         RegisterSnapshot(snapshot);
1662
1663         return heap_beginscan_internal(relation, snapshot, 0, NULL, parallel_scan,
1664                                                                    true, true, true, false, false, true);
1665 }
1666
1667 /* ----------------
1668  *              heap_parallelscan_nextpage - get the next page to scan
1669  *
1670  *              Get the next page to scan.  Even if there are no pages left to scan,
1671  *              another backend could have grabbed a page to scan and not yet finished
1672  *              looking at it, so it doesn't follow that the scan is done when the
1673  *              first backend gets an InvalidBlockNumber return.
1674  * ----------------
1675  */
1676 static BlockNumber
1677 heap_parallelscan_nextpage(HeapScanDesc scan)
1678 {
1679         BlockNumber page = InvalidBlockNumber;
1680         BlockNumber sync_startpage = InvalidBlockNumber;
1681         BlockNumber     report_page = InvalidBlockNumber;
1682         ParallelHeapScanDesc parallel_scan;
1683
1684         Assert(scan->rs_parallel);
1685         parallel_scan = scan->rs_parallel;
1686
1687 retry:
1688         /* Grab the spinlock. */
1689         SpinLockAcquire(&parallel_scan->phs_mutex);
1690
1691         /*
1692          * If the scan's startblock has not yet been initialized, we must do so
1693          * now.  If this is not a synchronized scan, we just start at block 0, but
1694          * if it is a synchronized scan, we must get the starting position from
1695          * the synchronized scan machinery.  We can't hold the spinlock while
1696          * doing that, though, so release the spinlock, get the information we
1697          * need, and retry.  If nobody else has initialized the scan in the
1698          * meantime, we'll fill in the value we fetched on the second time
1699          * through.
1700          */
1701         if (parallel_scan->phs_startblock == InvalidBlockNumber)
1702         {
1703                 if (!parallel_scan->phs_syncscan)
1704                         parallel_scan->phs_startblock = 0;
1705                 else if (sync_startpage != InvalidBlockNumber)
1706                         parallel_scan->phs_startblock = sync_startpage;
1707                 else
1708                 {
1709                         SpinLockRelease(&parallel_scan->phs_mutex);
1710                         sync_startpage = ss_get_location(scan->rs_rd, scan->rs_nblocks);
1711                         goto retry;
1712                 }
1713                 parallel_scan->phs_cblock = parallel_scan->phs_startblock;
1714         }
1715
1716         /*
1717          * The current block number is the next one that needs to be scanned,
1718          * unless it's InvalidBlockNumber already, in which case there are no more
1719          * blocks to scan.  After remembering the current value, we must advance
1720          * it so that the next call to this function returns the next block to be
1721          * scanned.
1722          */
1723         page = parallel_scan->phs_cblock;
1724         if (page != InvalidBlockNumber)
1725         {
1726                 parallel_scan->phs_cblock++;
1727                 if (parallel_scan->phs_cblock >= scan->rs_nblocks)
1728                         parallel_scan->phs_cblock = 0;
1729                 if (parallel_scan->phs_cblock == parallel_scan->phs_startblock)
1730                 {
1731                         parallel_scan->phs_cblock = InvalidBlockNumber;
1732                         report_page = parallel_scan->phs_startblock;
1733                 }
1734         }
1735
1736         /* Release the lock. */
1737         SpinLockRelease(&parallel_scan->phs_mutex);
1738
1739         /*
1740          * Report scan location.  Normally, we report the current page number.
1741          * When we reach the end of the scan, though, we report the starting page,
1742          * not the ending page, just so the starting positions for later scans
1743          * doesn't slew backwards.  We only report the position at the end of the
1744          * scan once, though: subsequent callers will have report nothing, since
1745          * they will have page == InvalidBlockNumber.
1746          */
1747         if (scan->rs_syncscan)
1748         {
1749                 if (report_page == InvalidBlockNumber)
1750                         report_page = page;
1751                 if (report_page != InvalidBlockNumber)
1752                         ss_report_location(scan->rs_rd, report_page);
1753         }
1754
1755         return page;
1756 }
1757
1758 /* ----------------
1759  *              heap_getnext    - retrieve next tuple in scan
1760  *
1761  *              Fix to work with index relations.
1762  *              We don't return the buffer anymore, but you can get it from the
1763  *              returned HeapTuple.
1764  * ----------------
1765  */
1766
1767 #ifdef HEAPDEBUGALL
1768 #define HEAPDEBUG_1 \
1769         elog(DEBUG2, "heap_getnext([%s,nkeys=%d],dir=%d) called", \
1770                  RelationGetRelationName(scan->rs_rd), scan->rs_nkeys, (int) direction)
1771 #define HEAPDEBUG_2 \
1772         elog(DEBUG2, "heap_getnext returning EOS")
1773 #define HEAPDEBUG_3 \
1774         elog(DEBUG2, "heap_getnext returning tuple")
1775 #else
1776 #define HEAPDEBUG_1
1777 #define HEAPDEBUG_2
1778 #define HEAPDEBUG_3
1779 #endif   /* !defined(HEAPDEBUGALL) */
1780
1781
1782 HeapTuple
1783 heap_getnext(HeapScanDesc scan, ScanDirection direction)
1784 {
1785         /* Note: no locking manipulations needed */
1786
1787         HEAPDEBUG_1;                            /* heap_getnext( info ) */
1788
1789         if (scan->rs_pageatatime)
1790                 heapgettup_pagemode(scan, direction,
1791                                                         scan->rs_nkeys, scan->rs_key);
1792         else
1793                 heapgettup(scan, direction, scan->rs_nkeys, scan->rs_key);
1794
1795         if (scan->rs_ctup.t_data == NULL)
1796         {
1797                 HEAPDEBUG_2;                    /* heap_getnext returning EOS */
1798                 return NULL;
1799         }
1800
1801         /*
1802          * if we get here it means we have a new current scan tuple, so point to
1803          * the proper return buffer and return the tuple.
1804          */
1805         HEAPDEBUG_3;                            /* heap_getnext returning tuple */
1806
1807         pgstat_count_heap_getnext(scan->rs_rd);
1808
1809         return &(scan->rs_ctup);
1810 }
1811
1812 /*
1813  *      heap_fetch              - retrieve tuple with given tid
1814  *
1815  * On entry, tuple->t_self is the TID to fetch.  We pin the buffer holding
1816  * the tuple, fill in the remaining fields of *tuple, and check the tuple
1817  * against the specified snapshot.
1818  *
1819  * If successful (tuple found and passes snapshot time qual), then *userbuf
1820  * is set to the buffer holding the tuple and TRUE is returned.  The caller
1821  * must unpin the buffer when done with the tuple.
1822  *
1823  * If the tuple is not found (ie, item number references a deleted slot),
1824  * then tuple->t_data is set to NULL and FALSE is returned.
1825  *
1826  * If the tuple is found but fails the time qual check, then FALSE is returned
1827  * but tuple->t_data is left pointing to the tuple.
1828  *
1829  * keep_buf determines what is done with the buffer in the FALSE-result cases.
1830  * When the caller specifies keep_buf = true, we retain the pin on the buffer
1831  * and return it in *userbuf (so the caller must eventually unpin it); when
1832  * keep_buf = false, the pin is released and *userbuf is set to InvalidBuffer.
1833  *
1834  * stats_relation is the relation to charge the heap_fetch operation against
1835  * for statistical purposes.  (This could be the heap rel itself, an
1836  * associated index, or NULL to not count the fetch at all.)
1837  *
1838  * heap_fetch does not follow HOT chains: only the exact TID requested will
1839  * be fetched.
1840  *
1841  * It is somewhat inconsistent that we ereport() on invalid block number but
1842  * return false on invalid item number.  There are a couple of reasons though.
1843  * One is that the caller can relatively easily check the block number for
1844  * validity, but cannot check the item number without reading the page
1845  * himself.  Another is that when we are following a t_ctid link, we can be
1846  * reasonably confident that the page number is valid (since VACUUM shouldn't
1847  * truncate off the destination page without having killed the referencing
1848  * tuple first), but the item number might well not be good.
1849  */
1850 bool
1851 heap_fetch(Relation relation,
1852                    Snapshot snapshot,
1853                    HeapTuple tuple,
1854                    Buffer *userbuf,
1855                    bool keep_buf,
1856                    Relation stats_relation)
1857 {
1858         ItemPointer tid = &(tuple->t_self);
1859         ItemId          lp;
1860         Buffer          buffer;
1861         Page            page;
1862         OffsetNumber offnum;
1863         bool            valid;
1864
1865         /*
1866          * Fetch and pin the appropriate page of the relation.
1867          */
1868         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1869
1870         /*
1871          * Need share lock on buffer to examine tuple commit status.
1872          */
1873         LockBuffer(buffer, BUFFER_LOCK_SHARE);
1874         page = BufferGetPage(buffer);
1875
1876         /*
1877          * We'd better check for out-of-range offnum in case of VACUUM since the
1878          * TID was obtained.
1879          */
1880         offnum = ItemPointerGetOffsetNumber(tid);
1881         if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1882         {
1883                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1884                 if (keep_buf)
1885                         *userbuf = buffer;
1886                 else
1887                 {
1888                         ReleaseBuffer(buffer);
1889                         *userbuf = InvalidBuffer;
1890                 }
1891                 tuple->t_data = NULL;
1892                 return false;
1893         }
1894
1895         /*
1896          * get the item line pointer corresponding to the requested tid
1897          */
1898         lp = PageGetItemId(page, offnum);
1899
1900         /*
1901          * Must check for deleted tuple.
1902          */
1903         if (!ItemIdIsNormal(lp))
1904         {
1905                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1906                 if (keep_buf)
1907                         *userbuf = buffer;
1908                 else
1909                 {
1910                         ReleaseBuffer(buffer);
1911                         *userbuf = InvalidBuffer;
1912                 }
1913                 tuple->t_data = NULL;
1914                 return false;
1915         }
1916
1917         /*
1918          * fill in *tuple fields
1919          */
1920         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1921         tuple->t_len = ItemIdGetLength(lp);
1922         tuple->t_tableOid = RelationGetRelid(relation);
1923
1924         /*
1925          * check time qualification of tuple, then release lock
1926          */
1927         valid = HeapTupleSatisfiesVisibility(tuple, snapshot, buffer);
1928
1929         if (valid)
1930                 PredicateLockTuple(relation, tuple, snapshot);
1931
1932         CheckForSerializableConflictOut(valid, relation, tuple, buffer, snapshot);
1933
1934         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1935
1936         if (valid)
1937         {
1938                 /*
1939                  * All checks passed, so return the tuple as valid. Caller is now
1940                  * responsible for releasing the buffer.
1941                  */
1942                 *userbuf = buffer;
1943
1944                 /* Count the successful fetch against appropriate rel, if any */
1945                 if (stats_relation != NULL)
1946                         pgstat_count_heap_fetch(stats_relation);
1947
1948                 return true;
1949         }
1950
1951         /* Tuple failed time qual, but maybe caller wants to see it anyway. */
1952         if (keep_buf)
1953                 *userbuf = buffer;
1954         else
1955         {
1956                 ReleaseBuffer(buffer);
1957                 *userbuf = InvalidBuffer;
1958         }
1959
1960         return false;
1961 }
1962
1963 /*
1964  *      heap_hot_search_buffer  - search HOT chain for tuple satisfying snapshot
1965  *
1966  * On entry, *tid is the TID of a tuple (either a simple tuple, or the root
1967  * of a HOT chain), and buffer is the buffer holding this tuple.  We search
1968  * for the first chain member satisfying the given snapshot.  If one is
1969  * found, we update *tid to reference that tuple's offset number, and
1970  * return TRUE.  If no match, return FALSE without modifying *tid.
1971  *
1972  * heapTuple is a caller-supplied buffer.  When a match is found, we return
1973  * the tuple here, in addition to updating *tid.  If no match is found, the
1974  * contents of this buffer on return are undefined.
1975  *
1976  * If all_dead is not NULL, we check non-visible tuples to see if they are
1977  * globally dead; *all_dead is set TRUE if all members of the HOT chain
1978  * are vacuumable, FALSE if not.
1979  *
1980  * Unlike heap_fetch, the caller must already have pin and (at least) share
1981  * lock on the buffer; it is still pinned/locked at exit.  Also unlike
1982  * heap_fetch, we do not report any pgstats count; caller may do so if wanted.
1983  */
1984 bool
1985 heap_hot_search_buffer(ItemPointer tid, Relation relation, Buffer buffer,
1986                                            Snapshot snapshot, HeapTuple heapTuple,
1987                                            bool *all_dead, bool first_call)
1988 {
1989         Page            dp = (Page) BufferGetPage(buffer);
1990         TransactionId prev_xmax = InvalidTransactionId;
1991         OffsetNumber offnum;
1992         bool            at_chain_start;
1993         bool            valid;
1994         bool            skip;
1995
1996         /* If this is not the first call, previous call returned a (live!) tuple */
1997         if (all_dead)
1998                 *all_dead = first_call;
1999
2000         Assert(TransactionIdIsValid(RecentGlobalXmin));
2001
2002         Assert(ItemPointerGetBlockNumber(tid) == BufferGetBlockNumber(buffer));
2003         offnum = ItemPointerGetOffsetNumber(tid);
2004         at_chain_start = first_call;
2005         skip = !first_call;
2006
2007         heapTuple->t_self = *tid;
2008
2009         /* Scan through possible multiple members of HOT-chain */
2010         for (;;)
2011         {
2012                 ItemId          lp;
2013
2014                 /* check for bogus TID */
2015                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(dp))
2016                         break;
2017
2018                 lp = PageGetItemId(dp, offnum);
2019
2020                 /* check for unused, dead, or redirected items */
2021                 if (!ItemIdIsNormal(lp))
2022                 {
2023                         /* We should only see a redirect at start of chain */
2024                         if (ItemIdIsRedirected(lp) && at_chain_start)
2025                         {
2026                                 /* Follow the redirect */
2027                                 offnum = ItemIdGetRedirect(lp);
2028                                 at_chain_start = false;
2029                                 continue;
2030                         }
2031                         /* else must be end of chain */
2032                         break;
2033                 }
2034
2035                 heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
2036                 heapTuple->t_len = ItemIdGetLength(lp);
2037                 heapTuple->t_tableOid = RelationGetRelid(relation);
2038                 ItemPointerSetOffsetNumber(&heapTuple->t_self, offnum);
2039
2040                 /*
2041                  * Shouldn't see a HEAP_ONLY tuple at chain start.
2042                  */
2043                 if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
2044                         break;
2045
2046                 /*
2047                  * The xmin should match the previous xmax value, else chain is
2048                  * broken.
2049                  */
2050                 if (TransactionIdIsValid(prev_xmax) &&
2051                         !TransactionIdEquals(prev_xmax,
2052                                                                  HeapTupleHeaderGetXmin(heapTuple->t_data)))
2053                         break;
2054
2055                 /*
2056                  * When first_call is true (and thus, skip is initially false) we'll
2057                  * return the first tuple we find.  But on later passes, heapTuple
2058                  * will initially be pointing to the tuple we returned last time.
2059                  * Returning it again would be incorrect (and would loop forever), so
2060                  * we skip it and return the next match we find.
2061                  */
2062                 if (!skip)
2063                 {
2064                         /*
2065                          * For the benefit of logical decoding, have t_self point at the
2066                          * element of the HOT chain we're currently investigating instead
2067                          * of the root tuple of the HOT chain. This is important because
2068                          * the *Satisfies routine for historical mvcc snapshots needs the
2069                          * correct tid to decide about the visibility in some cases.
2070                          */
2071                         ItemPointerSet(&(heapTuple->t_self), BufferGetBlockNumber(buffer), offnum);
2072
2073                         /* If it's visible per the snapshot, we must return it */
2074                         valid = HeapTupleSatisfiesVisibility(heapTuple, snapshot, buffer);
2075                         CheckForSerializableConflictOut(valid, relation, heapTuple,
2076                                                                                         buffer, snapshot);
2077                         /* reset to original, non-redirected, tid */
2078                         heapTuple->t_self = *tid;
2079
2080                         if (valid)
2081                         {
2082                                 ItemPointerSetOffsetNumber(tid, offnum);
2083                                 PredicateLockTuple(relation, heapTuple, snapshot);
2084                                 if (all_dead)
2085                                         *all_dead = false;
2086                                 return true;
2087                         }
2088                 }
2089                 skip = false;
2090
2091                 /*
2092                  * If we can't see it, maybe no one else can either.  At caller
2093                  * request, check whether all chain members are dead to all
2094                  * transactions.
2095                  */
2096                 if (all_dead && *all_dead &&
2097                         !HeapTupleIsSurelyDead(heapTuple, RecentGlobalXmin))
2098                         *all_dead = false;
2099
2100                 /*
2101                  * Check to see if HOT chain continues past this tuple; if so fetch
2102                  * the next offnum and loop around.
2103                  */
2104                 if (HeapTupleIsHotUpdated(heapTuple))
2105                 {
2106                         Assert(ItemPointerGetBlockNumber(&heapTuple->t_data->t_ctid) ==
2107                                    ItemPointerGetBlockNumber(tid));
2108                         offnum = ItemPointerGetOffsetNumber(&heapTuple->t_data->t_ctid);
2109                         at_chain_start = false;
2110                         prev_xmax = HeapTupleHeaderGetUpdateXid(heapTuple->t_data);
2111                 }
2112                 else
2113                         break;                          /* end of chain */
2114         }
2115
2116         return false;
2117 }
2118
2119 /*
2120  *      heap_hot_search         - search HOT chain for tuple satisfying snapshot
2121  *
2122  * This has the same API as heap_hot_search_buffer, except that the caller
2123  * does not provide the buffer containing the page, rather we access it
2124  * locally.
2125  */
2126 bool
2127 heap_hot_search(ItemPointer tid, Relation relation, Snapshot snapshot,
2128                                 bool *all_dead)
2129 {
2130         bool            result;
2131         Buffer          buffer;
2132         HeapTupleData heapTuple;
2133
2134         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
2135         LockBuffer(buffer, BUFFER_LOCK_SHARE);
2136         result = heap_hot_search_buffer(tid, relation, buffer, snapshot,
2137                                                                         &heapTuple, all_dead, true);
2138         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2139         ReleaseBuffer(buffer);
2140         return result;
2141 }
2142
2143 /*
2144  *      heap_get_latest_tid -  get the latest tid of a specified tuple
2145  *
2146  * Actually, this gets the latest version that is visible according to
2147  * the passed snapshot.  You can pass SnapshotDirty to get the very latest,
2148  * possibly uncommitted version.
2149  *
2150  * *tid is both an input and an output parameter: it is updated to
2151  * show the latest version of the row.  Note that it will not be changed
2152  * if no version of the row passes the snapshot test.
2153  */
2154 void
2155 heap_get_latest_tid(Relation relation,
2156                                         Snapshot snapshot,
2157                                         ItemPointer tid)
2158 {
2159         BlockNumber blk;
2160         ItemPointerData ctid;
2161         TransactionId priorXmax;
2162
2163         /* this is to avoid Assert failures on bad input */
2164         if (!ItemPointerIsValid(tid))
2165                 return;
2166
2167         /*
2168          * Since this can be called with user-supplied TID, don't trust the input
2169          * too much.  (RelationGetNumberOfBlocks is an expensive check, so we
2170          * don't check t_ctid links again this way.  Note that it would not do to
2171          * call it just once and save the result, either.)
2172          */
2173         blk = ItemPointerGetBlockNumber(tid);
2174         if (blk >= RelationGetNumberOfBlocks(relation))
2175                 elog(ERROR, "block number %u is out of range for relation \"%s\"",
2176                          blk, RelationGetRelationName(relation));
2177
2178         /*
2179          * Loop to chase down t_ctid links.  At top of loop, ctid is the tuple we
2180          * need to examine, and *tid is the TID we will return if ctid turns out
2181          * to be bogus.
2182          *
2183          * Note that we will loop until we reach the end of the t_ctid chain.
2184          * Depending on the snapshot passed, there might be at most one visible
2185          * version of the row, but we don't try to optimize for that.
2186          */
2187         ctid = *tid;
2188         priorXmax = InvalidTransactionId;       /* cannot check first XMIN */
2189         for (;;)
2190         {
2191                 Buffer          buffer;
2192                 Page            page;
2193                 OffsetNumber offnum;
2194                 ItemId          lp;
2195                 HeapTupleData tp;
2196                 bool            valid;
2197
2198                 /*
2199                  * Read, pin, and lock the page.
2200                  */
2201                 buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&ctid));
2202                 LockBuffer(buffer, BUFFER_LOCK_SHARE);
2203                 page = BufferGetPage(buffer);
2204
2205                 /*
2206                  * Check for bogus item number.  This is not treated as an error
2207                  * condition because it can happen while following a t_ctid link. We
2208                  * just assume that the prior tid is OK and return it unchanged.
2209                  */
2210                 offnum = ItemPointerGetOffsetNumber(&ctid);
2211                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
2212                 {
2213                         UnlockReleaseBuffer(buffer);
2214                         break;
2215                 }
2216                 lp = PageGetItemId(page, offnum);
2217                 if (!ItemIdIsNormal(lp))
2218                 {
2219                         UnlockReleaseBuffer(buffer);
2220                         break;
2221                 }
2222
2223                 /* OK to access the tuple */
2224                 tp.t_self = ctid;
2225                 tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
2226                 tp.t_len = ItemIdGetLength(lp);
2227                 tp.t_tableOid = RelationGetRelid(relation);
2228
2229                 /*
2230                  * After following a t_ctid link, we might arrive at an unrelated
2231                  * tuple.  Check for XMIN match.
2232                  */
2233                 if (TransactionIdIsValid(priorXmax) &&
2234                   !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(tp.t_data)))
2235                 {
2236                         UnlockReleaseBuffer(buffer);
2237                         break;
2238                 }
2239
2240                 /*
2241                  * Check time qualification of tuple; if visible, set it as the new
2242                  * result candidate.
2243                  */
2244                 valid = HeapTupleSatisfiesVisibility(&tp, snapshot, buffer);
2245                 CheckForSerializableConflictOut(valid, relation, &tp, buffer, snapshot);
2246                 if (valid)
2247                         *tid = ctid;
2248
2249                 /*
2250                  * If there's a valid t_ctid link, follow it, else we're done.
2251                  */
2252                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
2253                         HeapTupleHeaderIsOnlyLocked(tp.t_data) ||
2254                         ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
2255                 {
2256                         UnlockReleaseBuffer(buffer);
2257                         break;
2258                 }
2259
2260                 ctid = tp.t_data->t_ctid;
2261                 priorXmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
2262                 UnlockReleaseBuffer(buffer);
2263         }                                                       /* end of loop */
2264 }
2265
2266
2267 /*
2268  * UpdateXmaxHintBits - update tuple hint bits after xmax transaction ends
2269  *
2270  * This is called after we have waited for the XMAX transaction to terminate.
2271  * If the transaction aborted, we guarantee the XMAX_INVALID hint bit will
2272  * be set on exit.  If the transaction committed, we set the XMAX_COMMITTED
2273  * hint bit if possible --- but beware that that may not yet be possible,
2274  * if the transaction committed asynchronously.
2275  *
2276  * Note that if the transaction was a locker only, we set HEAP_XMAX_INVALID
2277  * even if it commits.
2278  *
2279  * Hence callers should look only at XMAX_INVALID.
2280  *
2281  * Note this is not allowed for tuples whose xmax is a multixact.
2282  */
2283 static void
2284 UpdateXmaxHintBits(HeapTupleHeader tuple, Buffer buffer, TransactionId xid)
2285 {
2286         Assert(TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple), xid));
2287         Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));
2288
2289         if (!(tuple->t_infomask & (HEAP_XMAX_COMMITTED | HEAP_XMAX_INVALID)))
2290         {
2291                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
2292                         TransactionIdDidCommit(xid))
2293                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_COMMITTED,
2294                                                                  xid);
2295                 else
2296                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_INVALID,
2297                                                                  InvalidTransactionId);
2298         }
2299 }
2300
2301
2302 /*
2303  * GetBulkInsertState - prepare status object for a bulk insert
2304  */
2305 BulkInsertState
2306 GetBulkInsertState(void)
2307 {
2308         BulkInsertState bistate;
2309
2310         bistate = (BulkInsertState) palloc(sizeof(BulkInsertStateData));
2311         bistate->strategy = GetAccessStrategy(BAS_BULKWRITE);
2312         bistate->current_buf = InvalidBuffer;
2313         return bistate;
2314 }
2315
2316 /*
2317  * FreeBulkInsertState - clean up after finishing a bulk insert
2318  */
2319 void
2320 FreeBulkInsertState(BulkInsertState bistate)
2321 {
2322         if (bistate->current_buf != InvalidBuffer)
2323                 ReleaseBuffer(bistate->current_buf);
2324         FreeAccessStrategy(bistate->strategy);
2325         pfree(bistate);
2326 }
2327
2328
2329 /*
2330  *      heap_insert             - insert tuple into a heap
2331  *
2332  * The new tuple is stamped with current transaction ID and the specified
2333  * command ID.
2334  *
2335  * If the HEAP_INSERT_SKIP_WAL option is specified, the new tuple is not
2336  * logged in WAL, even for a non-temp relation.  Safe usage of this behavior
2337  * requires that we arrange that all new tuples go into new pages not
2338  * containing any tuples from other transactions, and that the relation gets
2339  * fsync'd before commit.  (See also heap_sync() comments)
2340  *
2341  * The HEAP_INSERT_SKIP_FSM option is passed directly to
2342  * RelationGetBufferForTuple, which see for more info.
2343  *
2344  * HEAP_INSERT_FROZEN should only be specified for inserts into
2345  * relfilenodes created during the current subtransaction and when
2346  * there are no prior snapshots or pre-existing portals open.
2347  * This causes rows to be frozen, which is an MVCC violation and
2348  * requires explicit options chosen by user.
2349  *
2350  * HEAP_INSERT_IS_SPECULATIVE is used on so-called "speculative insertions",
2351  * which can be backed out afterwards without aborting the whole transaction.
2352  * Other sessions can wait for the speculative insertion to be confirmed,
2353  * turning it into a regular tuple, or aborted, as if it never existed.
2354  * Speculatively inserted tuples behave as "value locks" of short duration,
2355  * used to implement INSERT .. ON CONFLICT.
2356  *
2357  * Note that most of these options will be applied when inserting into the
2358  * heap's TOAST table, too, if the tuple requires any out-of-line data.  Only
2359  * HEAP_INSERT_IS_SPECULATIVE is explicitly ignored, as the toast data does
2360  * not partake in speculative insertion.
2361  *
2362  * The BulkInsertState object (if any; bistate can be NULL for default
2363  * behavior) is also just passed through to RelationGetBufferForTuple.
2364  *
2365  * The return value is the OID assigned to the tuple (either here or by the
2366  * caller), or InvalidOid if no OID.  The header fields of *tup are updated
2367  * to match the stored tuple; in particular tup->t_self receives the actual
2368  * TID where the tuple was stored.  But note that any toasting of fields
2369  * within the tuple data is NOT reflected into *tup.
2370  */
2371 Oid
2372 heap_insert(Relation relation, HeapTuple tup, CommandId cid,
2373                         int options, BulkInsertState bistate)
2374 {
2375         TransactionId xid = GetCurrentTransactionId();
2376         HeapTuple       heaptup;
2377         Buffer          buffer;
2378         Buffer          vmbuffer = InvalidBuffer;
2379         bool            all_visible_cleared = false;
2380
2381         /*
2382          * Fill in tuple header fields, assign an OID, and toast the tuple if
2383          * necessary.
2384          *
2385          * Note: below this point, heaptup is the data we actually intend to store
2386          * into the relation; tup is the caller's original untoasted data.
2387          */
2388         heaptup = heap_prepare_insert(relation, tup, xid, cid, options);
2389
2390         /*
2391          * Find buffer to insert this tuple into.  If the page is all visible,
2392          * this will also pin the requisite visibility map page.
2393          */
2394         buffer = RelationGetBufferForTuple(relation, heaptup->t_len,
2395                                                                            InvalidBuffer, options, bistate,
2396                                                                            &vmbuffer, NULL);
2397
2398         /*
2399          * We're about to do the actual insert -- but check for conflict first, to
2400          * avoid possibly having to roll back work we've just done.
2401          *
2402          * This is safe without a recheck as long as there is no possibility of
2403          * another process scanning the page between this check and the insert
2404          * being visible to the scan (i.e., an exclusive buffer content lock is
2405          * continuously held from this point until the tuple insert is visible).
2406          *
2407          * For a heap insert, we only need to check for table-level SSI locks. Our
2408          * new tuple can't possibly conflict with existing tuple locks, and heap
2409          * page locks are only consolidated versions of tuple locks; they do not
2410          * lock "gaps" as index page locks do.  So we don't need to specify a
2411          * buffer when making the call, which makes for a faster check.
2412          */
2413         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2414
2415         /* NO EREPORT(ERROR) from here till changes are logged */
2416         START_CRIT_SECTION();
2417
2418         RelationPutHeapTuple(relation, buffer, heaptup,
2419                                                  (options & HEAP_INSERT_SPECULATIVE) != 0);
2420
2421         if (PageIsAllVisible(BufferGetPage(buffer)))
2422         {
2423                 all_visible_cleared = true;
2424                 PageClearAllVisible(BufferGetPage(buffer));
2425                 visibilitymap_clear(relation,
2426                                                         ItemPointerGetBlockNumber(&(heaptup->t_self)),
2427                                                         vmbuffer);
2428         }
2429
2430         /*
2431          * XXX Should we set PageSetPrunable on this page ?
2432          *
2433          * The inserting transaction may eventually abort thus making this tuple
2434          * DEAD and hence available for pruning. Though we don't want to optimize
2435          * for aborts, if no other tuple in this page is UPDATEd/DELETEd, the
2436          * aborted tuple will never be pruned until next vacuum is triggered.
2437          *
2438          * If you do add PageSetPrunable here, add it in heap_xlog_insert too.
2439          */
2440
2441         MarkBufferDirty(buffer);
2442
2443         /* XLOG stuff */
2444         if (!(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation))
2445         {
2446                 xl_heap_insert xlrec;
2447                 xl_heap_header xlhdr;
2448                 XLogRecPtr      recptr;
2449                 Page            page = BufferGetPage(buffer);
2450                 uint8           info = XLOG_HEAP_INSERT;
2451                 int                     bufflags = 0;
2452
2453                 /*
2454                  * If this is a catalog, we need to transmit combocids to properly
2455                  * decode, so log that as well.
2456                  */
2457                 if (RelationIsAccessibleInLogicalDecoding(relation))
2458                         log_heap_new_cid(relation, heaptup);
2459
2460                 /*
2461                  * If this is the single and first tuple on page, we can reinit the
2462                  * page instead of restoring the whole thing.  Set flag, and hide
2463                  * buffer references from XLogInsert.
2464                  */
2465                 if (ItemPointerGetOffsetNumber(&(heaptup->t_self)) == FirstOffsetNumber &&
2466                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
2467                 {
2468                         info |= XLOG_HEAP_INIT_PAGE;
2469                         bufflags |= REGBUF_WILL_INIT;
2470                 }
2471
2472                 xlrec.offnum = ItemPointerGetOffsetNumber(&heaptup->t_self);
2473                 xlrec.flags = 0;
2474                 if (all_visible_cleared)
2475                         xlrec.flags |= XLH_INSERT_ALL_VISIBLE_CLEARED;
2476                 if (options & HEAP_INSERT_SPECULATIVE)
2477                         xlrec.flags |= XLH_INSERT_IS_SPECULATIVE;
2478                 Assert(ItemPointerGetBlockNumber(&heaptup->t_self) == BufferGetBlockNumber(buffer));
2479
2480                 /*
2481                  * For logical decoding, we need the tuple even if we're doing a full
2482                  * page write, so make sure it's included even if we take a full-page
2483                  * image. (XXX We could alternatively store a pointer into the FPW).
2484                  */
2485                 if (RelationIsLogicallyLogged(relation))
2486                 {
2487                         xlrec.flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2488                         bufflags |= REGBUF_KEEP_DATA;
2489                 }
2490
2491                 XLogBeginInsert();
2492                 XLogRegisterData((char *) &xlrec, SizeOfHeapInsert);
2493
2494                 xlhdr.t_infomask2 = heaptup->t_data->t_infomask2;
2495                 xlhdr.t_infomask = heaptup->t_data->t_infomask;
2496                 xlhdr.t_hoff = heaptup->t_data->t_hoff;
2497
2498                 /*
2499                  * note we mark xlhdr as belonging to buffer; if XLogInsert decides to
2500                  * write the whole page to the xlog, we don't need to store
2501                  * xl_heap_header in the xlog.
2502                  */
2503                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2504                 XLogRegisterBufData(0, (char *) &xlhdr, SizeOfHeapHeader);
2505                 /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
2506                 XLogRegisterBufData(0,
2507                                                         (char *) heaptup->t_data + SizeofHeapTupleHeader,
2508                                                         heaptup->t_len - SizeofHeapTupleHeader);
2509
2510                 /* filtering by origin on a row level is much more efficient */
2511                 XLogIncludeOrigin();
2512
2513                 recptr = XLogInsert(RM_HEAP_ID, info);
2514
2515                 PageSetLSN(page, recptr);
2516         }
2517
2518         END_CRIT_SECTION();
2519
2520         UnlockReleaseBuffer(buffer);
2521         if (vmbuffer != InvalidBuffer)
2522                 ReleaseBuffer(vmbuffer);
2523
2524         /*
2525          * If tuple is cachable, mark it for invalidation from the caches in case
2526          * we abort.  Note it is OK to do this after releasing the buffer, because
2527          * the heaptup data structure is all in local memory, not in the shared
2528          * buffer.
2529          */
2530         CacheInvalidateHeapTuple(relation, heaptup, NULL);
2531
2532         /* Note: speculative insertions are counted too, even if aborted later */
2533         pgstat_count_heap_insert(relation, 1);
2534
2535         /*
2536          * If heaptup is a private copy, release it.  Don't forget to copy t_self
2537          * back to the caller's image, too.
2538          */
2539         if (heaptup != tup)
2540         {
2541                 tup->t_self = heaptup->t_self;
2542                 heap_freetuple(heaptup);
2543         }
2544
2545         return HeapTupleGetOid(tup);
2546 }
2547
2548 /*
2549  * Subroutine for heap_insert(). Prepares a tuple for insertion. This sets the
2550  * tuple header fields, assigns an OID, and toasts the tuple if necessary.
2551  * Returns a toasted version of the tuple if it was toasted, or the original
2552  * tuple if not. Note that in any case, the header fields are also set in
2553  * the original tuple.
2554  */
2555 static HeapTuple
2556 heap_prepare_insert(Relation relation, HeapTuple tup, TransactionId xid,
2557                                         CommandId cid, int options)
2558 {
2559         /*
2560          * For now, parallel operations are required to be strictly read-only.
2561          * Unlike heap_update() and heap_delete(), an insert should never create a
2562          * combo CID, so it might be possible to relax this restriction, but not
2563          * without more thought and testing.
2564          */
2565         if (IsInParallelMode())
2566                 ereport(ERROR,
2567                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2568                                  errmsg("cannot insert tuples during a parallel operation")));
2569
2570         if (relation->rd_rel->relhasoids)
2571         {
2572 #ifdef NOT_USED
2573                 /* this is redundant with an Assert in HeapTupleSetOid */
2574                 Assert(tup->t_data->t_infomask & HEAP_HASOID);
2575 #endif
2576
2577                 /*
2578                  * If the object id of this tuple has already been assigned, trust the
2579                  * caller.  There are a couple of ways this can happen.  At initial db
2580                  * creation, the backend program sets oids for tuples. When we define
2581                  * an index, we set the oid.  Finally, in the future, we may allow
2582                  * users to set their own object ids in order to support a persistent
2583                  * object store (objects need to contain pointers to one another).
2584                  */
2585                 if (!OidIsValid(HeapTupleGetOid(tup)))
2586                         HeapTupleSetOid(tup, GetNewOid(relation));
2587         }
2588         else
2589         {
2590                 /* check there is not space for an OID */
2591                 Assert(!(tup->t_data->t_infomask & HEAP_HASOID));
2592         }
2593
2594         tup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
2595         tup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
2596         tup->t_data->t_infomask |= HEAP_XMAX_INVALID;
2597         HeapTupleHeaderSetXmin(tup->t_data, xid);
2598         if (options & HEAP_INSERT_FROZEN)
2599                 HeapTupleHeaderSetXminFrozen(tup->t_data);
2600
2601         HeapTupleHeaderSetCmin(tup->t_data, cid);
2602         HeapTupleHeaderSetXmax(tup->t_data, 0);         /* for cleanliness */
2603         tup->t_tableOid = RelationGetRelid(relation);
2604
2605         /*
2606          * If the new tuple is too big for storage or contains already toasted
2607          * out-of-line attributes from some other relation, invoke the toaster.
2608          */
2609         if (relation->rd_rel->relkind != RELKIND_RELATION &&
2610                 relation->rd_rel->relkind != RELKIND_MATVIEW)
2611         {
2612                 /* toast table entries should never be recursively toasted */
2613                 Assert(!HeapTupleHasExternal(tup));
2614                 return tup;
2615         }
2616         else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
2617                 return toast_insert_or_update(relation, tup, NULL, options);
2618         else
2619                 return tup;
2620 }
2621
2622 /*
2623  *      heap_multi_insert       - insert multiple tuple into a heap
2624  *
2625  * This is like heap_insert(), but inserts multiple tuples in one operation.
2626  * That's faster than calling heap_insert() in a loop, because when multiple
2627  * tuples can be inserted on a single page, we can write just a single WAL
2628  * record covering all of them, and only need to lock/unlock the page once.
2629  *
2630  * Note: this leaks memory into the current memory context. You can create a
2631  * temporary context before calling this, if that's a problem.
2632  */
2633 void
2634 heap_multi_insert(Relation relation, HeapTuple *tuples, int ntuples,
2635                                   CommandId cid, int options, BulkInsertState bistate)
2636 {
2637         TransactionId xid = GetCurrentTransactionId();
2638         HeapTuple  *heaptuples;
2639         int                     i;
2640         int                     ndone;
2641         char       *scratch = NULL;
2642         Page            page;
2643         bool            needwal;
2644         Size            saveFreeSpace;
2645         bool            need_tuple_data = RelationIsLogicallyLogged(relation);
2646         bool            need_cids = RelationIsAccessibleInLogicalDecoding(relation);
2647
2648         needwal = !(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation);
2649         saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
2650                                                                                                    HEAP_DEFAULT_FILLFACTOR);
2651
2652         /* Toast and set header data in all the tuples */
2653         heaptuples = palloc(ntuples * sizeof(HeapTuple));
2654         for (i = 0; i < ntuples; i++)
2655                 heaptuples[i] = heap_prepare_insert(relation, tuples[i],
2656                                                                                         xid, cid, options);
2657
2658         /*
2659          * Allocate some memory to use for constructing the WAL record. Using
2660          * palloc() within a critical section is not safe, so we allocate this
2661          * beforehand.
2662          */
2663         if (needwal)
2664                 scratch = palloc(BLCKSZ);
2665
2666         /*
2667          * We're about to do the actual inserts -- but check for conflict first,
2668          * to minimize the possibility of having to roll back work we've just
2669          * done.
2670          *
2671          * A check here does not definitively prevent a serialization anomaly;
2672          * that check MUST be done at least past the point of acquiring an
2673          * exclusive buffer content lock on every buffer that will be affected,
2674          * and MAY be done after all inserts are reflected in the buffers and
2675          * those locks are released; otherwise there race condition.  Since
2676          * multiple buffers can be locked and unlocked in the loop below, and it
2677          * would not be feasible to identify and lock all of those buffers before
2678          * the loop, we must do a final check at the end.
2679          *
2680          * The check here could be omitted with no loss of correctness; it is
2681          * present strictly as an optimization.
2682          *
2683          * For heap inserts, we only need to check for table-level SSI locks. Our
2684          * new tuples can't possibly conflict with existing tuple locks, and heap
2685          * page locks are only consolidated versions of tuple locks; they do not
2686          * lock "gaps" as index page locks do.  So we don't need to specify a
2687          * buffer when making the call, which makes for a faster check.
2688          */
2689         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2690
2691         ndone = 0;
2692         while (ndone < ntuples)
2693         {
2694                 Buffer          buffer;
2695                 Buffer          vmbuffer = InvalidBuffer;
2696                 bool            all_visible_cleared = false;
2697                 int                     nthispage;
2698
2699                 CHECK_FOR_INTERRUPTS();
2700
2701                 /*
2702                  * Find buffer where at least the next tuple will fit.  If the page is
2703                  * all-visible, this will also pin the requisite visibility map page.
2704                  */
2705                 buffer = RelationGetBufferForTuple(relation, heaptuples[ndone]->t_len,
2706                                                                                    InvalidBuffer, options, bistate,
2707                                                                                    &vmbuffer, NULL);
2708                 page = BufferGetPage(buffer);
2709
2710                 /* NO EREPORT(ERROR) from here till changes are logged */
2711                 START_CRIT_SECTION();
2712
2713                 /*
2714                  * RelationGetBufferForTuple has ensured that the first tuple fits.
2715                  * Put that on the page, and then as many other tuples as fit.
2716                  */
2717                 RelationPutHeapTuple(relation, buffer, heaptuples[ndone], false);
2718                 for (nthispage = 1; ndone + nthispage < ntuples; nthispage++)
2719                 {
2720                         HeapTuple       heaptup = heaptuples[ndone + nthispage];
2721
2722                         if (PageGetHeapFreeSpace(page) < MAXALIGN(heaptup->t_len) + saveFreeSpace)
2723                                 break;
2724
2725                         RelationPutHeapTuple(relation, buffer, heaptup, false);
2726
2727                         /*
2728                          * We don't use heap_multi_insert for catalog tuples yet, but
2729                          * better be prepared...
2730                          */
2731                         if (needwal && need_cids)
2732                                 log_heap_new_cid(relation, heaptup);
2733                 }
2734
2735                 if (PageIsAllVisible(page))
2736                 {
2737                         all_visible_cleared = true;
2738                         PageClearAllVisible(page);
2739                         visibilitymap_clear(relation,
2740                                                                 BufferGetBlockNumber(buffer),
2741                                                                 vmbuffer);
2742                 }
2743
2744                 /*
2745                  * XXX Should we set PageSetPrunable on this page ? See heap_insert()
2746                  */
2747
2748                 MarkBufferDirty(buffer);
2749
2750                 /* XLOG stuff */
2751                 if (needwal)
2752                 {
2753                         XLogRecPtr      recptr;
2754                         xl_heap_multi_insert *xlrec;
2755                         uint8           info = XLOG_HEAP2_MULTI_INSERT;
2756                         char       *tupledata;
2757                         int                     totaldatalen;
2758                         char       *scratchptr = scratch;
2759                         bool            init;
2760                         int                     bufflags = 0;
2761
2762                         /*
2763                          * If the page was previously empty, we can reinit the page
2764                          * instead of restoring the whole thing.
2765                          */
2766                         init = (ItemPointerGetOffsetNumber(&(heaptuples[ndone]->t_self)) == FirstOffsetNumber &&
2767                                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber + nthispage - 1);
2768
2769                         /* allocate xl_heap_multi_insert struct from the scratch area */
2770                         xlrec = (xl_heap_multi_insert *) scratchptr;
2771                         scratchptr += SizeOfHeapMultiInsert;
2772
2773                         /*
2774                          * Allocate offsets array. Unless we're reinitializing the page,
2775                          * in that case the tuples are stored in order starting at
2776                          * FirstOffsetNumber and we don't need to store the offsets
2777                          * explicitly.
2778                          */
2779                         if (!init)
2780                                 scratchptr += nthispage * sizeof(OffsetNumber);
2781
2782                         /* the rest of the scratch space is used for tuple data */
2783                         tupledata = scratchptr;
2784
2785                         xlrec->flags = all_visible_cleared ? XLH_INSERT_ALL_VISIBLE_CLEARED : 0;
2786                         xlrec->ntuples = nthispage;
2787
2788                         /*
2789                          * Write out an xl_multi_insert_tuple and the tuple data itself
2790                          * for each tuple.
2791                          */
2792                         for (i = 0; i < nthispage; i++)
2793                         {
2794                                 HeapTuple       heaptup = heaptuples[ndone + i];
2795                                 xl_multi_insert_tuple *tuphdr;
2796                                 int                     datalen;
2797
2798                                 if (!init)
2799                                         xlrec->offsets[i] = ItemPointerGetOffsetNumber(&heaptup->t_self);
2800                                 /* xl_multi_insert_tuple needs two-byte alignment. */
2801                                 tuphdr = (xl_multi_insert_tuple *) SHORTALIGN(scratchptr);
2802                                 scratchptr = ((char *) tuphdr) + SizeOfMultiInsertTuple;
2803
2804                                 tuphdr->t_infomask2 = heaptup->t_data->t_infomask2;
2805                                 tuphdr->t_infomask = heaptup->t_data->t_infomask;
2806                                 tuphdr->t_hoff = heaptup->t_data->t_hoff;
2807
2808                                 /* write bitmap [+ padding] [+ oid] + data */
2809                                 datalen = heaptup->t_len - SizeofHeapTupleHeader;
2810                                 memcpy(scratchptr,
2811                                            (char *) heaptup->t_data + SizeofHeapTupleHeader,
2812                                            datalen);
2813                                 tuphdr->datalen = datalen;
2814                                 scratchptr += datalen;
2815                         }
2816                         totaldatalen = scratchptr - tupledata;
2817                         Assert((scratchptr - scratch) < BLCKSZ);
2818
2819                         if (need_tuple_data)
2820                                 xlrec->flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2821
2822                         /*
2823                          * Signal that this is the last xl_heap_multi_insert record
2824                          * emitted by this call to heap_multi_insert(). Needed for logical
2825                          * decoding so it knows when to cleanup temporary data.
2826                          */
2827                         if (ndone + nthispage == ntuples)
2828                                 xlrec->flags |= XLH_INSERT_LAST_IN_MULTI;
2829
2830                         if (init)
2831                         {
2832                                 info |= XLOG_HEAP_INIT_PAGE;
2833                                 bufflags |= REGBUF_WILL_INIT;
2834                         }
2835
2836                         /*
2837                          * If we're doing logical decoding, include the new tuple data
2838                          * even if we take a full-page image of the page.
2839                          */
2840                         if (need_tuple_data)
2841                                 bufflags |= REGBUF_KEEP_DATA;
2842
2843                         XLogBeginInsert();
2844                         XLogRegisterData((char *) xlrec, tupledata - scratch);
2845                         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2846
2847                         XLogRegisterBufData(0, tupledata, totaldatalen);
2848
2849                         /* filtering by origin on a row level is much more efficient */
2850                         XLogIncludeOrigin();
2851
2852                         recptr = XLogInsert(RM_HEAP2_ID, info);
2853
2854                         PageSetLSN(page, recptr);
2855                 }
2856
2857                 END_CRIT_SECTION();
2858
2859                 UnlockReleaseBuffer(buffer);
2860                 if (vmbuffer != InvalidBuffer)
2861                         ReleaseBuffer(vmbuffer);
2862
2863                 ndone += nthispage;
2864         }
2865
2866         /*
2867          * We're done with the actual inserts.  Check for conflicts again, to
2868          * ensure that all rw-conflicts in to these inserts are detected.  Without
2869          * this final check, a sequential scan of the heap may have locked the
2870          * table after the "before" check, missing one opportunity to detect the
2871          * conflict, and then scanned the table before the new tuples were there,
2872          * missing the other chance to detect the conflict.
2873          *
2874          * For heap inserts, we only need to check for table-level SSI locks. Our
2875          * new tuples can't possibly conflict with existing tuple locks, and heap
2876          * page locks are only consolidated versions of tuple locks; they do not
2877          * lock "gaps" as index page locks do.  So we don't need to specify a
2878          * buffer when making the call.
2879          */
2880         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2881
2882         /*
2883          * If tuples are cachable, mark them for invalidation from the caches in
2884          * case we abort.  Note it is OK to do this after releasing the buffer,
2885          * because the heaptuples data structure is all in local memory, not in
2886          * the shared buffer.
2887          */
2888         if (IsCatalogRelation(relation))
2889         {
2890                 for (i = 0; i < ntuples; i++)
2891                         CacheInvalidateHeapTuple(relation, heaptuples[i], NULL);
2892         }
2893
2894         /*
2895          * Copy t_self fields back to the caller's original tuples. This does
2896          * nothing for untoasted tuples (tuples[i] == heaptuples[i)], but it's
2897          * probably faster to always copy than check.
2898          */
2899         for (i = 0; i < ntuples; i++)
2900                 tuples[i]->t_self = heaptuples[i]->t_self;
2901
2902         pgstat_count_heap_insert(relation, ntuples);
2903 }
2904
2905 /*
2906  *      simple_heap_insert - insert a tuple
2907  *
2908  * Currently, this routine differs from heap_insert only in supplying
2909  * a default command ID and not allowing access to the speedup options.
2910  *
2911  * This should be used rather than using heap_insert directly in most places
2912  * where we are modifying system catalogs.
2913  */
2914 Oid
2915 simple_heap_insert(Relation relation, HeapTuple tup)
2916 {
2917         return heap_insert(relation, tup, GetCurrentCommandId(true), 0, NULL);
2918 }
2919
2920 /*
2921  * Given infomask/infomask2, compute the bits that must be saved in the
2922  * "infobits" field of xl_heap_delete, xl_heap_update, xl_heap_lock,
2923  * xl_heap_lock_updated WAL records.
2924  *
2925  * See fix_infomask_from_infobits.
2926  */
2927 static uint8
2928 compute_infobits(uint16 infomask, uint16 infomask2)
2929 {
2930         return
2931                 ((infomask & HEAP_XMAX_IS_MULTI) != 0 ? XLHL_XMAX_IS_MULTI : 0) |
2932                 ((infomask & HEAP_XMAX_LOCK_ONLY) != 0 ? XLHL_XMAX_LOCK_ONLY : 0) |
2933                 ((infomask & HEAP_XMAX_EXCL_LOCK) != 0 ? XLHL_XMAX_EXCL_LOCK : 0) |
2934         /* note we ignore HEAP_XMAX_SHR_LOCK here */
2935                 ((infomask & HEAP_XMAX_KEYSHR_LOCK) != 0 ? XLHL_XMAX_KEYSHR_LOCK : 0) |
2936                 ((infomask2 & HEAP_KEYS_UPDATED) != 0 ?
2937                  XLHL_KEYS_UPDATED : 0);
2938 }
2939
2940 /*
2941  * Given two versions of the same t_infomask for a tuple, compare them and
2942  * return whether the relevant status for a tuple Xmax has changed.  This is
2943  * used after a buffer lock has been released and reacquired: we want to ensure
2944  * that the tuple state continues to be the same it was when we previously
2945  * examined it.
2946  *
2947  * Note the Xmax field itself must be compared separately.
2948  */
2949 static inline bool
2950 xmax_infomask_changed(uint16 new_infomask, uint16 old_infomask)
2951 {
2952         const uint16 interesting =
2953         HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY | HEAP_LOCK_MASK;
2954
2955         if ((new_infomask & interesting) != (old_infomask & interesting))
2956                 return true;
2957
2958         return false;
2959 }
2960
2961 /*
2962  *      heap_delete - delete a tuple
2963  *
2964  * NB: do not call this directly unless you are prepared to deal with
2965  * concurrent-update conditions.  Use simple_heap_delete instead.
2966  *
2967  *      relation - table to be modified (caller must hold suitable lock)
2968  *      tid - TID of tuple to be deleted
2969  *      cid - delete command ID (used for visibility test, and stored into
2970  *              cmax if successful)
2971  *      crosscheck - if not InvalidSnapshot, also check tuple against this
2972  *      wait - true if should wait for any conflicting update to commit/abort
2973  *      hufd - output parameter, filled in failure cases (see below)
2974  *
2975  * Normal, successful return value is HeapTupleMayBeUpdated, which
2976  * actually means we did delete it.  Failure return codes are
2977  * HeapTupleSelfUpdated, HeapTupleUpdated, or HeapTupleBeingUpdated
2978  * (the last only possible if wait == false).
2979  *
2980  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
2981  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
2982  * (the last only for HeapTupleSelfUpdated, since we
2983  * cannot obtain cmax from a combocid generated by another transaction).
2984  * See comments for struct HeapUpdateFailureData for additional info.
2985  */
2986 HTSU_Result
2987 heap_delete(Relation relation, ItemPointer tid,
2988                         CommandId cid, Snapshot crosscheck, bool wait,
2989                         HeapUpdateFailureData *hufd)
2990 {
2991         HTSU_Result result;
2992         TransactionId xid = GetCurrentTransactionId();
2993         ItemId          lp;
2994         HeapTupleData tp;
2995         Page            page;
2996         BlockNumber block;
2997         Buffer          buffer;
2998         Buffer          vmbuffer = InvalidBuffer;
2999         TransactionId new_xmax;
3000         uint16          new_infomask,
3001                                 new_infomask2;
3002         bool            have_tuple_lock = false;
3003         bool            iscombo;
3004         bool            all_visible_cleared = false;
3005         HeapTuple       old_key_tuple = NULL;   /* replica identity of the tuple */
3006         bool            old_key_copied = false;
3007
3008         Assert(ItemPointerIsValid(tid));
3009
3010         /*
3011          * Forbid this during a parallel operation, lets it allocate a combocid.
3012          * Other workers might need that combocid for visibility checks, and we
3013          * have no provision for broadcasting it to them.
3014          */
3015         if (IsInParallelMode())
3016                 ereport(ERROR,
3017                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
3018                                  errmsg("cannot delete tuples during a parallel operation")));
3019
3020         block = ItemPointerGetBlockNumber(tid);
3021         buffer = ReadBuffer(relation, block);
3022         page = BufferGetPage(buffer);
3023
3024         /*
3025          * Before locking the buffer, pin the visibility map page if it appears to
3026          * be necessary.  Since we haven't got the lock yet, someone else might be
3027          * in the middle of changing this, so we'll need to recheck after we have
3028          * the lock.
3029          */
3030         if (PageIsAllVisible(page))
3031                 visibilitymap_pin(relation, block, &vmbuffer);
3032
3033         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3034
3035         /*
3036          * If we didn't pin the visibility map page and the page has become all
3037          * visible while we were busy locking the buffer, we'll have to unlock and
3038          * re-lock, to avoid holding the buffer lock across an I/O.  That's a bit
3039          * unfortunate, but hopefully shouldn't happen often.
3040          */
3041         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3042         {
3043                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3044                 visibilitymap_pin(relation, block, &vmbuffer);
3045                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3046         }
3047
3048         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
3049         Assert(ItemIdIsNormal(lp));
3050
3051         tp.t_tableOid = RelationGetRelid(relation);
3052         tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3053         tp.t_len = ItemIdGetLength(lp);
3054         tp.t_self = *tid;
3055
3056 l1:
3057         result = HeapTupleSatisfiesUpdate(&tp, cid, buffer);
3058
3059         if (result == HeapTupleInvisible)
3060         {
3061                 UnlockReleaseBuffer(buffer);
3062                 ereport(ERROR,
3063                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3064                                  errmsg("attempted to delete invisible tuple")));
3065         }
3066         else if (result == HeapTupleBeingUpdated && wait)
3067         {
3068                 TransactionId xwait;
3069                 uint16          infomask;
3070
3071                 /* must copy state data before unlocking buffer */
3072                 xwait = HeapTupleHeaderGetRawXmax(tp.t_data);
3073                 infomask = tp.t_data->t_infomask;
3074
3075                 /*
3076                  * Sleep until concurrent transaction ends -- except when there's a
3077                  * single locker and it's our own transaction.  Note we don't care
3078                  * which lock mode the locker has, because we need the strongest one.
3079                  *
3080                  * Before sleeping, we need to acquire tuple lock to establish our
3081                  * priority for the tuple (see heap_lock_tuple).  LockTuple will
3082                  * release us when we are next-in-line for the tuple.
3083                  *
3084                  * If we are forced to "start over" below, we keep the tuple lock;
3085                  * this arranges that we stay at the head of the line while rechecking
3086                  * tuple state.
3087                  */
3088                 if (infomask & HEAP_XMAX_IS_MULTI)
3089                 {
3090                         /* wait for multixact */
3091                         if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
3092                                                                                 LockTupleExclusive))
3093                         {
3094                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3095
3096                                 /* acquire tuple lock, if necessary */
3097                                 heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
3098                                                                          LockWaitBlock, &have_tuple_lock);
3099
3100                                 /* wait for multixact */
3101                                 MultiXactIdWait((MultiXactId) xwait, MultiXactStatusUpdate, infomask,
3102                                                                 relation, &(tp.t_self), XLTW_Delete,
3103                                                                 NULL);
3104                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3105
3106                                 /*
3107                                  * If xwait had just locked the tuple then some other xact
3108                                  * could update this tuple before we get to this point.  Check
3109                                  * for xmax change, and start over if so.
3110                                  */
3111                                 if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
3112                                         !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
3113                                                                                  xwait))
3114                                         goto l1;
3115                         }
3116
3117                         /*
3118                          * You might think the multixact is necessarily done here, but not
3119                          * so: it could have surviving members, namely our own xact or
3120                          * other subxacts of this backend.  It is legal for us to delete
3121                          * the tuple in either case, however (the latter case is
3122                          * essentially a situation of upgrading our former shared lock to
3123                          * exclusive).  We don't bother changing the on-disk hint bits
3124                          * since we are about to overwrite the xmax altogether.
3125                          */
3126                 }
3127                 else if (!TransactionIdIsCurrentTransactionId(xwait))
3128                 {
3129                         /*
3130                          * Wait for regular transaction to end; but first, acquire tuple
3131                          * lock.
3132                          */
3133                         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3134                         heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
3135                                                                  LockWaitBlock, &have_tuple_lock);
3136                         XactLockTableWait(xwait, relation, &(tp.t_self), XLTW_Delete);
3137                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3138
3139                         /*
3140                          * xwait is done, but if xwait had just locked the tuple then some
3141                          * other xact could update this tuple before we get to this point.
3142                          * Check for xmax change, and start over if so.
3143                          */
3144                         if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
3145                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
3146                                                                          xwait))
3147                                 goto l1;
3148
3149                         /* Otherwise check if it committed or aborted */
3150                         UpdateXmaxHintBits(tp.t_data, buffer, xwait);
3151                 }
3152
3153                 /*
3154                  * We may overwrite if previous xmax aborted, or if it committed but
3155                  * only locked the tuple without updating it.
3156                  */
3157                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3158                         HEAP_XMAX_IS_LOCKED_ONLY(tp.t_data->t_infomask) ||
3159                         HeapTupleHeaderIsOnlyLocked(tp.t_data))
3160                         result = HeapTupleMayBeUpdated;
3161                 else
3162                         result = HeapTupleUpdated;
3163         }
3164
3165         if (crosscheck != InvalidSnapshot && result == HeapTupleMayBeUpdated)
3166         {
3167                 /* Perform additional check for transaction-snapshot mode RI updates */
3168                 if (!HeapTupleSatisfiesVisibility(&tp, crosscheck, buffer))
3169                         result = HeapTupleUpdated;
3170         }
3171
3172         if (result != HeapTupleMayBeUpdated)
3173         {
3174                 Assert(result == HeapTupleSelfUpdated ||
3175                            result == HeapTupleUpdated ||
3176                            result == HeapTupleBeingUpdated);
3177                 Assert(!(tp.t_data->t_infomask & HEAP_XMAX_INVALID));
3178                 hufd->ctid = tp.t_data->t_ctid;
3179                 hufd->xmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
3180                 if (result == HeapTupleSelfUpdated)
3181                         hufd->cmax = HeapTupleHeaderGetCmax(tp.t_data);
3182                 else
3183                         hufd->cmax = InvalidCommandId;
3184                 UnlockReleaseBuffer(buffer);
3185                 if (have_tuple_lock)
3186                         UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
3187                 if (vmbuffer != InvalidBuffer)
3188                         ReleaseBuffer(vmbuffer);
3189                 return result;
3190         }
3191
3192         /*
3193          * We're about to do the actual delete -- check for conflict first, to
3194          * avoid possibly having to roll back work we've just done.
3195          *
3196          * This is safe without a recheck as long as there is no possibility of
3197          * another process scanning the page between this check and the delete
3198          * being visible to the scan (i.e., an exclusive buffer content lock is
3199          * continuously held from this point until the tuple delete is visible).
3200          */
3201         CheckForSerializableConflictIn(relation, &tp, buffer);
3202
3203         /* replace cid with a combo cid if necessary */
3204         HeapTupleHeaderAdjustCmax(tp.t_data, &cid, &iscombo);
3205
3206         /*
3207          * Compute replica identity tuple before entering the critical section so
3208          * we don't PANIC upon a memory allocation failure.
3209          */
3210         old_key_tuple = ExtractReplicaIdentity(relation, &tp, true, &old_key_copied);
3211
3212         /*
3213          * If this is the first possibly-multixact-able operation in the current
3214          * transaction, set my per-backend OldestMemberMXactId setting. We can be
3215          * certain that the transaction will never become a member of any older
3216          * MultiXactIds than that.  (We have to do this even if we end up just
3217          * using our own TransactionId below, since some other backend could
3218          * incorporate our XID into a MultiXact immediately afterwards.)
3219          */
3220         MultiXactIdSetOldestMember();
3221
3222         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(tp.t_data),
3223                                                           tp.t_data->t_infomask, tp.t_data->t_infomask2,
3224                                                           xid, LockTupleExclusive, true,
3225                                                           &new_xmax, &new_infomask, &new_infomask2);
3226
3227         START_CRIT_SECTION();
3228
3229         /*
3230          * If this transaction commits, the tuple will become DEAD sooner or
3231          * later.  Set flag that this page is a candidate for pruning once our xid
3232          * falls below the OldestXmin horizon.  If the transaction finally aborts,
3233          * the subsequent page pruning will be a no-op and the hint will be
3234          * cleared.
3235          */
3236         PageSetPrunable(page, xid);
3237
3238         if (PageIsAllVisible(page))
3239         {
3240                 all_visible_cleared = true;
3241                 PageClearAllVisible(page);
3242                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
3243                                                         vmbuffer);
3244         }
3245
3246         /* store transaction information of xact deleting the tuple */
3247         tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3248         tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3249         tp.t_data->t_infomask |= new_infomask;
3250         tp.t_data->t_infomask2 |= new_infomask2;
3251         HeapTupleHeaderClearHotUpdated(tp.t_data);
3252         HeapTupleHeaderSetXmax(tp.t_data, new_xmax);
3253         HeapTupleHeaderSetCmax(tp.t_data, cid, iscombo);
3254         /* Make sure there is no forward chain link in t_ctid */
3255         tp.t_data->t_ctid = tp.t_self;
3256
3257         MarkBufferDirty(buffer);
3258
3259         /*
3260          * XLOG stuff
3261          *
3262          * NB: heap_abort_speculative() uses the same xlog record and replay
3263          * routines.
3264          */
3265         if (RelationNeedsWAL(relation))
3266         {
3267                 xl_heap_delete xlrec;
3268                 XLogRecPtr      recptr;
3269
3270                 /* For logical decode we need combocids to properly decode the catalog */
3271                 if (RelationIsAccessibleInLogicalDecoding(relation))
3272                         log_heap_new_cid(relation, &tp);
3273
3274                 xlrec.flags = all_visible_cleared ? XLH_DELETE_ALL_VISIBLE_CLEARED : 0;
3275                 xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
3276                                                                                           tp.t_data->t_infomask2);
3277                 xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
3278                 xlrec.xmax = new_xmax;
3279
3280                 if (old_key_tuple != NULL)
3281                 {
3282                         if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
3283                                 xlrec.flags |= XLH_DELETE_CONTAINS_OLD_TUPLE;
3284                         else
3285                                 xlrec.flags |= XLH_DELETE_CONTAINS_OLD_KEY;
3286                 }
3287
3288                 XLogBeginInsert();
3289                 XLogRegisterData((char *) &xlrec, SizeOfHeapDelete);
3290
3291                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3292
3293                 /*
3294                  * Log replica identity of the deleted tuple if there is one
3295                  */
3296                 if (old_key_tuple != NULL)
3297                 {
3298                         xl_heap_header xlhdr;
3299
3300                         xlhdr.t_infomask2 = old_key_tuple->t_data->t_infomask2;
3301                         xlhdr.t_infomask = old_key_tuple->t_data->t_infomask;
3302                         xlhdr.t_hoff = old_key_tuple->t_data->t_hoff;
3303
3304                         XLogRegisterData((char *) &xlhdr, SizeOfHeapHeader);
3305                         XLogRegisterData((char *) old_key_tuple->t_data
3306                                                          + SizeofHeapTupleHeader,
3307                                                          old_key_tuple->t_len
3308                                                          - SizeofHeapTupleHeader);
3309                 }
3310
3311                 /* filtering by origin on a row level is much more efficient */
3312                 XLogIncludeOrigin();
3313
3314                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
3315
3316                 PageSetLSN(page, recptr);
3317         }
3318
3319         END_CRIT_SECTION();
3320
3321         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3322
3323         if (vmbuffer != InvalidBuffer)
3324                 ReleaseBuffer(vmbuffer);
3325
3326         /*
3327          * If the tuple has toasted out-of-line attributes, we need to delete
3328          * those items too.  We have to do this before releasing the buffer
3329          * because we need to look at the contents of the tuple, but it's OK to
3330          * release the content lock on the buffer first.
3331          */
3332         if (relation->rd_rel->relkind != RELKIND_RELATION &&
3333                 relation->rd_rel->relkind != RELKIND_MATVIEW)
3334         {
3335                 /* toast table entries should never be recursively toasted */
3336                 Assert(!HeapTupleHasExternal(&tp));
3337         }
3338         else if (HeapTupleHasExternal(&tp))
3339                 toast_delete(relation, &tp);
3340
3341         /*
3342          * Mark tuple for invalidation from system caches at next command
3343          * boundary. We have to do this before releasing the buffer because we
3344          * need to look at the contents of the tuple.
3345          */
3346         CacheInvalidateHeapTuple(relation, &tp, NULL);
3347
3348         /* Now we can release the buffer */
3349         ReleaseBuffer(buffer);
3350
3351         /*
3352          * Release the lmgr tuple lock, if we had it.
3353          */
3354         if (have_tuple_lock)
3355                 UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
3356
3357         pgstat_count_heap_delete(relation);
3358
3359         if (old_key_tuple != NULL && old_key_copied)
3360                 heap_freetuple(old_key_tuple);
3361
3362         return HeapTupleMayBeUpdated;
3363 }
3364
3365 /*
3366  *      simple_heap_delete - delete a tuple
3367  *
3368  * This routine may be used to delete a tuple when concurrent updates of
3369  * the target tuple are not expected (for example, because we have a lock
3370  * on the relation associated with the tuple).  Any failure is reported
3371  * via ereport().
3372  */
3373 void
3374 simple_heap_delete(Relation relation, ItemPointer tid)
3375 {
3376         HTSU_Result result;
3377         HeapUpdateFailureData hufd;
3378
3379         result = heap_delete(relation, tid,
3380                                                  GetCurrentCommandId(true), InvalidSnapshot,
3381                                                  true /* wait for commit */ ,
3382                                                  &hufd);
3383         switch (result)
3384         {
3385                 case HeapTupleSelfUpdated:
3386                         /* Tuple was already updated in current command? */
3387                         elog(ERROR, "tuple already updated by self");
3388                         break;
3389
3390                 case HeapTupleMayBeUpdated:
3391                         /* done successfully */
3392                         break;
3393
3394                 case HeapTupleUpdated:
3395                         elog(ERROR, "tuple concurrently updated");
3396                         break;
3397
3398                 default:
3399                         elog(ERROR, "unrecognized heap_delete status: %u", result);
3400                         break;
3401         }
3402 }
3403
3404 /*
3405  *      heap_update - replace a tuple
3406  *
3407  * NB: do not call this directly unless you are prepared to deal with
3408  * concurrent-update conditions.  Use simple_heap_update instead.
3409  *
3410  *      relation - table to be modified (caller must hold suitable lock)
3411  *      otid - TID of old tuple to be replaced
3412  *      newtup - newly constructed tuple data to store
3413  *      cid - update command ID (used for visibility test, and stored into
3414  *              cmax/cmin if successful)
3415  *      crosscheck - if not InvalidSnapshot, also check old tuple against this
3416  *      wait - true if should wait for any conflicting update to commit/abort
3417  *      hufd - output parameter, filled in failure cases (see below)
3418  *      lockmode - output parameter, filled with lock mode acquired on tuple
3419  *
3420  * Normal, successful return value is HeapTupleMayBeUpdated, which
3421  * actually means we *did* update it.  Failure return codes are
3422  * HeapTupleSelfUpdated, HeapTupleUpdated, or HeapTupleBeingUpdated
3423  * (the last only possible if wait == false).
3424  *
3425  * On success, the header fields of *newtup are updated to match the new
3426  * stored tuple; in particular, newtup->t_self is set to the TID where the
3427  * new tuple was inserted, and its HEAP_ONLY_TUPLE flag is set iff a HOT
3428  * update was done.  However, any TOAST changes in the new tuple's
3429  * data are not reflected into *newtup.
3430  *
3431  * In the failure cases, the routine fills *hufd with the tuple's t_ctid,
3432  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax
3433  * (the last only for HeapTupleSelfUpdated, since we
3434  * cannot obtain cmax from a combocid generated by another transaction).
3435  * See comments for struct HeapUpdateFailureData for additional info.
3436  */
3437 HTSU_Result
3438 heap_update(Relation relation, ItemPointer otid, HeapTuple newtup,
3439                         CommandId cid, Snapshot crosscheck, bool wait,
3440                         HeapUpdateFailureData *hufd, LockTupleMode *lockmode)
3441 {
3442         HTSU_Result result;
3443         TransactionId xid = GetCurrentTransactionId();
3444         Bitmapset  *hot_attrs;
3445         Bitmapset  *key_attrs;
3446         Bitmapset  *id_attrs;
3447         ItemId          lp;
3448         HeapTupleData oldtup;
3449         HeapTuple       heaptup;
3450         HeapTuple       old_key_tuple = NULL;
3451         bool            old_key_copied = false;
3452         Page            page;
3453         BlockNumber block;
3454         MultiXactStatus mxact_status;
3455         Buffer          buffer,
3456                                 newbuf,
3457                                 vmbuffer = InvalidBuffer,
3458                                 vmbuffer_new = InvalidBuffer;
3459         bool            need_toast,
3460                                 already_marked;
3461         Size            newtupsize,
3462                                 pagefree;
3463         bool            have_tuple_lock = false;
3464         bool            iscombo;
3465         bool            satisfies_hot;
3466         bool            satisfies_key;
3467         bool            satisfies_id;
3468         bool            use_hot_update = false;
3469         bool            key_intact;
3470         bool            all_visible_cleared = false;
3471         bool            all_visible_cleared_new = false;
3472         bool            checked_lockers;
3473         bool            locker_remains;
3474         TransactionId xmax_new_tuple,
3475                                 xmax_old_tuple;
3476         uint16          infomask_old_tuple,
3477                                 infomask2_old_tuple,
3478                                 infomask_new_tuple,
3479                                 infomask2_new_tuple;
3480
3481         Assert(ItemPointerIsValid(otid));
3482
3483         /*
3484          * Forbid this during a parallel operation, lets it allocate a combocid.
3485          * Other workers might need that combocid for visibility checks, and we
3486          * have no provision for broadcasting it to them.
3487          */
3488         if (IsInParallelMode())
3489                 ereport(ERROR,
3490                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
3491                                  errmsg("cannot update tuples during a parallel operation")));
3492
3493         /*
3494          * Fetch the list of attributes to be checked for HOT update.  This is
3495          * wasted effort if we fail to update or have to put the new tuple on a
3496          * different page.  But we must compute the list before obtaining buffer
3497          * lock --- in the worst case, if we are doing an update on one of the
3498          * relevant system catalogs, we could deadlock if we try to fetch the list
3499          * later.  In any case, the relcache caches the data so this is usually
3500          * pretty cheap.
3501          *
3502          * Note that we get a copy here, so we need not worry about relcache flush
3503          * happening midway through.
3504          */
3505         hot_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_ALL);
3506         key_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_KEY);
3507         id_attrs = RelationGetIndexAttrBitmap(relation,
3508                                                                                   INDEX_ATTR_BITMAP_IDENTITY_KEY);
3509
3510         block = ItemPointerGetBlockNumber(otid);
3511         buffer = ReadBuffer(relation, block);
3512         page = BufferGetPage(buffer);
3513
3514         /*
3515          * Before locking the buffer, pin the visibility map page if it appears to
3516          * be necessary.  Since we haven't got the lock yet, someone else might be
3517          * in the middle of changing this, so we'll need to recheck after we have
3518          * the lock.
3519          */
3520         if (PageIsAllVisible(page))
3521                 visibilitymap_pin(relation, block, &vmbuffer);
3522
3523         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3524
3525         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(otid));
3526         Assert(ItemIdIsNormal(lp));
3527
3528         /*
3529          * Fill in enough data in oldtup for HeapSatisfiesHOTandKeyUpdate to work
3530          * properly.
3531          */
3532         oldtup.t_tableOid = RelationGetRelid(relation);
3533         oldtup.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3534         oldtup.t_len = ItemIdGetLength(lp);
3535         oldtup.t_self = *otid;
3536
3537         /* the new tuple is ready, except for this: */
3538         newtup->t_tableOid = RelationGetRelid(relation);
3539
3540         /* Fill in OID for newtup */
3541         if (relation->rd_rel->relhasoids)
3542         {
3543 #ifdef NOT_USED
3544                 /* this is redundant with an Assert in HeapTupleSetOid */
3545                 Assert(newtup->t_data->t_infomask & HEAP_HASOID);
3546 #endif
3547                 HeapTupleSetOid(newtup, HeapTupleGetOid(&oldtup));
3548         }
3549         else
3550         {
3551                 /* check there is not space for an OID */
3552                 Assert(!(newtup->t_data->t_infomask & HEAP_HASOID));
3553         }
3554
3555         /*
3556          * If we're not updating any "key" column, we can grab a weaker lock type.
3557          * This allows for more concurrency when we are running simultaneously
3558          * with foreign key checks.
3559          *
3560          * Note that if a column gets detoasted while executing the update, but
3561          * the value ends up being the same, this test will fail and we will use
3562          * the stronger lock.  This is acceptable; the important case to optimize
3563          * is updates that don't manipulate key columns, not those that
3564          * serendipitiously arrive at the same key values.
3565          */
3566         HeapSatisfiesHOTandKeyUpdate(relation, hot_attrs, key_attrs, id_attrs,
3567                                                                  &satisfies_hot, &satisfies_key,
3568                                                                  &satisfies_id, &oldtup, newtup);
3569         if (satisfies_key)
3570         {
3571                 *lockmode = LockTupleNoKeyExclusive;
3572                 mxact_status = MultiXactStatusNoKeyUpdate;
3573                 key_intact = true;
3574
3575                 /*
3576                  * If this is the first possibly-multixact-able operation in the
3577                  * current transaction, set my per-backend OldestMemberMXactId
3578                  * setting. We can be certain that the transaction will never become a
3579                  * member of any older MultiXactIds than that.  (We have to do this
3580                  * even if we end up just using our own TransactionId below, since
3581                  * some other backend could incorporate our XID into a MultiXact
3582                  * immediately afterwards.)
3583                  */
3584                 MultiXactIdSetOldestMember();
3585         }
3586         else
3587         {
3588                 *lockmode = LockTupleExclusive;
3589                 mxact_status = MultiXactStatusUpdate;
3590                 key_intact = false;
3591         }
3592
3593         /*
3594          * Note: beyond this point, use oldtup not otid to refer to old tuple.
3595          * otid may very well point at newtup->t_self, which we will overwrite
3596          * with the new tuple's location, so there's great risk of confusion if we
3597          * use otid anymore.
3598          */
3599
3600 l2:
3601         checked_lockers = false;
3602         locker_remains = false;
3603         result = HeapTupleSatisfiesUpdate(&oldtup, cid, buffer);
3604
3605         /* see below about the "no wait" case */
3606         Assert(result != HeapTupleBeingUpdated || wait);
3607
3608         if (result == HeapTupleInvisible)
3609         {
3610                 UnlockReleaseBuffer(buffer);
3611                 ereport(ERROR,
3612                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3613                                  errmsg("attempted to update invisible tuple")));
3614         }
3615         else if (result == HeapTupleBeingUpdated && wait)
3616         {
3617                 TransactionId xwait;
3618                 uint16          infomask;
3619                 bool            can_continue = false;
3620
3621                 /*
3622                  * XXX note that we don't consider the "no wait" case here.  This
3623                  * isn't a problem currently because no caller uses that case, but it
3624                  * should be fixed if such a caller is introduced.  It wasn't a
3625                  * problem previously because this code would always wait, but now
3626                  * that some tuple locks do not conflict with one of the lock modes we
3627                  * use, it is possible that this case is interesting to handle
3628                  * specially.
3629                  *
3630                  * This may cause failures with third-party code that calls
3631                  * heap_update directly.
3632                  */
3633
3634                 /* must copy state data before unlocking buffer */
3635                 xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3636                 infomask = oldtup.t_data->t_infomask;
3637
3638                 /*
3639                  * Now we have to do something about the existing locker.  If it's a
3640                  * multi, sleep on it; we might be awakened before it is completely
3641                  * gone (or even not sleep at all in some cases); we need to preserve
3642                  * it as locker, unless it is gone completely.
3643                  *
3644                  * If it's not a multi, we need to check for sleeping conditions
3645                  * before actually going to sleep.  If the update doesn't conflict
3646                  * with the locks, we just continue without sleeping (but making sure
3647                  * it is preserved).
3648                  *
3649                  * Before sleeping, we need to acquire tuple lock to establish our
3650                  * priority for the tuple (see heap_lock_tuple).  LockTuple will
3651                  * release us when we are next-in-line for the tuple.  Note we must
3652                  * not acquire the tuple lock until we're sure we're going to sleep;
3653                  * otherwise we're open for race conditions with other transactions
3654                  * holding the tuple lock which sleep on us.
3655                  *
3656                  * If we are forced to "start over" below, we keep the tuple lock;
3657                  * this arranges that we stay at the head of the line while rechecking
3658                  * tuple state.
3659                  */
3660                 if (infomask & HEAP_XMAX_IS_MULTI)
3661                 {
3662                         TransactionId update_xact;
3663                         int                     remain;
3664
3665                         if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
3666                                                                                 *lockmode))
3667                         {
3668                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3669
3670                                 /* acquire tuple lock, if necessary */
3671                                 heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3672                                                                          LockWaitBlock, &have_tuple_lock);
3673
3674                                 /* wait for multixact */
3675                                 MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
3676                                                                 relation, &oldtup.t_self, XLTW_Update,
3677                                                                 &remain);
3678                                 checked_lockers = true;
3679                                 locker_remains = remain != 0;
3680                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3681
3682                                 /*
3683                                  * If xwait had just locked the tuple then some other xact
3684                                  * could update this tuple before we get to this point.  Check
3685                                  * for xmax change, and start over if so.
3686                                  */
3687                                 if (xmax_infomask_changed(oldtup.t_data->t_infomask,
3688                                                                                   infomask) ||
3689                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3690                                                                          xwait))
3691                                         goto l2;
3692                         }
3693
3694                         /*
3695                          * Note that the multixact may not be done by now.  It could have
3696                          * surviving members; our own xact or other subxacts of this
3697                          * backend, and also any other concurrent transaction that locked
3698                          * the tuple with KeyShare if we only got TupleLockUpdate.  If
3699                          * this is the case, we have to be careful to mark the updated
3700                          * tuple with the surviving members in Xmax.
3701                          *
3702                          * Note that there could have been another update in the
3703                          * MultiXact. In that case, we need to check whether it committed
3704                          * or aborted. If it aborted we are safe to update it again;
3705                          * otherwise there is an update conflict, and we have to return
3706                          * HeapTupleUpdated below.
3707                          *
3708                          * In the LockTupleExclusive case, we still need to preserve the
3709                          * surviving members: those would include the tuple locks we had
3710                          * before this one, which are important to keep in case this
3711                          * subxact aborts.
3712                          */
3713                         if (!HEAP_XMAX_IS_LOCKED_ONLY(oldtup.t_data->t_infomask))
3714                                 update_xact = HeapTupleGetUpdateXid(oldtup.t_data);
3715                         else
3716                                 update_xact = InvalidTransactionId;
3717
3718                         /*
3719                          * There was no UPDATE in the MultiXact; or it aborted. No
3720                          * TransactionIdIsInProgress() call needed here, since we called
3721                          * MultiXactIdWait() above.
3722                          */
3723                         if (!TransactionIdIsValid(update_xact) ||
3724                                 TransactionIdDidAbort(update_xact))
3725                                 can_continue = true;
3726                 }
3727                 else if (TransactionIdIsCurrentTransactionId(xwait))
3728                 {
3729                         /*
3730                          * The only locker is ourselves; we can avoid grabbing the tuple
3731                          * lock here, but must preserve our locking information.
3732                          */
3733                         checked_lockers = true;
3734                         locker_remains = true;
3735                         can_continue = true;
3736                 }
3737                 else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) && key_intact)
3738                 {
3739                         /*
3740                          * If it's just a key-share locker, and we're not changing the key
3741                          * columns, we don't need to wait for it to end; but we need to
3742                          * preserve it as locker.
3743                          */
3744                         checked_lockers = true;
3745                         locker_remains = true;
3746                         can_continue = true;
3747                 }
3748                 else
3749                 {
3750                         /*
3751                          * Wait for regular transaction to end; but first, acquire tuple
3752                          * lock.
3753                          */
3754                         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3755                         heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3756                                                                  LockWaitBlock, &have_tuple_lock);
3757                         XactLockTableWait(xwait, relation, &oldtup.t_self,
3758                                                           XLTW_Update);
3759                         checked_lockers = true;
3760                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3761
3762                         /*
3763                          * xwait is done, but if xwait had just locked the tuple then some
3764                          * other xact could update this tuple before we get to this point.
3765                          * Check for xmax change, and start over if so.
3766                          */
3767                         if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3768                                 !TransactionIdEquals(xwait,
3769                                                                    HeapTupleHeaderGetRawXmax(oldtup.t_data)))
3770                                 goto l2;
3771
3772                         /* Otherwise check if it committed or aborted */
3773                         UpdateXmaxHintBits(oldtup.t_data, buffer, xwait);
3774                         if (oldtup.t_data->t_infomask & HEAP_XMAX_INVALID)
3775                                 can_continue = true;
3776                 }
3777
3778                 result = can_continue ? HeapTupleMayBeUpdated : HeapTupleUpdated;
3779         }
3780
3781         if (crosscheck != InvalidSnapshot && result == HeapTupleMayBeUpdated)
3782         {
3783                 /* Perform additional check for transaction-snapshot mode RI updates */
3784                 if (!HeapTupleSatisfiesVisibility(&oldtup, crosscheck, buffer))
3785                         result = HeapTupleUpdated;
3786         }
3787
3788         if (result != HeapTupleMayBeUpdated)
3789         {
3790                 Assert(result == HeapTupleSelfUpdated ||
3791                            result == HeapTupleUpdated ||
3792                            result == HeapTupleBeingUpdated);
3793                 Assert(!(oldtup.t_data->t_infomask & HEAP_XMAX_INVALID));
3794                 hufd->ctid = oldtup.t_data->t_ctid;
3795                 hufd->xmax = HeapTupleHeaderGetUpdateXid(oldtup.t_data);
3796                 if (result == HeapTupleSelfUpdated)
3797                         hufd->cmax = HeapTupleHeaderGetCmax(oldtup.t_data);
3798                 else
3799                         hufd->cmax = InvalidCommandId;
3800                 UnlockReleaseBuffer(buffer);
3801                 if (have_tuple_lock)
3802                         UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3803                 if (vmbuffer != InvalidBuffer)
3804                         ReleaseBuffer(vmbuffer);
3805                 bms_free(hot_attrs);
3806                 bms_free(key_attrs);
3807                 return result;
3808         }
3809
3810         /*
3811          * If we didn't pin the visibility map page and the page has become all
3812          * visible while we were busy locking the buffer, or during some
3813          * subsequent window during which we had it unlocked, we'll have to unlock
3814          * and re-lock, to avoid holding the buffer lock across an I/O.  That's a
3815          * bit unfortunate, especially since we'll now have to recheck whether the
3816          * tuple has been locked or updated under us, but hopefully it won't
3817          * happen very often.
3818          */
3819         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3820         {
3821                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3822                 visibilitymap_pin(relation, block, &vmbuffer);
3823                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3824                 goto l2;
3825         }
3826
3827         /* Fill in transaction status data */
3828
3829         /*
3830          * If the tuple we're updating is locked, we need to preserve the locking
3831          * info in the old tuple's Xmax.  Prepare a new Xmax value for this.
3832          */
3833         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3834                                                           oldtup.t_data->t_infomask,
3835                                                           oldtup.t_data->t_infomask2,
3836                                                           xid, *lockmode, true,
3837                                                           &xmax_old_tuple, &infomask_old_tuple,
3838                                                           &infomask2_old_tuple);
3839
3840         /*
3841          * And also prepare an Xmax value for the new copy of the tuple.  If there
3842          * was no xmax previously, or there was one but all lockers are now gone,
3843          * then use InvalidXid; otherwise, get the xmax from the old tuple.  (In
3844          * rare cases that might also be InvalidXid and yet not have the
3845          * HEAP_XMAX_INVALID bit set; that's fine.)
3846          */
3847         if ((oldtup.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3848                 (checked_lockers && !locker_remains))
3849                 xmax_new_tuple = InvalidTransactionId;
3850         else
3851                 xmax_new_tuple = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3852
3853         if (!TransactionIdIsValid(xmax_new_tuple))
3854         {
3855                 infomask_new_tuple = HEAP_XMAX_INVALID;
3856                 infomask2_new_tuple = 0;
3857         }
3858         else
3859         {
3860                 /*
3861                  * If we found a valid Xmax for the new tuple, then the infomask bits
3862                  * to use on the new tuple depend on what was there on the old one.
3863                  * Note that since we're doing an update, the only possibility is that
3864                  * the lockers had FOR KEY SHARE lock.
3865                  */
3866                 if (oldtup.t_data->t_infomask & HEAP_XMAX_IS_MULTI)
3867                 {
3868                         GetMultiXactIdHintBits(xmax_new_tuple, &infomask_new_tuple,
3869                                                                    &infomask2_new_tuple);
3870                 }
3871                 else
3872                 {
3873                         infomask_new_tuple = HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_LOCK_ONLY;
3874                         infomask2_new_tuple = 0;
3875                 }
3876         }
3877
3878         /*
3879          * Prepare the new tuple with the appropriate initial values of Xmin and
3880          * Xmax, as well as initial infomask bits as computed above.
3881          */
3882         newtup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
3883         newtup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
3884         HeapTupleHeaderSetXmin(newtup->t_data, xid);
3885         HeapTupleHeaderSetCmin(newtup->t_data, cid);
3886         newtup->t_data->t_infomask |= HEAP_UPDATED | infomask_new_tuple;
3887         newtup->t_data->t_infomask2 |= infomask2_new_tuple;
3888         HeapTupleHeaderSetXmax(newtup->t_data, xmax_new_tuple);
3889
3890         /*
3891          * Replace cid with a combo cid if necessary.  Note that we already put
3892          * the plain cid into the new tuple.
3893          */
3894         HeapTupleHeaderAdjustCmax(oldtup.t_data, &cid, &iscombo);
3895
3896         /*
3897          * If the toaster needs to be activated, OR if the new tuple will not fit
3898          * on the same page as the old, then we need to release the content lock
3899          * (but not the pin!) on the old tuple's buffer while we are off doing
3900          * TOAST and/or table-file-extension work.  We must mark the old tuple to
3901          * show that it's already being updated, else other processes may try to
3902          * update it themselves.
3903          *
3904          * We need to invoke the toaster if there are already any out-of-line
3905          * toasted values present, or if the new tuple is over-threshold.
3906          */
3907         if (relation->rd_rel->relkind != RELKIND_RELATION &&
3908                 relation->rd_rel->relkind != RELKIND_MATVIEW)
3909         {
3910                 /* toast table entries should never be recursively toasted */
3911                 Assert(!HeapTupleHasExternal(&oldtup));
3912                 Assert(!HeapTupleHasExternal(newtup));
3913                 need_toast = false;
3914         }
3915         else
3916                 need_toast = (HeapTupleHasExternal(&oldtup) ||
3917                                           HeapTupleHasExternal(newtup) ||
3918                                           newtup->t_len > TOAST_TUPLE_THRESHOLD);
3919
3920         pagefree = PageGetHeapFreeSpace(page);
3921
3922         newtupsize = MAXALIGN(newtup->t_len);
3923
3924         if (need_toast || newtupsize > pagefree)
3925         {
3926                 /* Clear obsolete visibility flags ... */
3927                 oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3928                 oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3929                 HeapTupleClearHotUpdated(&oldtup);
3930                 /* ... and store info about transaction updating this tuple */
3931                 Assert(TransactionIdIsValid(xmax_old_tuple));
3932                 HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
3933                 oldtup.t_data->t_infomask |= infomask_old_tuple;
3934                 oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
3935                 HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3936                 /* temporarily make it look not-updated */
3937                 oldtup.t_data->t_ctid = oldtup.t_self;
3938                 already_marked = true;
3939                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3940
3941                 /*
3942                  * Let the toaster do its thing, if needed.
3943                  *
3944                  * Note: below this point, heaptup is the data we actually intend to
3945                  * store into the relation; newtup is the caller's original untoasted
3946                  * data.
3947                  */
3948                 if (need_toast)
3949                 {
3950                         /* Note we always use WAL and FSM during updates */
3951                         heaptup = toast_insert_or_update(relation, newtup, &oldtup, 0);
3952                         newtupsize = MAXALIGN(heaptup->t_len);
3953                 }
3954                 else
3955                         heaptup = newtup;
3956
3957                 /*
3958                  * Now, do we need a new page for the tuple, or not?  This is a bit
3959                  * tricky since someone else could have added tuples to the page while
3960                  * we weren't looking.  We have to recheck the available space after
3961                  * reacquiring the buffer lock.  But don't bother to do that if the
3962                  * former amount of free space is still not enough; it's unlikely
3963                  * there's more free now than before.
3964                  *
3965                  * What's more, if we need to get a new page, we will need to acquire
3966                  * buffer locks on both old and new pages.  To avoid deadlock against
3967                  * some other backend trying to get the same two locks in the other
3968                  * order, we must be consistent about the order we get the locks in.
3969                  * We use the rule "lock the lower-numbered page of the relation
3970                  * first".  To implement this, we must do RelationGetBufferForTuple
3971                  * while not holding the lock on the old page, and we must rely on it
3972                  * to get the locks on both pages in the correct order.
3973                  */
3974                 if (newtupsize > pagefree)
3975                 {
3976                         /* Assume there's no chance to put heaptup on same page. */
3977                         newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3978                                                                                            buffer, 0, NULL,
3979                                                                                            &vmbuffer_new, &vmbuffer);
3980                 }
3981                 else
3982                 {
3983                         /* Re-acquire the lock on the old tuple's page. */
3984                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3985                         /* Re-check using the up-to-date free space */
3986                         pagefree = PageGetHeapFreeSpace(page);
3987                         if (newtupsize > pagefree)
3988                         {
3989                                 /*
3990                                  * Rats, it doesn't fit anymore.  We must now unlock and
3991                                  * relock to avoid deadlock.  Fortunately, this path should
3992                                  * seldom be taken.
3993                                  */
3994                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3995                                 newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3996                                                                                                    buffer, 0, NULL,
3997                                                                                                    &vmbuffer_new, &vmbuffer);
3998                         }
3999                         else
4000                         {
4001                                 /* OK, it fits here, so we're done. */
4002                                 newbuf = buffer;
4003                         }
4004                 }
4005         }
4006         else
4007         {
4008                 /* No TOAST work needed, and it'll fit on same page */
4009                 already_marked = false;
4010                 newbuf = buffer;
4011                 heaptup = newtup;
4012         }
4013
4014         /*
4015          * We're about to do the actual update -- check for conflict first, to
4016          * avoid possibly having to roll back work we've just done.
4017          *
4018          * This is safe without a recheck as long as there is no possibility of
4019          * another process scanning the pages between this check and the update
4020          * being visible to the scan (i.e., exclusive buffer content lock(s) are
4021          * continuously held from this point until the tuple update is visible).
4022          *
4023          * For the new tuple the only check needed is at the relation level, but
4024          * since both tuples are in the same relation and the check for oldtup
4025          * will include checking the relation level, there is no benefit to a
4026          * separate check for the new tuple.
4027          */
4028         CheckForSerializableConflictIn(relation, &oldtup, buffer);
4029
4030         /*
4031          * At this point newbuf and buffer are both pinned and locked, and newbuf
4032          * has enough space for the new tuple.  If they are the same buffer, only
4033          * one pin is held.
4034          */
4035
4036         if (newbuf == buffer)
4037         {
4038                 /*
4039                  * Since the new tuple is going into the same page, we might be able
4040                  * to do a HOT update.  Check if any of the index columns have been
4041                  * changed.  If not, then HOT update is possible.
4042                  */
4043                 if (satisfies_hot)
4044                         use_hot_update = true;
4045         }
4046         else
4047         {
4048                 /* Set a hint that the old page could use prune/defrag */
4049                 PageSetFull(page);
4050         }
4051
4052         /*
4053          * Compute replica identity tuple before entering the critical section so
4054          * we don't PANIC upon a memory allocation failure.
4055          * ExtractReplicaIdentity() will return NULL if nothing needs to be
4056          * logged.
4057          */
4058         old_key_tuple = ExtractReplicaIdentity(relation, &oldtup, !satisfies_id, &old_key_copied);
4059
4060         /* NO EREPORT(ERROR) from here till changes are logged */
4061         START_CRIT_SECTION();
4062
4063         /*
4064          * If this transaction commits, the old tuple will become DEAD sooner or
4065          * later.  Set flag that this page is a candidate for pruning once our xid
4066          * falls below the OldestXmin horizon.  If the transaction finally aborts,
4067          * the subsequent page pruning will be a no-op and the hint will be
4068          * cleared.
4069          *
4070          * XXX Should we set hint on newbuf as well?  If the transaction aborts,
4071          * there would be a prunable tuple in the newbuf; but for now we choose
4072          * not to optimize for aborts.  Note that heap_xlog_update must be kept in
4073          * sync if this decision changes.
4074          */
4075         PageSetPrunable(page, xid);
4076
4077         if (use_hot_update)
4078         {
4079                 /* Mark the old tuple as HOT-updated */
4080                 HeapTupleSetHotUpdated(&oldtup);
4081                 /* And mark the new tuple as heap-only */
4082                 HeapTupleSetHeapOnly(heaptup);
4083                 /* Mark the caller's copy too, in case different from heaptup */
4084                 HeapTupleSetHeapOnly(newtup);
4085         }
4086         else
4087         {
4088                 /* Make sure tuples are correctly marked as not-HOT */
4089                 HeapTupleClearHotUpdated(&oldtup);
4090                 HeapTupleClearHeapOnly(heaptup);
4091                 HeapTupleClearHeapOnly(newtup);
4092         }
4093
4094         RelationPutHeapTuple(relation, newbuf, heaptup, false);         /* insert new tuple */
4095
4096         if (!already_marked)
4097         {
4098                 /* Clear obsolete visibility flags ... */
4099                 oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
4100                 oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
4101                 /* ... and store info about transaction updating this tuple */
4102                 Assert(TransactionIdIsValid(xmax_old_tuple));
4103                 HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
4104                 oldtup.t_data->t_infomask |= infomask_old_tuple;
4105                 oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
4106                 HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
4107         }
4108
4109         /* record address of new tuple in t_ctid of old one */
4110         oldtup.t_data->t_ctid = heaptup->t_self;
4111
4112         /* clear PD_ALL_VISIBLE flags */
4113         if (PageIsAllVisible(BufferGetPage(buffer)))
4114         {
4115                 all_visible_cleared = true;
4116                 PageClearAllVisible(BufferGetPage(buffer));
4117                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
4118                                                         vmbuffer);
4119         }
4120         if (newbuf != buffer && PageIsAllVisible(BufferGetPage(newbuf)))
4121         {
4122                 all_visible_cleared_new = true;
4123                 PageClearAllVisible(BufferGetPage(newbuf));
4124                 visibilitymap_clear(relation, BufferGetBlockNumber(newbuf),
4125                                                         vmbuffer_new);
4126         }
4127
4128         if (newbuf != buffer)
4129                 MarkBufferDirty(newbuf);
4130         MarkBufferDirty(buffer);
4131
4132         /* XLOG stuff */
4133         if (RelationNeedsWAL(relation))
4134         {
4135                 XLogRecPtr      recptr;
4136
4137                 /*
4138                  * For logical decoding we need combocids to properly decode the
4139                  * catalog.
4140                  */
4141                 if (RelationIsAccessibleInLogicalDecoding(relation))
4142                 {
4143                         log_heap_new_cid(relation, &oldtup);
4144                         log_heap_new_cid(relation, heaptup);
4145                 }
4146
4147                 recptr = log_heap_update(relation, buffer,
4148                                                                  newbuf, &oldtup, heaptup,
4149                                                                  old_key_tuple,
4150                                                                  all_visible_cleared,
4151                                                                  all_visible_cleared_new);
4152                 if (newbuf != buffer)
4153                 {
4154                         PageSetLSN(BufferGetPage(newbuf), recptr);
4155                 }
4156                 PageSetLSN(BufferGetPage(buffer), recptr);
4157         }
4158
4159         END_CRIT_SECTION();
4160
4161         if (newbuf != buffer)
4162                 LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
4163         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
4164
4165         /*
4166          * Mark old tuple for invalidation from system caches at next command
4167          * boundary, and mark the new tuple for invalidation in case we abort. We
4168          * have to do this before releasing the buffer because oldtup is in the
4169          * buffer.  (heaptup is all in local memory, but it's necessary to process
4170          * both tuple versions in one call to inval.c so we can avoid redundant
4171          * sinval messages.)
4172          */
4173         CacheInvalidateHeapTuple(relation, &oldtup, heaptup);
4174
4175         /* Now we can release the buffer(s) */
4176         if (newbuf != buffer)
4177                 ReleaseBuffer(newbuf);
4178         ReleaseBuffer(buffer);
4179         if (BufferIsValid(vmbuffer_new))
4180                 ReleaseBuffer(vmbuffer_new);
4181         if (BufferIsValid(vmbuffer))
4182                 ReleaseBuffer(vmbuffer);
4183
4184         /*
4185          * Release the lmgr tuple lock, if we had it.
4186          */
4187         if (have_tuple_lock)
4188                 UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
4189
4190         pgstat_count_heap_update(relation, use_hot_update);
4191
4192         /*
4193          * If heaptup is a private copy, release it.  Don't forget to copy t_self
4194          * back to the caller's image, too.
4195          */
4196         if (heaptup != newtup)
4197         {
4198                 newtup->t_self = heaptup->t_self;
4199                 heap_freetuple(heaptup);
4200         }
4201
4202         if (old_key_tuple != NULL && old_key_copied)
4203                 heap_freetuple(old_key_tuple);
4204
4205         bms_free(hot_attrs);
4206         bms_free(key_attrs);
4207
4208         return HeapTupleMayBeUpdated;
4209 }
4210
4211 /*
4212  * Check if the specified attribute's value is same in both given tuples.
4213  * Subroutine for HeapSatisfiesHOTandKeyUpdate.
4214  */
4215 static bool
4216 heap_tuple_attr_equals(TupleDesc tupdesc, int attrnum,
4217                                            HeapTuple tup1, HeapTuple tup2)
4218 {
4219         Datum           value1,
4220                                 value2;
4221         bool            isnull1,
4222                                 isnull2;
4223         Form_pg_attribute att;
4224
4225         /*
4226          * If it's a whole-tuple reference, say "not equal".  It's not really
4227          * worth supporting this case, since it could only succeed after a no-op
4228          * update, which is hardly a case worth optimizing for.
4229          */
4230         if (attrnum == 0)
4231                 return false;
4232
4233         /*
4234          * Likewise, automatically say "not equal" for any system attribute other
4235          * than OID and tableOID; we cannot expect these to be consistent in a HOT
4236          * chain, or even to be set correctly yet in the new tuple.
4237          */
4238         if (attrnum < 0)
4239         {
4240                 if (attrnum != ObjectIdAttributeNumber &&
4241                         attrnum != TableOidAttributeNumber)
4242                         return false;
4243         }
4244
4245         /*
4246          * Extract the corresponding values.  XXX this is pretty inefficient if
4247          * there are many indexed columns.  Should HeapSatisfiesHOTandKeyUpdate do
4248          * a single heap_deform_tuple call on each tuple, instead?      But that
4249          * doesn't work for system columns ...
4250          */
4251         value1 = heap_getattr(tup1, attrnum, tupdesc, &isnull1);
4252         value2 = heap_getattr(tup2, attrnum, tupdesc, &isnull2);
4253
4254         /*
4255          * If one value is NULL and other is not, then they are certainly not
4256          * equal
4257          */
4258         if (isnull1 != isnull2)
4259                 return false;
4260
4261         /*
4262          * If both are NULL, they can be considered equal.
4263          */
4264         if (isnull1)
4265                 return true;
4266
4267         /*
4268          * We do simple binary comparison of the two datums.  This may be overly
4269          * strict because there can be multiple binary representations for the
4270          * same logical value.  But we should be OK as long as there are no false
4271          * positives.  Using a type-specific equality operator is messy because
4272          * there could be multiple notions of equality in different operator
4273          * classes; furthermore, we cannot safely invoke user-defined functions
4274          * while holding exclusive buffer lock.
4275          */
4276         if (attrnum <= 0)
4277         {
4278                 /* The only allowed system columns are OIDs, so do this */
4279                 return (DatumGetObjectId(value1) == DatumGetObjectId(value2));
4280         }
4281         else
4282         {
4283                 Assert(attrnum <= tupdesc->natts);
4284                 att = tupdesc->attrs[attrnum - 1];
4285                 return datumIsEqual(value1, value2, att->attbyval, att->attlen);
4286         }
4287 }
4288
4289 /*
4290  * Check which columns are being updated.
4291  *
4292  * This simultaneously checks conditions for HOT updates, for FOR KEY
4293  * SHARE updates, and REPLICA IDENTITY concerns.  Since much of the time they
4294  * will be checking very similar sets of columns, and doing the same tests on
4295  * them, it makes sense to optimize and do them together.
4296  *
4297  * We receive three bitmapsets comprising the three sets of columns we're
4298  * interested in.  Note these are destructively modified; that is OK since
4299  * this is invoked at most once in heap_update.
4300  *
4301  * hot_result is set to TRUE if it's okay to do a HOT update (i.e. it does not
4302  * modified indexed columns); key_result is set to TRUE if the update does not
4303  * modify columns used in the key; id_result is set to TRUE if the update does
4304  * not modify columns in any index marked as the REPLICA IDENTITY.
4305  */
4306 static void
4307 HeapSatisfiesHOTandKeyUpdate(Relation relation, Bitmapset *hot_attrs,
4308                                                          Bitmapset *key_attrs, Bitmapset *id_attrs,
4309                                                          bool *satisfies_hot, bool *satisfies_key,
4310                                                          bool *satisfies_id,
4311                                                          HeapTuple oldtup, HeapTuple newtup)
4312 {
4313         int                     next_hot_attnum;
4314         int                     next_key_attnum;
4315         int                     next_id_attnum;
4316         bool            hot_result = true;
4317         bool            key_result = true;
4318         bool            id_result = true;
4319
4320         /* If REPLICA IDENTITY is set to FULL, id_attrs will be empty. */
4321         Assert(bms_is_subset(id_attrs, key_attrs));
4322         Assert(bms_is_subset(key_attrs, hot_attrs));
4323
4324         /*
4325          * If one of these sets contains no remaining bits, bms_first_member will
4326          * return -1, and after adding FirstLowInvalidHeapAttributeNumber (which
4327          * is negative!)  we'll get an attribute number that can't possibly be
4328          * real, and thus won't match any actual attribute number.
4329          */
4330         next_hot_attnum = bms_first_member(hot_attrs);
4331         next_hot_attnum += FirstLowInvalidHeapAttributeNumber;
4332         next_key_attnum = bms_first_member(key_attrs);
4333         next_key_attnum += FirstLowInvalidHeapAttributeNumber;
4334         next_id_attnum = bms_first_member(id_attrs);
4335         next_id_attnum += FirstLowInvalidHeapAttributeNumber;
4336
4337         for (;;)
4338         {
4339                 bool            changed;
4340                 int                     check_now;
4341
4342                 /*
4343                  * Since the HOT attributes are a superset of the key attributes and
4344                  * the key attributes are a superset of the id attributes, this logic
4345                  * is guaranteed to identify the next column that needs to be checked.
4346                  */
4347                 if (hot_result && next_hot_attnum > FirstLowInvalidHeapAttributeNumber)
4348                         check_now = next_hot_attnum;
4349                 else if (key_result && next_key_attnum > FirstLowInvalidHeapAttributeNumber)
4350                         check_now = next_key_attnum;
4351                 else if (id_result && next_id_attnum > FirstLowInvalidHeapAttributeNumber)
4352                         check_now = next_id_attnum;
4353                 else
4354                         break;
4355
4356                 /* See whether it changed. */
4357                 changed = !heap_tuple_attr_equals(RelationGetDescr(relation),
4358                                                                                   check_now, oldtup, newtup);
4359                 if (changed)
4360                 {
4361                         if (check_now == next_hot_attnum)
4362                                 hot_result = false;
4363                         if (check_now == next_key_attnum)
4364                                 key_result = false;
4365                         if (check_now == next_id_attnum)
4366                                 id_result = false;
4367
4368                         /* if all are false now, we can stop checking */
4369                         if (!hot_result && !key_result && !id_result)
4370                                 break;
4371                 }
4372
4373                 /*
4374                  * Advance the next attribute numbers for the sets that contain the
4375                  * attribute we just checked.  As we work our way through the columns,
4376                  * the next_attnum values will rise; but when each set becomes empty,
4377                  * bms_first_member() will return -1 and the attribute number will end
4378                  * up with a value less than FirstLowInvalidHeapAttributeNumber.
4379                  */
4380                 if (hot_result && check_now == next_hot_attnum)
4381                 {
4382                         next_hot_attnum = bms_first_member(hot_attrs);
4383                         next_hot_attnum += FirstLowInvalidHeapAttributeNumber;
4384                 }
4385                 if (key_result && check_now == next_key_attnum)
4386                 {
4387                         next_key_attnum = bms_first_member(key_attrs);
4388                         next_key_attnum += FirstLowInvalidHeapAttributeNumber;
4389                 }
4390                 if (id_result && check_now == next_id_attnum)
4391                 {
4392                         next_id_attnum = bms_first_member(id_attrs);
4393                         next_id_attnum += FirstLowInvalidHeapAttributeNumber;
4394                 }
4395         }
4396
4397         *satisfies_hot = hot_result;
4398         *satisfies_key = key_result;
4399         *satisfies_id = id_result;
4400 }
4401
4402 /*
4403  *      simple_heap_update - replace a tuple
4404  *
4405  * This routine may be used to update a tuple when concurrent updates of
4406  * the target tuple are not expected (for example, because we have a lock
4407  * on the relation associated with the tuple).  Any failure is reported
4408  * via ereport().
4409  */
4410 void
4411 simple_heap_update(Relation relation, ItemPointer otid, HeapTuple tup)
4412 {
4413         HTSU_Result result;
4414         HeapUpdateFailureData hufd;
4415         LockTupleMode lockmode;
4416
4417         result = heap_update(relation, otid, tup,
4418                                                  GetCurrentCommandId(true), InvalidSnapshot,
4419                                                  true /* wait for commit */ ,
4420                                                  &hufd, &lockmode);
4421         switch (result)
4422         {
4423                 case HeapTupleSelfUpdated:
4424                         /* Tuple was already updated in current command? */
4425                         elog(ERROR, "tuple already updated by self");
4426                         break;
4427
4428                 case HeapTupleMayBeUpdated:
4429                         /* done successfully */
4430                         break;
4431
4432                 case HeapTupleUpdated:
4433                         elog(ERROR, "tuple concurrently updated");
4434                         break;
4435
4436                 default:
4437                         elog(ERROR, "unrecognized heap_update status: %u", result);
4438                         break;
4439         }
4440 }
4441
4442
4443 /*
4444  * Return the MultiXactStatus corresponding to the given tuple lock mode.
4445  */
4446 static MultiXactStatus
4447 get_mxact_status_for_lock(LockTupleMode mode, bool is_update)
4448 {
4449         int                     retval;
4450
4451         if (is_update)
4452                 retval = tupleLockExtraInfo[mode].updstatus;
4453         else
4454                 retval = tupleLockExtraInfo[mode].lockstatus;
4455
4456         if (retval == -1)
4457                 elog(ERROR, "invalid lock tuple mode %d/%s", mode,
4458                          is_update ? "true" : "false");
4459
4460         return (MultiXactStatus) retval;
4461 }
4462
4463 /*
4464  *      heap_lock_tuple - lock a tuple in shared or exclusive mode
4465  *
4466  * Note that this acquires a buffer pin, which the caller must release.
4467  *
4468  * Input parameters:
4469  *      relation: relation containing tuple (caller must hold suitable lock)
4470  *      tuple->t_self: TID of tuple to lock (rest of struct need not be valid)
4471  *      cid: current command ID (used for visibility test, and stored into
4472  *              tuple's cmax if lock is successful)
4473  *      mode: indicates if shared or exclusive tuple lock is desired
4474  *      wait_policy: what to do if tuple lock is not available
4475  *      follow_updates: if true, follow the update chain to also lock descendant
4476  *              tuples.
4477  *
4478  * Output parameters:
4479  *      *tuple: all fields filled in
4480  *      *buffer: set to buffer holding tuple (pinned but not locked at exit)
4481  *      *hufd: filled in failure cases (see below)
4482  *
4483  * Function result may be:
4484  *      HeapTupleMayBeUpdated: lock was successfully acquired
4485  *      HeapTupleInvisible: lock failed because tuple was never visible to us
4486  *      HeapTupleSelfUpdated: lock failed because tuple updated by self
4487  *      HeapTupleUpdated: lock failed because tuple updated by other xact
4488  *      HeapTupleWouldBlock: lock couldn't be acquired and wait_policy is skip
4489  *
4490  * In the failure cases other than HeapTupleInvisible, the routine fills
4491  * *hufd with the tuple's t_ctid, t_xmax (resolving a possible MultiXact,
4492  * if necessary), and t_cmax (the last only for HeapTupleSelfUpdated,
4493  * since we cannot obtain cmax from a combocid generated by another
4494  * transaction).
4495  * See comments for struct HeapUpdateFailureData for additional info.
4496  *
4497  * See README.tuplock for a thorough explanation of this mechanism.
4498  */
4499 HTSU_Result
4500 heap_lock_tuple(Relation relation, HeapTuple tuple,
4501                                 CommandId cid, LockTupleMode mode, LockWaitPolicy wait_policy,
4502                                 bool follow_updates,
4503                                 Buffer *buffer, HeapUpdateFailureData *hufd)
4504 {
4505         HTSU_Result result;
4506         ItemPointer tid = &(tuple->t_self);
4507         ItemId          lp;
4508         Page            page;
4509         TransactionId xid,
4510                                 xmax;
4511         uint16          old_infomask,
4512                                 new_infomask,
4513                                 new_infomask2;
4514         bool            first_time = true;
4515         bool            have_tuple_lock = false;
4516
4517         *buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
4518         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4519
4520         page = BufferGetPage(*buffer);
4521         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
4522         Assert(ItemIdIsNormal(lp));
4523
4524         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
4525         tuple->t_len = ItemIdGetLength(lp);
4526         tuple->t_tableOid = RelationGetRelid(relation);
4527
4528 l3:
4529         result = HeapTupleSatisfiesUpdate(tuple, cid, *buffer);
4530
4531         if (result == HeapTupleInvisible)
4532         {
4533                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4534
4535                 /*
4536                  * This is possible, but only when locking a tuple for ON CONFLICT
4537                  * UPDATE.  We return this value here rather than throwing an error in
4538                  * order to give that case the opportunity to throw a more specific
4539                  * error.
4540                  */
4541                 return HeapTupleInvisible;
4542         }
4543         else if (result == HeapTupleBeingUpdated)
4544         {
4545                 TransactionId xwait;
4546                 uint16          infomask;
4547                 uint16          infomask2;
4548                 bool            require_sleep;
4549                 ItemPointerData t_ctid;
4550
4551                 /* must copy state data before unlocking buffer */
4552                 xwait = HeapTupleHeaderGetRawXmax(tuple->t_data);
4553                 infomask = tuple->t_data->t_infomask;
4554                 infomask2 = tuple->t_data->t_infomask2;
4555                 ItemPointerCopy(&tuple->t_data->t_ctid, &t_ctid);
4556
4557                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4558
4559                 /*
4560                  * If any subtransaction of the current top transaction already holds
4561                  * a lock as strong as or stronger than what we're requesting, we
4562                  * effectively hold the desired lock already.  We *must* succeed
4563                  * without trying to take the tuple lock, else we will deadlock
4564                  * against anyone wanting to acquire a stronger lock.
4565                  *
4566                  * Note we only do this the first time we loop on the HTSU result;
4567                  * there is no point in testing in subsequent passes, because
4568                  * evidently our own transaction cannot have acquired a new lock after
4569                  * the first time we checked.
4570                  */
4571                 if (first_time)
4572                 {
4573                         first_time = false;
4574
4575                         if (infomask & HEAP_XMAX_IS_MULTI)
4576                         {
4577                                 int                     i;
4578                                 int                     nmembers;
4579                                 MultiXactMember *members;
4580
4581                                 /*
4582                                  * We don't need to allow old multixacts here; if that had
4583                                  * been the case, HeapTupleSatisfiesUpdate would have returned
4584                                  * MayBeUpdated and we wouldn't be here.
4585                                  */
4586                                 nmembers =
4587                                         GetMultiXactIdMembers(xwait, &members, false,
4588                                                                                   HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4589
4590                                 for (i = 0; i < nmembers; i++)
4591                                 {
4592                                         /* only consider members of our own transaction */
4593                                         if (!TransactionIdIsCurrentTransactionId(members[i].xid))
4594                                                 continue;
4595
4596                                         if (TUPLOCK_from_mxstatus(members[i].status) >= mode)
4597                                         {
4598                                                 pfree(members);
4599                                                 return HeapTupleMayBeUpdated;
4600                                         }
4601                                 }
4602
4603                                 if (members)
4604                                         pfree(members);
4605                         }
4606                         else if (TransactionIdIsCurrentTransactionId(xwait))
4607                         {
4608                                 switch (mode)
4609                                 {
4610                                         case LockTupleKeyShare:
4611                                                 Assert(HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) ||
4612                                                            HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4613                                                            HEAP_XMAX_IS_EXCL_LOCKED(infomask));
4614                                                 return HeapTupleMayBeUpdated;
4615                                                 break;
4616                                         case LockTupleShare:
4617                                                 if (HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4618                                                         HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4619                                                         return HeapTupleMayBeUpdated;
4620                                                 break;
4621                                         case LockTupleNoKeyExclusive:
4622                                                 if (HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4623                                                         return HeapTupleMayBeUpdated;
4624                                                 break;
4625                                         case LockTupleExclusive:
4626                                                 if (HEAP_XMAX_IS_EXCL_LOCKED(infomask) &&
4627                                                         infomask2 & HEAP_KEYS_UPDATED)
4628                                                         return HeapTupleMayBeUpdated;
4629                                                 break;
4630                                 }
4631                         }
4632                 }
4633
4634                 /*
4635                  * Initially assume that we will have to wait for the locking
4636                  * transaction(s) to finish.  We check various cases below in which
4637                  * this can be turned off.
4638                  */
4639                 require_sleep = true;
4640                 if (mode == LockTupleKeyShare)
4641                 {
4642                         /*
4643                          * If we're requesting KeyShare, and there's no update present, we
4644                          * don't need to wait.  Even if there is an update, we can still
4645                          * continue if the key hasn't been modified.
4646                          *
4647                          * However, if there are updates, we need to walk the update chain
4648                          * to mark future versions of the row as locked, too.  That way,
4649                          * if somebody deletes that future version, we're protected
4650                          * against the key going away.  This locking of future versions
4651                          * could block momentarily, if a concurrent transaction is
4652                          * deleting a key; or it could return a value to the effect that
4653                          * the transaction deleting the key has already committed.  So we
4654                          * do this before re-locking the buffer; otherwise this would be
4655                          * prone to deadlocks.
4656                          *
4657                          * Note that the TID we're locking was grabbed before we unlocked
4658                          * the buffer.  For it to change while we're not looking, the
4659                          * other properties we're testing for below after re-locking the
4660                          * buffer would also change, in which case we would restart this
4661                          * loop above.
4662                          */
4663                         if (!(infomask2 & HEAP_KEYS_UPDATED))
4664                         {
4665                                 bool            updated;
4666
4667                                 updated = !HEAP_XMAX_IS_LOCKED_ONLY(infomask);
4668
4669                                 /*
4670                                  * If there are updates, follow the update chain; bail out if
4671                                  * that cannot be done.
4672                                  */
4673                                 if (follow_updates && updated)
4674                                 {
4675                                         HTSU_Result res;
4676
4677                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4678                                                                                                   GetCurrentTransactionId(),
4679                                                                                                   mode);
4680                                         if (res != HeapTupleMayBeUpdated)
4681                                         {
4682                                                 result = res;
4683                                                 /* recovery code expects to have buffer lock held */
4684                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4685                                                 goto failed;
4686                                         }
4687                                 }
4688
4689                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4690
4691                                 /*
4692                                  * Make sure it's still an appropriate lock, else start over.
4693                                  * Also, if it wasn't updated before we released the lock, but
4694                                  * is updated now, we start over too; the reason is that we
4695                                  * now need to follow the update chain to lock the new
4696                                  * versions.
4697                                  */
4698                                 if (!HeapTupleHeaderIsOnlyLocked(tuple->t_data) &&
4699                                         ((tuple->t_data->t_infomask2 & HEAP_KEYS_UPDATED) ||
4700                                          !updated))
4701                                         goto l3;
4702
4703                                 /* Things look okay, so we can skip sleeping */
4704                                 require_sleep = false;
4705
4706                                 /*
4707                                  * Note we allow Xmax to change here; other updaters/lockers
4708                                  * could have modified it before we grabbed the buffer lock.
4709                                  * However, this is not a problem, because with the recheck we
4710                                  * just did we ensure that they still don't conflict with the
4711                                  * lock we want.
4712                                  */
4713                         }
4714                 }
4715                 else if (mode == LockTupleShare)
4716                 {
4717                         /*
4718                          * If we're requesting Share, we can similarly avoid sleeping if
4719                          * there's no update and no exclusive lock present.
4720                          */
4721                         if (HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
4722                                 !HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4723                         {
4724                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4725
4726                                 /*
4727                                  * Make sure it's still an appropriate lock, else start over.
4728                                  * See above about allowing xmax to change.
4729                                  */
4730                                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4731                                         HEAP_XMAX_IS_EXCL_LOCKED(tuple->t_data->t_infomask))
4732                                         goto l3;
4733                                 require_sleep = false;
4734                         }
4735                 }
4736                 else if (mode == LockTupleNoKeyExclusive)
4737                 {
4738                         /*
4739                          * If we're requesting NoKeyExclusive, we might also be able to
4740                          * avoid sleeping; just ensure that there no conflicting lock
4741                          * already acquired.
4742                          */
4743                         if (infomask & HEAP_XMAX_IS_MULTI)
4744                         {
4745                                 if (!DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
4746                                                                                          mode))
4747                                 {
4748                                         /*
4749                                          * No conflict, but if the xmax changed under us in the
4750                                          * meantime, start over.
4751                                          */
4752                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4753                                         if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4754                                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4755                                                                                          xwait))
4756                                                 goto l3;
4757
4758                                         /* otherwise, we're good */
4759                                         require_sleep = false;
4760                                 }
4761                         }
4762                         else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
4763                         {
4764                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4765
4766                                 /* if the xmax changed in the meantime, start over */
4767                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4768                                         !TransactionIdEquals(
4769                                                                         HeapTupleHeaderGetRawXmax(tuple->t_data),
4770                                                                                  xwait))
4771                                         goto l3;
4772                                 /* otherwise, we're good */
4773                                 require_sleep = false;
4774                         }
4775                 }
4776
4777                 /*
4778                  * As a check independent from those above, we can also avoid sleeping
4779                  * if the current transaction is the sole locker of the tuple.  Note
4780                  * that the strength of the lock already held is irrelevant; this is
4781                  * not about recording the lock in Xmax (which will be done regardless
4782                  * of this optimization, below).  Also, note that the cases where we
4783                  * hold a lock stronger than we are requesting are already handled
4784                  * above by not doing anything.
4785                  *
4786                  * Note we only deal with the non-multixact case here; MultiXactIdWait
4787                  * is well equipped to deal with this situation on its own.
4788                  */
4789                 if (require_sleep && !(infomask & HEAP_XMAX_IS_MULTI) &&
4790                         TransactionIdIsCurrentTransactionId(xwait))
4791                 {
4792                         /* ... but if the xmax changed in the meantime, start over */
4793                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4794                         if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4795                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4796                                                                          xwait))
4797                                 goto l3;
4798                         Assert(HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask));
4799                         require_sleep = false;
4800                 }
4801
4802                 /*
4803                  * By here, we either have already acquired the buffer exclusive lock,
4804                  * or we must wait for the locking transaction or multixact; so below
4805                  * we ensure that we grab buffer lock after the sleep.
4806                  */
4807
4808                 if (require_sleep)
4809                 {
4810                         /*
4811                          * Acquire tuple lock to establish our priority for the tuple, or
4812                          * die trying.  LockTuple will release us when we are next-in-line
4813                          * for the tuple.  We must do this even if we are share-locking.
4814                          *
4815                          * If we are forced to "start over" below, we keep the tuple lock;
4816                          * this arranges that we stay at the head of the line while
4817                          * rechecking tuple state.
4818                          */
4819                         if (!heap_acquire_tuplock(relation, tid, mode, wait_policy,
4820                                                                           &have_tuple_lock))
4821                         {
4822                                 /*
4823                                  * This can only happen if wait_policy is Skip and the lock
4824                                  * couldn't be obtained.
4825                                  */
4826                                 result = HeapTupleWouldBlock;
4827                                 /* recovery code expects to have buffer lock held */
4828                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4829                                 goto failed;
4830                         }
4831
4832                         if (infomask & HEAP_XMAX_IS_MULTI)
4833                         {
4834                                 MultiXactStatus status = get_mxact_status_for_lock(mode, false);
4835
4836                                 /* We only ever lock tuples, never update them */
4837                                 if (status >= MultiXactStatusNoKeyUpdate)
4838                                         elog(ERROR, "invalid lock mode in heap_lock_tuple");
4839
4840                                 /* wait for multixact to end, or die trying  */
4841                                 switch (wait_policy)
4842                                 {
4843                                         case LockWaitBlock:
4844                                                 MultiXactIdWait((MultiXactId) xwait, status, infomask,
4845                                                                   relation, &tuple->t_self, XLTW_Lock, NULL);
4846                                                 break;
4847                                         case LockWaitSkip:
4848                                                 if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4849                                                                                                   status, infomask, relation,
4850                                                                                                                 NULL))
4851                                                 {
4852                                                         result = HeapTupleWouldBlock;
4853                                                         /* recovery code expects to have buffer lock held */
4854                                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4855                                                         goto failed;
4856                                                 }
4857                                                 break;
4858                                         case LockWaitError:
4859                                                 if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4860                                                                                                   status, infomask, relation,
4861                                                                                                                 NULL))
4862                                                         ereport(ERROR,
4863                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4864                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
4865                                                                                 RelationGetRelationName(relation))));
4866
4867                                                 break;
4868                                 }
4869
4870                                 /*
4871                                  * Of course, the multixact might not be done here: if we're
4872                                  * requesting a light lock mode, other transactions with light
4873                                  * locks could still be alive, as well as locks owned by our
4874                                  * own xact or other subxacts of this backend.  We need to
4875                                  * preserve the surviving MultiXact members.  Note that it
4876                                  * isn't absolutely necessary in the latter case, but doing so
4877                                  * is simpler.
4878                                  */
4879                         }
4880                         else
4881                         {
4882                                 /* wait for regular transaction to end, or die trying */
4883                                 switch (wait_policy)
4884                                 {
4885                                         case LockWaitBlock:
4886                                                 XactLockTableWait(xwait, relation, &tuple->t_self,
4887                                                                                   XLTW_Lock);
4888                                                 break;
4889                                         case LockWaitSkip:
4890                                                 if (!ConditionalXactLockTableWait(xwait))
4891                                                 {
4892                                                         result = HeapTupleWouldBlock;
4893                                                         /* recovery code expects to have buffer lock held */
4894                                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4895                                                         goto failed;
4896                                                 }
4897                                                 break;
4898                                         case LockWaitError:
4899                                                 if (!ConditionalXactLockTableWait(xwait))
4900                                                         ereport(ERROR,
4901                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4902                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
4903                                                                                 RelationGetRelationName(relation))));
4904                                                 break;
4905                                 }
4906                         }
4907
4908                         /* if there are updates, follow the update chain */
4909                         if (follow_updates && !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
4910                         {
4911                                 HTSU_Result res;
4912
4913                                 res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4914                                                                                           GetCurrentTransactionId(),
4915                                                                                           mode);
4916                                 if (res != HeapTupleMayBeUpdated)
4917                                 {
4918                                         result = res;
4919                                         /* recovery code expects to have buffer lock held */
4920                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4921                                         goto failed;
4922                                 }
4923                         }
4924
4925                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4926
4927                         /*
4928                          * xwait is done, but if xwait had just locked the tuple then some
4929                          * other xact could update this tuple before we get to this point.
4930                          * Check for xmax change, and start over if so.
4931                          */
4932                         if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4933                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4934                                                                          xwait))
4935                                 goto l3;
4936
4937                         if (!(infomask & HEAP_XMAX_IS_MULTI))
4938                         {
4939                                 /*
4940                                  * Otherwise check if it committed or aborted.  Note we cannot
4941                                  * be here if the tuple was only locked by somebody who didn't
4942                                  * conflict with us; that would have been handled above.  So
4943                                  * that transaction must necessarily be gone by now.  But
4944                                  * don't check for this in the multixact case, because some
4945                                  * locker transactions might still be running.
4946                                  */
4947                                 UpdateXmaxHintBits(tuple->t_data, *buffer, xwait);
4948                         }
4949                 }
4950
4951                 /* By here, we're certain that we hold buffer exclusive lock again */
4952
4953                 /*
4954                  * We may lock if previous xmax aborted, or if it committed but only
4955                  * locked the tuple without updating it; or if we didn't have to wait
4956                  * at all for whatever reason.
4957                  */
4958                 if (!require_sleep ||
4959                         (tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
4960                         HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4961                         HeapTupleHeaderIsOnlyLocked(tuple->t_data))
4962                         result = HeapTupleMayBeUpdated;
4963                 else
4964                         result = HeapTupleUpdated;
4965         }
4966
4967 failed:
4968         if (result != HeapTupleMayBeUpdated)
4969         {
4970                 Assert(result == HeapTupleSelfUpdated || result == HeapTupleUpdated ||
4971                            result == HeapTupleWouldBlock);
4972                 Assert(!(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));
4973                 hufd->ctid = tuple->t_data->t_ctid;
4974                 hufd->xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
4975                 if (result == HeapTupleSelfUpdated)
4976                         hufd->cmax = HeapTupleHeaderGetCmax(tuple->t_data);
4977                 else
4978                         hufd->cmax = InvalidCommandId;
4979                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4980                 if (have_tuple_lock)
4981                         UnlockTupleTuplock(relation, tid, mode);
4982                 return result;
4983         }
4984
4985         xmax = HeapTupleHeaderGetRawXmax(tuple->t_data);
4986         old_infomask = tuple->t_data->t_infomask;
4987
4988         /*
4989          * If this is the first possibly-multixact-able operation in the current
4990          * transaction, set my per-backend OldestMemberMXactId setting. We can be
4991          * certain that the transaction will never become a member of any older
4992          * MultiXactIds than that.  (We have to do this even if we end up just
4993          * using our own TransactionId below, since some other backend could
4994          * incorporate our XID into a MultiXact immediately afterwards.)
4995          */
4996         MultiXactIdSetOldestMember();
4997
4998         /*
4999          * Compute the new xmax and infomask to store into the tuple.  Note we do
5000          * not modify the tuple just yet, because that would leave it in the wrong
5001          * state if multixact.c elogs.
5002          */
5003         compute_new_xmax_infomask(xmax, old_infomask, tuple->t_data->t_infomask2,
5004                                                           GetCurrentTransactionId(), mode, false,
5005                                                           &xid, &new_infomask, &new_infomask2);
5006
5007         START_CRIT_SECTION();
5008
5009         /*
5010          * Store transaction information of xact locking the tuple.
5011          *
5012          * Note: Cmax is meaningless in this context, so don't set it; this avoids
5013          * possibly generating a useless combo CID.  Moreover, if we're locking a
5014          * previously updated tuple, it's important to preserve the Cmax.
5015          *
5016          * Also reset the HOT UPDATE bit, but only if there's no update; otherwise
5017          * we would break the HOT chain.
5018          */
5019         tuple->t_data->t_infomask &= ~HEAP_XMAX_BITS;
5020         tuple->t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5021         tuple->t_data->t_infomask |= new_infomask;
5022         tuple->t_data->t_infomask2 |= new_infomask2;
5023         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5024                 HeapTupleHeaderClearHotUpdated(tuple->t_data);
5025         HeapTupleHeaderSetXmax(tuple->t_data, xid);
5026
5027         /*
5028          * Make sure there is no forward chain link in t_ctid.  Note that in the
5029          * cases where the tuple has been updated, we must not overwrite t_ctid,
5030          * because it was set by the updater.  Moreover, if the tuple has been
5031          * updated, we need to follow the update chain to lock the new versions of
5032          * the tuple as well.
5033          */
5034         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
5035                 tuple->t_data->t_ctid = *tid;
5036
5037         MarkBufferDirty(*buffer);
5038
5039         /*
5040          * XLOG stuff.  You might think that we don't need an XLOG record because
5041          * there is no state change worth restoring after a crash.  You would be
5042          * wrong however: we have just written either a TransactionId or a
5043          * MultiXactId that may never have been seen on disk before, and we need
5044          * to make sure that there are XLOG entries covering those ID numbers.
5045          * Else the same IDs might be re-used after a crash, which would be
5046          * disastrous if this page made it to disk before the crash.  Essentially
5047          * we have to enforce the WAL log-before-data rule even in this case.
5048          * (Also, in a PITR log-shipping or 2PC environment, we have to have XLOG
5049          * entries for everything anyway.)
5050          */
5051         if (RelationNeedsWAL(relation))
5052         {
5053                 xl_heap_lock xlrec;
5054                 XLogRecPtr      recptr;
5055
5056                 XLogBeginInsert();
5057                 XLogRegisterBuffer(0, *buffer, REGBUF_STANDARD);
5058
5059                 xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
5060                 xlrec.locking_xid = xid;
5061                 xlrec.infobits_set = compute_infobits(new_infomask,
5062                                                                                           tuple->t_data->t_infomask2);
5063                 XLogRegisterData((char *) &xlrec, SizeOfHeapLock);
5064
5065                 /* we don't decode row locks atm, so no need to log the origin */
5066
5067                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
5068
5069                 PageSetLSN(page, recptr);
5070         }
5071
5072         END_CRIT_SECTION();
5073
5074         LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
5075
5076         /*
5077          * Don't update the visibility map here. Locking a tuple doesn't change
5078          * visibility info.
5079          */
5080
5081         /*
5082          * Now that we have successfully marked the tuple as locked, we can
5083          * release the lmgr tuple lock, if we had it.
5084          */
5085         if (have_tuple_lock)
5086                 UnlockTupleTuplock(relation, tid, mode);
5087
5088         return HeapTupleMayBeUpdated;
5089 }
5090
5091 /*
5092  * Acquire heavyweight lock on the given tuple, in preparation for acquiring
5093  * its normal, Xmax-based tuple lock.
5094  *
5095  * have_tuple_lock is an input and output parameter: on input, it indicates
5096  * whether the lock has previously been acquired (and this function does
5097  * nothing in that case).  If this function returns success, have_tuple_lock
5098  * has been flipped to true.
5099  *
5100  * Returns false if it was unable to obtain the lock; this can only happen if
5101  * wait_policy is Skip.
5102  */
5103 static bool
5104 heap_acquire_tuplock(Relation relation, ItemPointer tid, LockTupleMode mode,
5105                                          LockWaitPolicy wait_policy, bool *have_tuple_lock)
5106 {
5107         if (*have_tuple_lock)
5108                 return true;
5109
5110         switch (wait_policy)
5111         {
5112                 case LockWaitBlock:
5113                         LockTupleTuplock(relation, tid, mode);
5114                         break;
5115
5116                 case LockWaitSkip:
5117                         if (!ConditionalLockTupleTuplock(relation, tid, mode))
5118                                 return false;
5119                         break;
5120
5121                 case LockWaitError:
5122                         if (!ConditionalLockTupleTuplock(relation, tid, mode))
5123                                 ereport(ERROR,
5124                                                 (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
5125                                         errmsg("could not obtain lock on row in relation \"%s\"",
5126                                                    RelationGetRelationName(relation))));
5127                         break;
5128         }
5129         *have_tuple_lock = true;
5130
5131         return true;
5132 }
5133
5134 /*
5135  * Given an original set of Xmax and infomask, and a transaction (identified by
5136  * add_to_xmax) acquiring a new lock of some mode, compute the new Xmax and
5137  * corresponding infomasks to use on the tuple.
5138  *
5139  * Note that this might have side effects such as creating a new MultiXactId.
5140  *
5141  * Most callers will have called HeapTupleSatisfiesUpdate before this function;
5142  * that will have set the HEAP_XMAX_INVALID bit if the xmax was a MultiXactId
5143  * but it was not running anymore. There is a race condition, which is that the
5144  * MultiXactId may have finished since then, but that uncommon case is handled
5145  * either here, or within MultiXactIdExpand.
5146  *
5147  * There is a similar race condition possible when the old xmax was a regular
5148  * TransactionId.  We test TransactionIdIsInProgress again just to narrow the
5149  * window, but it's still possible to end up creating an unnecessary
5150  * MultiXactId.  Fortunately this is harmless.
5151  */
5152 static void
5153 compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
5154                                                   uint16 old_infomask2, TransactionId add_to_xmax,
5155                                                   LockTupleMode mode, bool is_update,
5156                                                   TransactionId *result_xmax, uint16 *result_infomask,
5157                                                   uint16 *result_infomask2)
5158 {
5159         TransactionId new_xmax;
5160         uint16          new_infomask,
5161                                 new_infomask2;
5162
5163         Assert(TransactionIdIsCurrentTransactionId(add_to_xmax));
5164
5165 l5:
5166         new_infomask = 0;
5167         new_infomask2 = 0;
5168         if (old_infomask & HEAP_XMAX_INVALID)
5169         {
5170                 /*
5171                  * No previous locker; we just insert our own TransactionId.
5172                  *
5173                  * Note that it's critical that this case be the first one checked,
5174                  * because there are several blocks below that come back to this one
5175                  * to implement certain optimizations; old_infomask might contain
5176                  * other dirty bits in those cases, but we don't really care.
5177                  */
5178                 if (is_update)
5179                 {
5180                         new_xmax = add_to_xmax;
5181                         if (mode == LockTupleExclusive)
5182                                 new_infomask2 |= HEAP_KEYS_UPDATED;
5183                 }
5184                 else
5185                 {
5186                         new_infomask |= HEAP_XMAX_LOCK_ONLY;
5187                         switch (mode)
5188                         {
5189                                 case LockTupleKeyShare:
5190                                         new_xmax = add_to_xmax;
5191                                         new_infomask |= HEAP_XMAX_KEYSHR_LOCK;
5192                                         break;
5193                                 case LockTupleShare:
5194                                         new_xmax = add_to_xmax;
5195                                         new_infomask |= HEAP_XMAX_SHR_LOCK;
5196                                         break;
5197                                 case LockTupleNoKeyExclusive:
5198                                         new_xmax = add_to_xmax;
5199                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
5200                                         break;
5201                                 case LockTupleExclusive:
5202                                         new_xmax = add_to_xmax;
5203                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
5204                                         new_infomask2 |= HEAP_KEYS_UPDATED;
5205                                         break;
5206                                 default:
5207                                         new_xmax = InvalidTransactionId;        /* silence compiler */
5208                                         elog(ERROR, "invalid lock mode");
5209                         }
5210                 }
5211         }
5212         else if (old_infomask & HEAP_XMAX_IS_MULTI)
5213         {
5214                 MultiXactStatus new_status;
5215
5216                 /*
5217                  * Currently we don't allow XMAX_COMMITTED to be set for multis, so
5218                  * cross-check.
5219                  */
5220                 Assert(!(old_infomask & HEAP_XMAX_COMMITTED));
5221
5222                 /*
5223                  * A multixact together with LOCK_ONLY set but neither lock bit set
5224                  * (i.e. a pg_upgraded share locked tuple) cannot possibly be running
5225                  * anymore.  This check is critical for databases upgraded by
5226                  * pg_upgrade; both MultiXactIdIsRunning and MultiXactIdExpand assume
5227                  * that such multis are never passed.
5228                  */
5229                 if (!(old_infomask & HEAP_LOCK_MASK) &&
5230                         HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5231                 {
5232                         old_infomask &= ~HEAP_XMAX_IS_MULTI;
5233                         old_infomask |= HEAP_XMAX_INVALID;
5234                         goto l5;
5235                 }
5236
5237                 /*
5238                  * If the XMAX is already a MultiXactId, then we need to expand it to
5239                  * include add_to_xmax; but if all the members were lockers and are
5240                  * all gone, we can do away with the IS_MULTI bit and just set
5241                  * add_to_xmax as the only locker/updater.  If all lockers are gone
5242                  * and we have an updater that aborted, we can also do without a
5243                  * multi.
5244                  *
5245                  * The cost of doing GetMultiXactIdMembers would be paid by
5246                  * MultiXactIdExpand if we weren't to do this, so this check is not
5247                  * incurring extra work anyhow.
5248                  */
5249                 if (!MultiXactIdIsRunning(xmax, HEAP_XMAX_IS_LOCKED_ONLY(old_infomask)))
5250                 {
5251                         if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) ||
5252                                 !TransactionIdDidCommit(MultiXactIdGetUpdateXid(xmax,
5253                                                                                                                           old_infomask)))
5254                         {
5255                                 /*
5256                                  * Reset these bits and restart; otherwise fall through to
5257                                  * create a new multi below.
5258                                  */
5259                                 old_infomask &= ~HEAP_XMAX_IS_MULTI;
5260                                 old_infomask |= HEAP_XMAX_INVALID;
5261                                 goto l5;
5262                         }
5263                 }
5264
5265                 new_status = get_mxact_status_for_lock(mode, is_update);
5266
5267                 new_xmax = MultiXactIdExpand((MultiXactId) xmax, add_to_xmax,
5268                                                                          new_status);
5269                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5270         }
5271         else if (old_infomask & HEAP_XMAX_COMMITTED)
5272         {
5273                 /*
5274                  * It's a committed update, so we need to preserve him as updater of
5275                  * the tuple.
5276                  */
5277                 MultiXactStatus status;
5278                 MultiXactStatus new_status;
5279
5280                 if (old_infomask2 & HEAP_KEYS_UPDATED)
5281                         status = MultiXactStatusUpdate;
5282                 else
5283                         status = MultiXactStatusNoKeyUpdate;
5284
5285                 new_status = get_mxact_status_for_lock(mode, is_update);
5286
5287                 /*
5288                  * since it's not running, it's obviously impossible for the old
5289                  * updater to be identical to the current one, so we need not check
5290                  * for that case as we do in the block above.
5291                  */
5292                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5293                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5294         }
5295         else if (TransactionIdIsInProgress(xmax))
5296         {
5297                 /*
5298                  * If the XMAX is a valid, in-progress TransactionId, then we need to
5299                  * create a new MultiXactId that includes both the old locker or
5300                  * updater and our own TransactionId.
5301                  */
5302                 MultiXactStatus new_status;
5303                 MultiXactStatus old_status;
5304                 LockTupleMode old_mode;
5305
5306                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5307                 {
5308                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5309                                 old_status = MultiXactStatusForKeyShare;
5310                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5311                                 old_status = MultiXactStatusForShare;
5312                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5313                         {
5314                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
5315                                         old_status = MultiXactStatusForUpdate;
5316                                 else
5317                                         old_status = MultiXactStatusForNoKeyUpdate;
5318                         }
5319                         else
5320                         {
5321                                 /*
5322                                  * LOCK_ONLY can be present alone only when a page has been
5323                                  * upgraded by pg_upgrade.  But in that case,
5324                                  * TransactionIdIsInProgress() should have returned false.  We
5325                                  * assume it's no longer locked in this case.
5326                                  */
5327                                 elog(WARNING, "LOCK_ONLY found for Xid in progress %u", xmax);
5328                                 old_infomask |= HEAP_XMAX_INVALID;
5329                                 old_infomask &= ~HEAP_XMAX_LOCK_ONLY;
5330                                 goto l5;
5331                         }
5332                 }
5333                 else
5334                 {
5335                         /* it's an update, but which kind? */
5336                         if (old_infomask2 & HEAP_KEYS_UPDATED)
5337                                 old_status = MultiXactStatusUpdate;
5338                         else
5339                                 old_status = MultiXactStatusNoKeyUpdate;
5340                 }
5341
5342                 old_mode = TUPLOCK_from_mxstatus(old_status);
5343
5344                 /*
5345                  * If the lock to be acquired is for the same TransactionId as the
5346                  * existing lock, there's an optimization possible: consider only the
5347                  * strongest of both locks as the only one present, and restart.
5348                  */
5349                 if (xmax == add_to_xmax)
5350                 {
5351                         /*
5352                          * Note that it's not possible for the original tuple to be
5353                          * updated: we wouldn't be here because the tuple would have been
5354                          * invisible and we wouldn't try to update it.  As a subtlety,
5355                          * this code can also run when traversing an update chain to lock
5356                          * future versions of a tuple.  But we wouldn't be here either,
5357                          * because the add_to_xmax would be different from the original
5358                          * updater.
5359                          */
5360                         Assert(HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5361
5362                         /* acquire the strongest of both */
5363                         if (mode < old_mode)
5364                                 mode = old_mode;
5365                         /* mustn't touch is_update */
5366
5367                         old_infomask |= HEAP_XMAX_INVALID;
5368                         goto l5;
5369                 }
5370
5371                 /* otherwise, just fall back to creating a new multixact */
5372                 new_status = get_mxact_status_for_lock(mode, is_update);
5373                 new_xmax = MultiXactIdCreate(xmax, old_status,
5374                                                                          add_to_xmax, new_status);
5375                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5376         }
5377         else if (!HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) &&
5378                          TransactionIdDidCommit(xmax))
5379         {
5380                 /*
5381                  * It's a committed update, so we gotta preserve him as updater of the
5382                  * tuple.
5383                  */
5384                 MultiXactStatus status;
5385                 MultiXactStatus new_status;
5386
5387                 if (old_infomask2 & HEAP_KEYS_UPDATED)
5388                         status = MultiXactStatusUpdate;
5389                 else
5390                         status = MultiXactStatusNoKeyUpdate;
5391
5392                 new_status = get_mxact_status_for_lock(mode, is_update);
5393
5394                 /*
5395                  * since it's not running, it's obviously impossible for the old
5396                  * updater to be identical to the current one, so we need not check
5397                  * for that case as we do in the block above.
5398                  */
5399                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
5400                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
5401         }
5402         else
5403         {
5404                 /*
5405                  * Can get here iff the locking/updating transaction was running when
5406                  * the infomask was extracted from the tuple, but finished before
5407                  * TransactionIdIsInProgress got to run.  Deal with it as if there was
5408                  * no locker at all in the first place.
5409                  */
5410                 old_infomask |= HEAP_XMAX_INVALID;
5411                 goto l5;
5412         }
5413
5414         *result_infomask = new_infomask;
5415         *result_infomask2 = new_infomask2;
5416         *result_xmax = new_xmax;
5417 }
5418
5419 /*
5420  * Subroutine for heap_lock_updated_tuple_rec.
5421  *
5422  * Given a hypothetical multixact status held by the transaction identified
5423  * with the given xid, does the current transaction need to wait, fail, or can
5424  * it continue if it wanted to acquire a lock of the given mode?  "needwait"
5425  * is set to true if waiting is necessary; if it can continue, then
5426  * HeapTupleMayBeUpdated is returned.  In case of a conflict, a different
5427  * HeapTupleSatisfiesUpdate return code is returned.
5428  *
5429  * The held status is said to be hypothetical because it might correspond to a
5430  * lock held by a single Xid, i.e. not a real MultiXactId; we express it this
5431  * way for simplicity of API.
5432  */
5433 static HTSU_Result
5434 test_lockmode_for_conflict(MultiXactStatus status, TransactionId xid,
5435                                                    LockTupleMode mode, bool *needwait)
5436 {
5437         MultiXactStatus wantedstatus;
5438
5439         *needwait = false;
5440         wantedstatus = get_mxact_status_for_lock(mode, false);
5441
5442         /*
5443          * Note: we *must* check TransactionIdIsInProgress before
5444          * TransactionIdDidAbort/Commit; see comment at top of tqual.c for an
5445          * explanation.
5446          */
5447         if (TransactionIdIsCurrentTransactionId(xid))
5448         {
5449                 /*
5450                  * Updated by our own transaction? Just return failure.  This
5451                  * shouldn't normally happen.
5452                  */
5453                 return HeapTupleSelfUpdated;
5454         }
5455         else if (TransactionIdIsInProgress(xid))
5456         {
5457                 /*
5458                  * If the locking transaction is running, what we do depends on
5459                  * whether the lock modes conflict: if they do, then we must wait for
5460                  * it to finish; otherwise we can fall through to lock this tuple
5461                  * version without waiting.
5462                  */
5463                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5464                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5465                 {
5466                         *needwait = true;
5467                 }
5468
5469                 /*
5470                  * If we set needwait above, then this value doesn't matter;
5471                  * otherwise, this value signals to caller that it's okay to proceed.
5472                  */
5473                 return HeapTupleMayBeUpdated;
5474         }
5475         else if (TransactionIdDidAbort(xid))
5476                 return HeapTupleMayBeUpdated;
5477         else if (TransactionIdDidCommit(xid))
5478         {
5479                 /*
5480                  * The other transaction committed.  If it was only a locker, then the
5481                  * lock is completely gone now and we can return success; but if it
5482                  * was an update, then what we do depends on whether the two lock
5483                  * modes conflict.  If they conflict, then we must report error to
5484                  * caller. But if they don't, we can fall through to allow the current
5485                  * transaction to lock the tuple.
5486                  *
5487                  * Note: the reason we worry about ISUPDATE here is because as soon as
5488                  * a transaction ends, all its locks are gone and meaningless, and
5489                  * thus we can ignore them; whereas its updates persist.  In the
5490                  * TransactionIdIsInProgress case, above, we don't need to check
5491                  * because we know the lock is still "alive" and thus a conflict needs
5492                  * always be checked.
5493                  */
5494                 if (!ISUPDATE_from_mxstatus(status))
5495                         return HeapTupleMayBeUpdated;
5496
5497                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5498                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5499                         /* bummer */
5500                         return HeapTupleUpdated;
5501
5502                 return HeapTupleMayBeUpdated;
5503         }
5504
5505         /* Not in progress, not aborted, not committed -- must have crashed */
5506         return HeapTupleMayBeUpdated;
5507 }
5508
5509
5510 /*
5511  * Recursive part of heap_lock_updated_tuple
5512  *
5513  * Fetch the tuple pointed to by tid in rel, and mark it as locked by the given
5514  * xid with the given mode; if this tuple is updated, recurse to lock the new
5515  * version as well.
5516  */
5517 static HTSU_Result
5518 heap_lock_updated_tuple_rec(Relation rel, ItemPointer tid, TransactionId xid,
5519                                                         LockTupleMode mode)
5520 {
5521         ItemPointerData tupid;
5522         HeapTupleData mytup;
5523         Buffer          buf;
5524         uint16          new_infomask,
5525                                 new_infomask2,
5526                                 old_infomask,
5527                                 old_infomask2;
5528         TransactionId xmax,
5529                                 new_xmax;
5530         TransactionId priorXmax = InvalidTransactionId;
5531
5532         ItemPointerCopy(tid, &tupid);
5533
5534         for (;;)
5535         {
5536                 new_infomask = 0;
5537                 new_xmax = InvalidTransactionId;
5538                 ItemPointerCopy(&tupid, &(mytup.t_self));
5539
5540                 if (!heap_fetch(rel, SnapshotAny, &mytup, &buf, false, NULL))
5541                 {
5542                         /*
5543                          * if we fail to find the updated version of the tuple, it's
5544                          * because it was vacuumed/pruned away after its creator
5545                          * transaction aborted.  So behave as if we got to the end of the
5546                          * chain, and there's no further tuple to lock: return success to
5547                          * caller.
5548                          */
5549                         return HeapTupleMayBeUpdated;
5550                 }
5551
5552 l4:
5553                 CHECK_FOR_INTERRUPTS();
5554                 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5555
5556                 /*
5557                  * Check the tuple XMIN against prior XMAX, if any.  If we reached the
5558                  * end of the chain, we're done, so return success.
5559                  */
5560                 if (TransactionIdIsValid(priorXmax) &&
5561                         !TransactionIdEquals(HeapTupleHeaderGetXmin(mytup.t_data),
5562                                                                  priorXmax))
5563                 {
5564                         UnlockReleaseBuffer(buf);
5565                         return HeapTupleMayBeUpdated;
5566                 }
5567
5568                 old_infomask = mytup.t_data->t_infomask;
5569                 old_infomask2 = mytup.t_data->t_infomask2;
5570                 xmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5571
5572                 /*
5573                  * If this tuple version has been updated or locked by some concurrent
5574                  * transaction(s), what we do depends on whether our lock mode
5575                  * conflicts with what those other transactions hold, and also on the
5576                  * status of them.
5577                  */
5578                 if (!(old_infomask & HEAP_XMAX_INVALID))
5579                 {
5580                         TransactionId rawxmax;
5581                         bool            needwait;
5582
5583                         rawxmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5584                         if (old_infomask & HEAP_XMAX_IS_MULTI)
5585                         {
5586                                 int                     nmembers;
5587                                 int                     i;
5588                                 MultiXactMember *members;
5589
5590                                 nmembers = GetMultiXactIdMembers(rawxmax, &members, false,
5591                                                                          HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5592                                 for (i = 0; i < nmembers; i++)
5593                                 {
5594                                         HTSU_Result res;
5595
5596                                         res = test_lockmode_for_conflict(members[i].status,
5597                                                                                                          members[i].xid,
5598                                                                                                          mode, &needwait);
5599
5600                                         if (needwait)
5601                                         {
5602                                                 LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5603                                                 XactLockTableWait(members[i].xid, rel,
5604                                                                                   &mytup.t_self,
5605                                                                                   XLTW_LockUpdated);
5606                                                 pfree(members);
5607                                                 goto l4;
5608                                         }
5609                                         if (res != HeapTupleMayBeUpdated)
5610                                         {
5611                                                 UnlockReleaseBuffer(buf);
5612                                                 pfree(members);
5613                                                 return res;
5614                                         }
5615                                 }
5616                                 if (members)
5617                                         pfree(members);
5618                         }
5619                         else
5620                         {
5621                                 HTSU_Result res;
5622                                 MultiXactStatus status;
5623
5624                                 /*
5625                                  * For a non-multi Xmax, we first need to compute the
5626                                  * corresponding MultiXactStatus by using the infomask bits.
5627                                  */
5628                                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5629                                 {
5630                                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5631                                                 status = MultiXactStatusForKeyShare;
5632                                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5633                                                 status = MultiXactStatusForShare;
5634                                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5635                                         {
5636                                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
5637                                                         status = MultiXactStatusForUpdate;
5638                                                 else
5639                                                         status = MultiXactStatusForNoKeyUpdate;
5640                                         }
5641                                         else
5642                                         {
5643                                                 /*
5644                                                  * LOCK_ONLY present alone (a pg_upgraded tuple marked
5645                                                  * as share-locked in the old cluster) shouldn't be
5646                                                  * seen in the middle of an update chain.
5647                                                  */
5648                                                 elog(ERROR, "invalid lock status in tuple");
5649                                         }
5650                                 }
5651                                 else
5652                                 {
5653                                         /* it's an update, but which kind? */
5654                                         if (old_infomask2 & HEAP_KEYS_UPDATED)
5655                                                 status = MultiXactStatusUpdate;
5656                                         else
5657                                                 status = MultiXactStatusNoKeyUpdate;
5658                                 }
5659
5660                                 res = test_lockmode_for_conflict(status, rawxmax, mode,
5661                                                                                                  &needwait);
5662                                 if (needwait)
5663                                 {
5664                                         LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5665                                         XactLockTableWait(rawxmax, rel, &mytup.t_self,
5666                                                                           XLTW_LockUpdated);
5667                                         goto l4;
5668                                 }
5669                                 if (res != HeapTupleMayBeUpdated)
5670                                 {
5671                                         UnlockReleaseBuffer(buf);
5672                                         return res;
5673                                 }
5674                         }
5675                 }
5676
5677                 /* compute the new Xmax and infomask values for the tuple ... */
5678                 compute_new_xmax_infomask(xmax, old_infomask, mytup.t_data->t_infomask2,
5679                                                                   xid, mode, false,
5680                                                                   &new_xmax, &new_infomask, &new_infomask2);
5681
5682                 START_CRIT_SECTION();
5683
5684                 /* ... and set them */
5685                 HeapTupleHeaderSetXmax(mytup.t_data, new_xmax);
5686                 mytup.t_data->t_infomask &= ~HEAP_XMAX_BITS;
5687                 mytup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5688                 mytup.t_data->t_infomask |= new_infomask;
5689                 mytup.t_data->t_infomask2 |= new_infomask2;
5690
5691                 MarkBufferDirty(buf);
5692
5693                 /* XLOG stuff */
5694                 if (RelationNeedsWAL(rel))
5695                 {
5696                         xl_heap_lock_updated xlrec;
5697                         XLogRecPtr      recptr;
5698                         Page            page = BufferGetPage(buf);
5699
5700                         XLogBeginInsert();
5701                         XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
5702
5703                         xlrec.offnum = ItemPointerGetOffsetNumber(&mytup.t_self);
5704                         xlrec.xmax = new_xmax;
5705                         xlrec.infobits_set = compute_infobits(new_infomask, new_infomask2);
5706
5707                         XLogRegisterData((char *) &xlrec, SizeOfHeapLockUpdated);
5708
5709                         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_LOCK_UPDATED);
5710
5711                         PageSetLSN(page, recptr);
5712                 }
5713
5714                 END_CRIT_SECTION();
5715
5716                 /* if we find the end of update chain, we're done. */
5717                 if (mytup.t_data->t_infomask & HEAP_XMAX_INVALID ||
5718                         ItemPointerEquals(&mytup.t_self, &mytup.t_data->t_ctid) ||
5719                         HeapTupleHeaderIsOnlyLocked(mytup.t_data))
5720                 {
5721                         UnlockReleaseBuffer(buf);
5722                         return HeapTupleMayBeUpdated;
5723                 }
5724
5725                 /* tail recursion */
5726                 priorXmax = HeapTupleHeaderGetUpdateXid(mytup.t_data);
5727                 ItemPointerCopy(&(mytup.t_data->t_ctid), &tupid);
5728                 UnlockReleaseBuffer(buf);
5729         }
5730 }
5731
5732 /*
5733  * heap_lock_updated_tuple
5734  *              Follow update chain when locking an updated tuple, acquiring locks (row
5735  *              marks) on the updated versions.
5736  *
5737  * The initial tuple is assumed to be already locked.
5738  *
5739  * This function doesn't check visibility, it just unconditionally marks the
5740  * tuple(s) as locked.  If any tuple in the updated chain is being deleted
5741  * concurrently (or updated with the key being modified), sleep until the
5742  * transaction doing it is finished.
5743  *
5744  * Note that we don't acquire heavyweight tuple locks on the tuples we walk
5745  * when we have to wait for other transactions to release them, as opposed to
5746  * what heap_lock_tuple does.  The reason is that having more than one
5747  * transaction walking the chain is probably uncommon enough that risk of
5748  * starvation is not likely: one of the preconditions for being here is that
5749  * the snapshot in use predates the update that created this tuple (because we
5750  * started at an earlier version of the tuple), but at the same time such a
5751  * transaction cannot be using repeatable read or serializable isolation
5752  * levels, because that would lead to a serializability failure.
5753  */
5754 static HTSU_Result
5755 heap_lock_updated_tuple(Relation rel, HeapTuple tuple, ItemPointer ctid,
5756                                                 TransactionId xid, LockTupleMode mode)
5757 {
5758         if (!ItemPointerEquals(&tuple->t_self, ctid))
5759         {
5760                 /*
5761                  * If this is the first possibly-multixact-able operation in the
5762                  * current transaction, set my per-backend OldestMemberMXactId
5763                  * setting. We can be certain that the transaction will never become a
5764                  * member of any older MultiXactIds than that.  (We have to do this
5765                  * even if we end up just using our own TransactionId below, since
5766                  * some other backend could incorporate our XID into a MultiXact
5767                  * immediately afterwards.)
5768                  */
5769                 MultiXactIdSetOldestMember();
5770
5771                 return heap_lock_updated_tuple_rec(rel, ctid, xid, mode);
5772         }
5773
5774         /* nothing to lock */
5775         return HeapTupleMayBeUpdated;
5776 }
5777
5778 /*
5779  *      heap_finish_speculative - mark speculative insertion as successful
5780  *
5781  * To successfully finish a speculative insertion we have to clear speculative
5782  * token from tuple.  To do so the t_ctid field, which will contain a
5783  * speculative token value, is modified in place to point to the tuple itself,
5784  * which is characteristic of a newly inserted ordinary tuple.
5785  *
5786  * NB: It is not ok to commit without either finishing or aborting a
5787  * speculative insertion.  We could treat speculative tuples of committed
5788  * transactions implicitly as completed, but then we would have to be prepared
5789  * to deal with speculative tokens on committed tuples.  That wouldn't be
5790  * difficult - no-one looks at the ctid field of a tuple with invalid xmax -
5791  * but clearing the token at completion isn't very expensive either.
5792  * An explicit confirmation WAL record also makes logical decoding simpler.
5793  */
5794 void
5795 heap_finish_speculative(Relation relation, HeapTuple tuple)
5796 {
5797         Buffer          buffer;
5798         Page            page;
5799         OffsetNumber offnum;
5800         ItemId          lp = NULL;
5801         HeapTupleHeader htup;
5802
5803         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&(tuple->t_self)));
5804         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5805         page = (Page) BufferGetPage(buffer);
5806
5807         offnum = ItemPointerGetOffsetNumber(&(tuple->t_self));
5808         if (PageGetMaxOffsetNumber(page) >= offnum)
5809                 lp = PageGetItemId(page, offnum);
5810
5811         if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
5812                 elog(ERROR, "invalid lp");
5813
5814         htup = (HeapTupleHeader) PageGetItem(page, lp);
5815
5816         /* SpecTokenOffsetNumber should be distinguishable from any real offset */
5817         StaticAssertStmt(MaxOffsetNumber < SpecTokenOffsetNumber,
5818                                          "invalid speculative token constant");
5819
5820         /* NO EREPORT(ERROR) from here till changes are logged */
5821         START_CRIT_SECTION();
5822
5823         Assert(HeapTupleHeaderIsSpeculative(tuple->t_data));
5824
5825         MarkBufferDirty(buffer);
5826
5827         /*
5828          * Replace the speculative insertion token with a real t_ctid, pointing to
5829          * itself like it does on regular tuples.
5830          */
5831         htup->t_ctid = tuple->t_self;
5832
5833         /* XLOG stuff */
5834         if (RelationNeedsWAL(relation))
5835         {
5836                 xl_heap_confirm xlrec;
5837                 XLogRecPtr      recptr;
5838
5839                 xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
5840
5841                 XLogBeginInsert();
5842
5843                 /* We want the same filtering on this as on a plain insert */
5844                 XLogIncludeOrigin();
5845
5846                 XLogRegisterData((char *) &xlrec, SizeOfHeapConfirm);
5847                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
5848
5849                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_CONFIRM);
5850
5851                 PageSetLSN(page, recptr);
5852         }
5853
5854         END_CRIT_SECTION();
5855
5856         UnlockReleaseBuffer(buffer);
5857 }
5858
5859 /*
5860  *      heap_abort_speculative - kill a speculatively inserted tuple
5861  *
5862  * Marks a tuple that was speculatively inserted in the same command as dead,
5863  * by setting its xmin as invalid.  That makes it immediately appear as dead
5864  * to all transactions, including our own.  In particular, it makes
5865  * HeapTupleSatisfiesDirty() regard the tuple as dead, so that another backend
5866  * inserting a duplicate key value won't unnecessarily wait for our whole
5867  * transaction to finish (it'll just wait for our speculative insertion to
5868  * finish).
5869  *
5870  * Killing the tuple prevents "unprincipled deadlocks", which are deadlocks
5871  * that arise due to a mutual dependency that is not user visible.  By
5872  * definition, unprincipled deadlocks cannot be prevented by the user
5873  * reordering lock acquisition in client code, because the implementation level
5874  * lock acquisitions are not under the user's direct control.  If speculative
5875  * inserters did not take this precaution, then under high concurrency they
5876  * could deadlock with each other, which would not be acceptable.
5877  *
5878  * This is somewhat redundant with heap_delete, but we prefer to have a
5879  * dedicated routine with stripped down requirements.
5880  *
5881  * This routine does not affect logical decoding as it only looks at
5882  * confirmation records.
5883  */
5884 void
5885 heap_abort_speculative(Relation relation, HeapTuple tuple)
5886 {
5887         TransactionId xid = GetCurrentTransactionId();
5888         ItemPointer tid = &(tuple->t_self);
5889         ItemId          lp;
5890         HeapTupleData tp;
5891         Page            page;
5892         BlockNumber block;
5893         Buffer          buffer;
5894
5895         Assert(ItemPointerIsValid(tid));
5896
5897         block = ItemPointerGetBlockNumber(tid);
5898         buffer = ReadBuffer(relation, block);
5899         page = BufferGetPage(buffer);
5900
5901         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5902
5903         /*
5904          * Page can't be all visible, we just inserted into it, and are still
5905          * running.
5906          */
5907         Assert(!PageIsAllVisible(page));
5908
5909         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
5910         Assert(ItemIdIsNormal(lp));
5911
5912         tp.t_tableOid = RelationGetRelid(relation);
5913         tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
5914         tp.t_len = ItemIdGetLength(lp);
5915         tp.t_self = *tid;
5916
5917         /*
5918          * Sanity check that the tuple really is a speculatively inserted tuple,
5919          * inserted by us.
5920          */
5921         if (tp.t_data->t_choice.t_heap.t_xmin != xid)
5922                 elog(ERROR, "attempted to kill a tuple inserted by another transaction");
5923         if (!HeapTupleHeaderIsSpeculative(tp.t_data))
5924                 elog(ERROR, "attempted to kill a non-speculative tuple");
5925         Assert(!HeapTupleHeaderIsHeapOnly(tp.t_data));
5926
5927         /*
5928          * No need to check for serializable conflicts here.  There is never a
5929          * need for a combocid, either.  No need to extract replica identity, or
5930          * do anything special with infomask bits.
5931          */
5932
5933         START_CRIT_SECTION();
5934
5935         /*
5936          * The tuple will become DEAD immediately.  Flag that this page
5937          * immediately is a candidate for pruning by setting xmin to
5938          * RecentGlobalXmin.  That's not pretty, but it doesn't seem worth
5939          * inventing a nicer API for this.
5940          */
5941         Assert(TransactionIdIsValid(RecentGlobalXmin));
5942         PageSetPrunable(page, RecentGlobalXmin);
5943
5944         /* store transaction information of xact deleting the tuple */
5945         tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
5946         tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5947
5948         /*
5949          * Set the tuple header xmin to InvalidTransactionId.  This makes the
5950          * tuple immediately invisible everyone.  (In particular, to any
5951          * transactions waiting on the speculative token, woken up later.)
5952          */
5953         HeapTupleHeaderSetXmin(tp.t_data, InvalidTransactionId);
5954
5955         /* Clear the speculative insertion token too */
5956         tp.t_data->t_ctid = tp.t_self;
5957
5958         MarkBufferDirty(buffer);
5959
5960         /*
5961          * XLOG stuff
5962          *
5963          * The WAL records generated here match heap_delete().  The same recovery
5964          * routines are used.
5965          */
5966         if (RelationNeedsWAL(relation))
5967         {
5968                 xl_heap_delete xlrec;
5969                 XLogRecPtr      recptr;
5970
5971                 xlrec.flags = XLH_DELETE_IS_SUPER;
5972                 xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
5973                                                                                           tp.t_data->t_infomask2);
5974                 xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
5975                 xlrec.xmax = xid;
5976
5977                 XLogBeginInsert();
5978                 XLogRegisterData((char *) &xlrec, SizeOfHeapDelete);
5979                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
5980
5981                 /* No replica identity & replication origin logged */
5982
5983                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
5984
5985                 PageSetLSN(page, recptr);
5986         }
5987
5988         END_CRIT_SECTION();
5989
5990         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5991
5992         if (HeapTupleHasExternal(&tp))
5993                 toast_delete(relation, &tp);
5994
5995         /*
5996          * Never need to mark tuple for invalidation, since catalogs don't support
5997          * speculative insertion
5998          */
5999
6000         /* Now we can release the buffer */
6001         ReleaseBuffer(buffer);
6002
6003         /* count deletion, as we counted the insertion too */
6004         pgstat_count_heap_delete(relation);
6005 }
6006
6007 /*
6008  * heap_inplace_update - update a tuple "in place" (ie, overwrite it)
6009  *
6010  * Overwriting violates both MVCC and transactional safety, so the uses
6011  * of this function in Postgres are extremely limited.  Nonetheless we
6012  * find some places to use it.
6013  *
6014  * The tuple cannot change size, and therefore it's reasonable to assume
6015  * that its null bitmap (if any) doesn't change either.  So we just
6016  * overwrite the data portion of the tuple without touching the null
6017  * bitmap or any of the header fields.
6018  *
6019  * tuple is an in-memory tuple structure containing the data to be written
6020  * over the target tuple.  Also, tuple->t_self identifies the target tuple.
6021  */
6022 void
6023 heap_inplace_update(Relation relation, HeapTuple tuple)
6024 {
6025         Buffer          buffer;
6026         Page            page;
6027         OffsetNumber offnum;
6028         ItemId          lp = NULL;
6029         HeapTupleHeader htup;
6030         uint32          oldlen;
6031         uint32          newlen;
6032
6033         /*
6034          * For now, parallel operations are required to be strictly read-only.
6035          * Unlike a regular update, this should never create a combo CID, so it
6036          * might be possible to relax this restriction, but not without more
6037          * thought and testing.  It's not clear that it would be useful, anyway.
6038          */
6039         if (IsInParallelMode())
6040                 ereport(ERROR,
6041                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
6042                                  errmsg("cannot update tuples during a parallel operation")));
6043
6044         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&(tuple->t_self)));
6045         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
6046         page = (Page) BufferGetPage(buffer);
6047
6048         offnum = ItemPointerGetOffsetNumber(&(tuple->t_self));
6049         if (PageGetMaxOffsetNumber(page) >= offnum)
6050                 lp = PageGetItemId(page, offnum);
6051
6052         if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
6053                 elog(ERROR, "invalid lp");
6054
6055         htup = (HeapTupleHeader) PageGetItem(page, lp);
6056
6057         oldlen = ItemIdGetLength(lp) - htup->t_hoff;
6058         newlen = tuple->t_len - tuple->t_data->t_hoff;
6059         if (oldlen != newlen || htup->t_hoff != tuple->t_data->t_hoff)
6060                 elog(ERROR, "wrong tuple length");
6061
6062         /* NO EREPORT(ERROR) from here till changes are logged */
6063         START_CRIT_SECTION();
6064
6065         memcpy((char *) htup + htup->t_hoff,
6066                    (char *) tuple->t_data + tuple->t_data->t_hoff,
6067                    newlen);
6068
6069         MarkBufferDirty(buffer);
6070
6071         /* XLOG stuff */
6072         if (RelationNeedsWAL(relation))
6073         {
6074                 xl_heap_inplace xlrec;
6075                 XLogRecPtr      recptr;
6076
6077                 xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
6078
6079                 XLogBeginInsert();
6080                 XLogRegisterData((char *) &xlrec, SizeOfHeapInplace);
6081
6082                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
6083                 XLogRegisterBufData(0, (char *) htup + htup->t_hoff, newlen);
6084
6085                 /* inplace updates aren't decoded atm, don't log the origin */
6086
6087                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_INPLACE);
6088
6089                 PageSetLSN(page, recptr);
6090         }
6091
6092         END_CRIT_SECTION();
6093
6094         UnlockReleaseBuffer(buffer);
6095
6096         /*
6097          * Send out shared cache inval if necessary.  Note that because we only
6098          * pass the new version of the tuple, this mustn't be used for any
6099          * operations that could change catcache lookup keys.  But we aren't
6100          * bothering with index updates either, so that's true a fortiori.
6101          */
6102         if (!IsBootstrapProcessingMode())
6103                 CacheInvalidateHeapTuple(relation, tuple, NULL);
6104 }
6105
6106 #define         FRM_NOOP                                0x0001
6107 #define         FRM_INVALIDATE_XMAX             0x0002
6108 #define         FRM_RETURN_IS_XID               0x0004
6109 #define         FRM_RETURN_IS_MULTI             0x0008
6110 #define         FRM_MARK_COMMITTED              0x0010
6111
6112 /*
6113  * FreezeMultiXactId
6114  *              Determine what to do during freezing when a tuple is marked by a
6115  *              MultiXactId.
6116  *
6117  * NB -- this might have the side-effect of creating a new MultiXactId!
6118  *
6119  * "flags" is an output value; it's used to tell caller what to do on return.
6120  * Possible flags are:
6121  * FRM_NOOP
6122  *              don't do anything -- keep existing Xmax
6123  * FRM_INVALIDATE_XMAX
6124  *              mark Xmax as InvalidTransactionId and set XMAX_INVALID flag.
6125  * FRM_RETURN_IS_XID
6126  *              The Xid return value is a single update Xid to set as xmax.
6127  * FRM_MARK_COMMITTED
6128  *              Xmax can be marked as HEAP_XMAX_COMMITTED
6129  * FRM_RETURN_IS_MULTI
6130  *              The return value is a new MultiXactId to set as new Xmax.
6131  *              (caller must obtain proper infomask bits using GetMultiXactIdHintBits)
6132  */
6133 static TransactionId
6134 FreezeMultiXactId(MultiXactId multi, uint16 t_infomask,
6135                                   TransactionId cutoff_xid, MultiXactId cutoff_multi,
6136                                   uint16 *flags)
6137 {
6138         TransactionId xid = InvalidTransactionId;
6139         int                     i;
6140         MultiXactMember *members;
6141         int                     nmembers;
6142         bool            need_replace;
6143         int                     nnewmembers;
6144         MultiXactMember *newmembers;
6145         bool            has_lockers;
6146         TransactionId update_xid;
6147         bool            update_committed;
6148         bool            allow_old;
6149
6150         *flags = 0;
6151
6152         /* We should only be called in Multis */
6153         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6154
6155         if (!MultiXactIdIsValid(multi))
6156         {
6157                 /* Ensure infomask bits are appropriately set/reset */
6158                 *flags |= FRM_INVALIDATE_XMAX;
6159                 return InvalidTransactionId;
6160         }
6161         else if (MultiXactIdPrecedes(multi, cutoff_multi))
6162         {
6163                 /*
6164                  * This old multi cannot possibly have members still running.  If it
6165                  * was a locker only, it can be removed without any further
6166                  * consideration; but if it contained an update, we might need to
6167                  * preserve it.
6168                  *
6169                  * Don't assert MultiXactIdIsRunning if the multi came from a
6170                  * pg_upgrade'd share-locked tuple, though, as doing that causes an
6171                  * error to be raised unnecessarily.
6172                  */
6173                 Assert((!(t_infomask & HEAP_LOCK_MASK) &&
6174                                 HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)) ||
6175                            !MultiXactIdIsRunning(multi,
6176                                                                          HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)));
6177                 if (HEAP_XMAX_IS_LOCKED_ONLY(t_infomask))
6178                 {
6179                         *flags |= FRM_INVALIDATE_XMAX;
6180                         xid = InvalidTransactionId; /* not strictly necessary */
6181                 }
6182                 else
6183                 {
6184                         /* replace multi by update xid */
6185                         xid = MultiXactIdGetUpdateXid(multi, t_infomask);
6186
6187                         /* wasn't only a lock, xid needs to be valid */
6188                         Assert(TransactionIdIsValid(xid));
6189
6190                         /*
6191                          * If the xid is older than the cutoff, it has to have aborted,
6192                          * otherwise the tuple would have gotten pruned away.
6193                          */
6194                         if (TransactionIdPrecedes(xid, cutoff_xid))
6195                         {
6196                                 Assert(!TransactionIdDidCommit(xid));
6197                                 *flags |= FRM_INVALIDATE_XMAX;
6198                                 xid = InvalidTransactionId;             /* not strictly necessary */
6199                         }
6200                         else
6201                         {
6202                                 *flags |= FRM_RETURN_IS_XID;
6203                         }
6204                 }
6205
6206                 return xid;
6207         }
6208
6209         /*
6210          * This multixact might have or might not have members still running, but
6211          * we know it's valid and is newer than the cutoff point for multis.
6212          * However, some member(s) of it may be below the cutoff for Xids, so we
6213          * need to walk the whole members array to figure out what to do, if
6214          * anything.
6215          */
6216
6217         allow_old = !(t_infomask & HEAP_LOCK_MASK) &&
6218                 HEAP_XMAX_IS_LOCKED_ONLY(t_infomask);
6219         nmembers =
6220                 GetMultiXactIdMembers(multi, &members, allow_old,
6221                                                           HEAP_XMAX_IS_LOCKED_ONLY(t_infomask));
6222         if (nmembers <= 0)
6223         {
6224                 /* Nothing worth keeping */
6225                 *flags |= FRM_INVALIDATE_XMAX;
6226                 return InvalidTransactionId;
6227         }
6228
6229         /* is there anything older than the cutoff? */
6230         need_replace = false;
6231         for (i = 0; i < nmembers; i++)
6232         {
6233                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
6234                 {
6235                         need_replace = true;
6236                         break;
6237                 }
6238         }
6239
6240         /*
6241          * In the simplest case, there is no member older than the cutoff; we can
6242          * keep the existing MultiXactId as is.
6243          */
6244         if (!need_replace)
6245         {
6246                 *flags |= FRM_NOOP;
6247                 pfree(members);
6248                 return InvalidTransactionId;
6249         }
6250
6251         /*
6252          * If the multi needs to be updated, figure out which members do we need
6253          * to keep.
6254          */
6255         nnewmembers = 0;
6256         newmembers = palloc(sizeof(MultiXactMember) * nmembers);
6257         has_lockers = false;
6258         update_xid = InvalidTransactionId;
6259         update_committed = false;
6260
6261         for (i = 0; i < nmembers; i++)
6262         {
6263                 /*
6264                  * Determine whether to keep this member or ignore it.
6265                  */
6266                 if (ISUPDATE_from_mxstatus(members[i].status))
6267                 {
6268                         TransactionId xid = members[i].xid;
6269
6270                         /*
6271                          * It's an update; should we keep it?  If the transaction is known
6272                          * aborted or crashed then it's okay to ignore it, otherwise not.
6273                          * Note that an updater older than cutoff_xid cannot possibly be
6274                          * committed, because HeapTupleSatisfiesVacuum would have returned
6275                          * HEAPTUPLE_DEAD and we would not be trying to freeze the tuple.
6276                          *
6277                          * As with all tuple visibility routines, it's critical to test
6278                          * TransactionIdIsInProgress before TransactionIdDidCommit,
6279                          * because of race conditions explained in detail in tqual.c.
6280                          */
6281                         if (TransactionIdIsCurrentTransactionId(xid) ||
6282                                 TransactionIdIsInProgress(xid))
6283                         {
6284                                 Assert(!TransactionIdIsValid(update_xid));
6285                                 update_xid = xid;
6286                         }
6287                         else if (TransactionIdDidCommit(xid))
6288                         {
6289                                 /*
6290                                  * The transaction committed, so we can tell caller to set
6291                                  * HEAP_XMAX_COMMITTED.  (We can only do this because we know
6292                                  * the transaction is not running.)
6293                                  */
6294                                 Assert(!TransactionIdIsValid(update_xid));
6295                                 update_committed = true;
6296                                 update_xid = xid;
6297                         }
6298
6299                         /*
6300                          * Not in progress, not committed -- must be aborted or crashed;
6301                          * we can ignore it.
6302                          */
6303
6304                         /*
6305                          * Since the tuple wasn't marked HEAPTUPLE_DEAD by vacuum, the
6306                          * update Xid cannot possibly be older than the xid cutoff.
6307                          */
6308                         Assert(!TransactionIdIsValid(update_xid) ||
6309                                    !TransactionIdPrecedes(update_xid, cutoff_xid));
6310
6311                         /*
6312                          * If we determined that it's an Xid corresponding to an update
6313                          * that must be retained, additionally add it to the list of
6314                          * members of the new Multi, in case we end up using that.  (We
6315                          * might still decide to use only an update Xid and not a multi,
6316                          * but it's easier to maintain the list as we walk the old members
6317                          * list.)
6318                          */
6319                         if (TransactionIdIsValid(update_xid))
6320                                 newmembers[nnewmembers++] = members[i];
6321                 }
6322                 else
6323                 {
6324                         /* We only keep lockers if they are still running */
6325                         if (TransactionIdIsCurrentTransactionId(members[i].xid) ||
6326                                 TransactionIdIsInProgress(members[i].xid))
6327                         {
6328                                 /* running locker cannot possibly be older than the cutoff */
6329                                 Assert(!TransactionIdPrecedes(members[i].xid, cutoff_xid));
6330                                 newmembers[nnewmembers++] = members[i];
6331                                 has_lockers = true;
6332                         }
6333                 }
6334         }
6335
6336         pfree(members);
6337
6338         if (nnewmembers == 0)
6339         {
6340                 /* nothing worth keeping!? Tell caller to remove the whole thing */
6341                 *flags |= FRM_INVALIDATE_XMAX;
6342                 xid = InvalidTransactionId;
6343         }
6344         else if (TransactionIdIsValid(update_xid) && !has_lockers)
6345         {
6346                 /*
6347                  * If there's a single member and it's an update, pass it back alone
6348                  * without creating a new Multi.  (XXX we could do this when there's a
6349                  * single remaining locker, too, but that would complicate the API too
6350                  * much; moreover, the case with the single updater is more
6351                  * interesting, because those are longer-lived.)
6352                  */
6353                 Assert(nnewmembers == 1);
6354                 *flags |= FRM_RETURN_IS_XID;
6355                 if (update_committed)
6356                         *flags |= FRM_MARK_COMMITTED;
6357                 xid = update_xid;
6358         }
6359         else
6360         {
6361                 /*
6362                  * Create a new multixact with the surviving members of the previous
6363                  * one, to set as new Xmax in the tuple.
6364                  */
6365                 xid = MultiXactIdCreateFromMembers(nnewmembers, newmembers);
6366                 *flags |= FRM_RETURN_IS_MULTI;
6367         }
6368
6369         pfree(newmembers);
6370
6371         return xid;
6372 }
6373
6374 /*
6375  * heap_prepare_freeze_tuple
6376  *
6377  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6378  * are older than the specified cutoff XID and cutoff MultiXactId.  If so,
6379  * setup enough state (in the *frz output argument) to later execute and
6380  * WAL-log what we would need to do, and return TRUE.  Return FALSE if nothing
6381  * is to be changed.
6382  *
6383  * Caller is responsible for setting the offset field, if appropriate.
6384  *
6385  * It is assumed that the caller has checked the tuple with
6386  * HeapTupleSatisfiesVacuum() and determined that it is not HEAPTUPLE_DEAD
6387  * (else we should be removing the tuple, not freezing it).
6388  *
6389  * NB: cutoff_xid *must* be <= the current global xmin, to ensure that any
6390  * XID older than it could neither be running nor seen as running by any
6391  * open transaction.  This ensures that the replacement will not change
6392  * anyone's idea of the tuple state.
6393  * Similarly, cutoff_multi must be less than or equal to the smallest
6394  * MultiXactId used by any transaction currently open.
6395  *
6396  * If the tuple is in a shared buffer, caller must hold an exclusive lock on
6397  * that buffer.
6398  *
6399  * NB: It is not enough to set hint bits to indicate something is
6400  * committed/invalid -- they might not be set on a standby, or after crash
6401  * recovery.  We really need to remove old xids.
6402  */
6403 bool
6404 heap_prepare_freeze_tuple(HeapTupleHeader tuple, TransactionId cutoff_xid,
6405                                                   TransactionId cutoff_multi,
6406                                                   xl_heap_freeze_tuple *frz)
6407
6408 {
6409         bool            changed = false;
6410         bool            freeze_xmax = false;
6411         TransactionId xid;
6412
6413         frz->frzflags = 0;
6414         frz->t_infomask2 = tuple->t_infomask2;
6415         frz->t_infomask = tuple->t_infomask;
6416         frz->xmax = HeapTupleHeaderGetRawXmax(tuple);
6417
6418         /* Process xmin */
6419         xid = HeapTupleHeaderGetXmin(tuple);
6420         if (TransactionIdIsNormal(xid) &&
6421                 TransactionIdPrecedes(xid, cutoff_xid))
6422         {
6423                 frz->t_infomask |= HEAP_XMIN_FROZEN;
6424                 changed = true;
6425         }
6426
6427         /*
6428          * Process xmax.  To thoroughly examine the current Xmax value we need to
6429          * resolve a MultiXactId to its member Xids, in case some of them are
6430          * below the given cutoff for Xids.  In that case, those values might need
6431          * freezing, too.  Also, if a multi needs freezing, we cannot simply take
6432          * it out --- if there's a live updater Xid, it needs to be kept.
6433          *
6434          * Make sure to keep heap_tuple_needs_freeze in sync with this.
6435          */
6436         xid = HeapTupleHeaderGetRawXmax(tuple);
6437
6438         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
6439         {
6440                 TransactionId newxmax;
6441                 uint16          flags;
6442
6443                 newxmax = FreezeMultiXactId(xid, tuple->t_infomask,
6444                                                                         cutoff_xid, cutoff_multi, &flags);
6445
6446                 if (flags & FRM_INVALIDATE_XMAX)
6447                         freeze_xmax = true;
6448                 else if (flags & FRM_RETURN_IS_XID)
6449                 {
6450                         /*
6451                          * NB -- some of these transformations are only valid because we
6452                          * know the return Xid is a tuple updater (i.e. not merely a
6453                          * locker.) Also note that the only reason we don't explicitly
6454                          * worry about HEAP_KEYS_UPDATED is because it lives in
6455                          * t_infomask2 rather than t_infomask.
6456                          */
6457                         frz->t_infomask &= ~HEAP_XMAX_BITS;
6458                         frz->xmax = newxmax;
6459                         if (flags & FRM_MARK_COMMITTED)
6460                                 frz->t_infomask &= HEAP_XMAX_COMMITTED;
6461                         changed = true;
6462                 }
6463                 else if (flags & FRM_RETURN_IS_MULTI)
6464                 {
6465                         uint16          newbits;
6466                         uint16          newbits2;
6467
6468                         /*
6469                          * We can't use GetMultiXactIdHintBits directly on the new multi
6470                          * here; that routine initializes the masks to all zeroes, which
6471                          * would lose other bits we need.  Doing it this way ensures all
6472                          * unrelated bits remain untouched.
6473                          */
6474                         frz->t_infomask &= ~HEAP_XMAX_BITS;
6475                         frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
6476                         GetMultiXactIdHintBits(newxmax, &newbits, &newbits2);
6477                         frz->t_infomask |= newbits;
6478                         frz->t_infomask2 |= newbits2;
6479
6480                         frz->xmax = newxmax;
6481
6482                         changed = true;
6483                 }
6484                 else
6485                 {
6486                         Assert(flags & FRM_NOOP);
6487                 }
6488         }
6489         else if (TransactionIdIsNormal(xid) &&
6490                          TransactionIdPrecedes(xid, cutoff_xid))
6491         {
6492                 freeze_xmax = true;
6493         }
6494
6495         if (freeze_xmax)
6496         {
6497                 frz->xmax = InvalidTransactionId;
6498
6499                 /*
6500                  * The tuple might be marked either XMAX_INVALID or XMAX_COMMITTED +
6501                  * LOCKED.  Normalize to INVALID just to be sure no one gets confused.
6502                  * Also get rid of the HEAP_KEYS_UPDATED bit.
6503                  */
6504                 frz->t_infomask &= ~HEAP_XMAX_BITS;
6505                 frz->t_infomask |= HEAP_XMAX_INVALID;
6506                 frz->t_infomask2 &= ~HEAP_HOT_UPDATED;
6507                 frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
6508                 changed = true;
6509         }
6510
6511         /*
6512          * Old-style VACUUM FULL is gone, but we have to keep this code as long as
6513          * we support having MOVED_OFF/MOVED_IN tuples in the database.
6514          */
6515         if (tuple->t_infomask & HEAP_MOVED)
6516         {
6517                 xid = HeapTupleHeaderGetXvac(tuple);
6518                 if (TransactionIdIsNormal(xid) &&
6519                         TransactionIdPrecedes(xid, cutoff_xid))
6520                 {
6521                         /*
6522                          * If a MOVED_OFF tuple is not dead, the xvac transaction must
6523                          * have failed; whereas a non-dead MOVED_IN tuple must mean the
6524                          * xvac transaction succeeded.
6525                          */
6526                         if (tuple->t_infomask & HEAP_MOVED_OFF)
6527                                 frz->frzflags |= XLH_INVALID_XVAC;
6528                         else
6529                                 frz->frzflags |= XLH_FREEZE_XVAC;
6530
6531                         /*
6532                          * Might as well fix the hint bits too; usually XMIN_COMMITTED
6533                          * will already be set here, but there's a small chance not.
6534                          */
6535                         Assert(!(tuple->t_infomask & HEAP_XMIN_INVALID));
6536                         frz->t_infomask |= HEAP_XMIN_COMMITTED;
6537                         changed = true;
6538                 }
6539         }
6540
6541         return changed;
6542 }
6543
6544 /*
6545  * heap_execute_freeze_tuple
6546  *              Execute the prepared freezing of a tuple.
6547  *
6548  * Caller is responsible for ensuring that no other backend can access the
6549  * storage underlying this tuple, either by holding an exclusive lock on the
6550  * buffer containing it (which is what lazy VACUUM does), or by having it be
6551  * in private storage (which is what CLUSTER and friends do).
6552  *
6553  * Note: it might seem we could make the changes without exclusive lock, since
6554  * TransactionId read/write is assumed atomic anyway.  However there is a race
6555  * condition: someone who just fetched an old XID that we overwrite here could
6556  * conceivably not finish checking the XID against pg_clog before we finish
6557  * the VACUUM and perhaps truncate off the part of pg_clog he needs.  Getting
6558  * exclusive lock ensures no other backend is in process of checking the
6559  * tuple status.  Also, getting exclusive lock makes it safe to adjust the
6560  * infomask bits.
6561  *
6562  * NB: All code in here must be safe to execute during crash recovery!
6563  */
6564 void
6565 heap_execute_freeze_tuple(HeapTupleHeader tuple, xl_heap_freeze_tuple *frz)
6566 {
6567         HeapTupleHeaderSetXmax(tuple, frz->xmax);
6568
6569         if (frz->frzflags & XLH_FREEZE_XVAC)
6570                 HeapTupleHeaderSetXvac(tuple, FrozenTransactionId);
6571
6572         if (frz->frzflags & XLH_INVALID_XVAC)
6573                 HeapTupleHeaderSetXvac(tuple, InvalidTransactionId);
6574
6575         tuple->t_infomask = frz->t_infomask;
6576         tuple->t_infomask2 = frz->t_infomask2;
6577 }
6578
6579 /*
6580  * heap_freeze_tuple
6581  *              Freeze tuple in place, without WAL logging.
6582  *
6583  * Useful for callers like CLUSTER that perform their own WAL logging.
6584  */
6585 bool
6586 heap_freeze_tuple(HeapTupleHeader tuple, TransactionId cutoff_xid,
6587                                   TransactionId cutoff_multi)
6588 {
6589         xl_heap_freeze_tuple frz;
6590         bool            do_freeze;
6591
6592         do_freeze = heap_prepare_freeze_tuple(tuple, cutoff_xid, cutoff_multi,
6593                                                                                   &frz);
6594
6595         /*
6596          * Note that because this is not a WAL-logged operation, we don't need to
6597          * fill in the offset in the freeze record.
6598          */
6599
6600         if (do_freeze)
6601                 heap_execute_freeze_tuple(tuple, &frz);
6602         return do_freeze;
6603 }
6604
6605 /*
6606  * For a given MultiXactId, return the hint bits that should be set in the
6607  * tuple's infomask.
6608  *
6609  * Normally this should be called for a multixact that was just created, and
6610  * so is on our local cache, so the GetMembers call is fast.
6611  */
6612 static void
6613 GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
6614                                            uint16 *new_infomask2)
6615 {
6616         int                     nmembers;
6617         MultiXactMember *members;
6618         int                     i;
6619         uint16          bits = HEAP_XMAX_IS_MULTI;
6620         uint16          bits2 = 0;
6621         bool            has_update = false;
6622         LockTupleMode strongest = LockTupleKeyShare;
6623
6624         /*
6625          * We only use this in multis we just created, so they cannot be values
6626          * pre-pg_upgrade.
6627          */
6628         nmembers = GetMultiXactIdMembers(multi, &members, false, false);
6629
6630         for (i = 0; i < nmembers; i++)
6631         {
6632                 LockTupleMode mode;
6633
6634                 /*
6635                  * Remember the strongest lock mode held by any member of the
6636                  * multixact.
6637                  */
6638                 mode = TUPLOCK_from_mxstatus(members[i].status);
6639                 if (mode > strongest)
6640                         strongest = mode;
6641
6642                 /* See what other bits we need */
6643                 switch (members[i].status)
6644                 {
6645                         case MultiXactStatusForKeyShare:
6646                         case MultiXactStatusForShare:
6647                         case MultiXactStatusForNoKeyUpdate:
6648                                 break;
6649
6650                         case MultiXactStatusForUpdate:
6651                                 bits2 |= HEAP_KEYS_UPDATED;
6652                                 break;
6653
6654                         case MultiXactStatusNoKeyUpdate:
6655                                 has_update = true;
6656                                 break;
6657
6658                         case MultiXactStatusUpdate:
6659                                 bits2 |= HEAP_KEYS_UPDATED;
6660                                 has_update = true;
6661                                 break;
6662                 }
6663         }
6664
6665         if (strongest == LockTupleExclusive ||
6666                 strongest == LockTupleNoKeyExclusive)
6667                 bits |= HEAP_XMAX_EXCL_LOCK;
6668         else if (strongest == LockTupleShare)
6669                 bits |= HEAP_XMAX_SHR_LOCK;
6670         else if (strongest == LockTupleKeyShare)
6671                 bits |= HEAP_XMAX_KEYSHR_LOCK;
6672
6673         if (!has_update)
6674                 bits |= HEAP_XMAX_LOCK_ONLY;
6675
6676         if (nmembers > 0)
6677                 pfree(members);
6678
6679         *new_infomask = bits;
6680         *new_infomask2 = bits2;
6681 }
6682
6683 /*
6684  * MultiXactIdGetUpdateXid
6685  *
6686  * Given a multixact Xmax and corresponding infomask, which does not have the
6687  * HEAP_XMAX_LOCK_ONLY bit set, obtain and return the Xid of the updating
6688  * transaction.
6689  *
6690  * Caller is expected to check the status of the updating transaction, if
6691  * necessary.
6692  */
6693 static TransactionId
6694 MultiXactIdGetUpdateXid(TransactionId xmax, uint16 t_infomask)
6695 {
6696         TransactionId update_xact = InvalidTransactionId;
6697         MultiXactMember *members;
6698         int                     nmembers;
6699
6700         Assert(!(t_infomask & HEAP_XMAX_LOCK_ONLY));
6701         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6702
6703         /*
6704          * Since we know the LOCK_ONLY bit is not set, this cannot be a multi from
6705          * pre-pg_upgrade.
6706          */
6707         nmembers = GetMultiXactIdMembers(xmax, &members, false, false);
6708
6709         if (nmembers > 0)
6710         {
6711                 int                     i;
6712
6713                 for (i = 0; i < nmembers; i++)
6714                 {
6715                         /* Ignore lockers */
6716                         if (!ISUPDATE_from_mxstatus(members[i].status))
6717                                 continue;
6718
6719                         /* there can be at most one updater */
6720                         Assert(update_xact == InvalidTransactionId);
6721                         update_xact = members[i].xid;
6722 #ifndef USE_ASSERT_CHECKING
6723
6724                         /*
6725                          * in an assert-enabled build, walk the whole array to ensure
6726                          * there's no other updater.
6727                          */
6728                         break;
6729 #endif
6730                 }
6731
6732                 pfree(members);
6733         }
6734
6735         return update_xact;
6736 }
6737
6738 /*
6739  * HeapTupleGetUpdateXid
6740  *              As above, but use a HeapTupleHeader
6741  *
6742  * See also HeapTupleHeaderGetUpdateXid, which can be used without previously
6743  * checking the hint bits.
6744  */
6745 TransactionId
6746 HeapTupleGetUpdateXid(HeapTupleHeader tuple)
6747 {
6748         return MultiXactIdGetUpdateXid(HeapTupleHeaderGetRawXmax(tuple),
6749                                                                    tuple->t_infomask);
6750 }
6751
6752 /*
6753  * Does the given multixact conflict with the current transaction grabbing a
6754  * tuple lock of the given strength?
6755  *
6756  * The passed infomask pairs up with the given multixact in the tuple header.
6757  */
6758 static bool
6759 DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
6760                                                 LockTupleMode lockmode)
6761 {
6762         bool            allow_old;
6763         int                     nmembers;
6764         MultiXactMember *members;
6765         bool            result = false;
6766         LOCKMODE        wanted = tupleLockExtraInfo[lockmode].hwlock;
6767
6768         allow_old = !(infomask & HEAP_LOCK_MASK) && HEAP_XMAX_IS_LOCKED_ONLY(infomask);
6769         nmembers = GetMultiXactIdMembers(multi, &members, allow_old,
6770                                                                          HEAP_XMAX_IS_LOCKED_ONLY(infomask));
6771         if (nmembers >= 0)
6772         {
6773                 int                     i;
6774
6775                 for (i = 0; i < nmembers; i++)
6776                 {
6777                         TransactionId memxid;
6778                         LOCKMODE        memlockmode;
6779
6780                         memlockmode = LOCKMODE_from_mxstatus(members[i].status);
6781
6782                         /* ignore members that don't conflict with the lock we want */
6783                         if (!DoLockModesConflict(memlockmode, wanted))
6784                                 continue;
6785
6786                         /* ignore members from current xact */
6787                         memxid = members[i].xid;
6788                         if (TransactionIdIsCurrentTransactionId(memxid))
6789                                 continue;
6790
6791                         if (ISUPDATE_from_mxstatus(members[i].status))
6792                         {
6793                                 /* ignore aborted updaters */
6794                                 if (TransactionIdDidAbort(memxid))
6795                                         continue;
6796                         }
6797                         else
6798                         {
6799                                 /* ignore lockers-only that are no longer in progress */
6800                                 if (!TransactionIdIsInProgress(memxid))
6801                                         continue;
6802                         }
6803
6804                         /*
6805                          * Whatever remains are either live lockers that conflict with our
6806                          * wanted lock, and updaters that are not aborted.  Those conflict
6807                          * with what we want, so return true.
6808                          */
6809                         result = true;
6810                         break;
6811                 }
6812                 pfree(members);
6813         }
6814
6815         return result;
6816 }
6817
6818 /*
6819  * Do_MultiXactIdWait
6820  *              Actual implementation for the two functions below.
6821  *
6822  * 'multi', 'status' and 'infomask' indicate what to sleep on (the status is
6823  * needed to ensure we only sleep on conflicting members, and the infomask is
6824  * used to optimize multixact access in case it's a lock-only multi); 'nowait'
6825  * indicates whether to use conditional lock acquisition, to allow callers to
6826  * fail if lock is unavailable.  'rel', 'ctid' and 'oper' are used to set up
6827  * context information for error messages.  'remaining', if not NULL, receives
6828  * the number of members that are still running, including any (non-aborted)
6829  * subtransactions of our own transaction.
6830  *
6831  * We do this by sleeping on each member using XactLockTableWait.  Any
6832  * members that belong to the current backend are *not* waited for, however;
6833  * this would not merely be useless but would lead to Assert failure inside
6834  * XactLockTableWait.  By the time this returns, it is certain that all
6835  * transactions *of other backends* that were members of the MultiXactId
6836  * that conflict with the requested status are dead (and no new ones can have
6837  * been added, since it is not legal to add members to an existing
6838  * MultiXactId).
6839  *
6840  * But by the time we finish sleeping, someone else may have changed the Xmax
6841  * of the containing tuple, so the caller needs to iterate on us somehow.
6842  *
6843  * Note that in case we return false, the number of remaining members is
6844  * not to be trusted.
6845  */
6846 static bool
6847 Do_MultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6848                                    uint16 infomask, bool nowait,
6849                                    Relation rel, ItemPointer ctid, XLTW_Oper oper,
6850                                    int *remaining)
6851 {
6852         bool            allow_old;
6853         bool            result = true;
6854         MultiXactMember *members;
6855         int                     nmembers;
6856         int                     remain = 0;
6857
6858         allow_old = !(infomask & HEAP_LOCK_MASK) && HEAP_XMAX_IS_LOCKED_ONLY(infomask);
6859         nmembers = GetMultiXactIdMembers(multi, &members, allow_old,
6860                                                                          HEAP_XMAX_IS_LOCKED_ONLY(infomask));
6861
6862         if (nmembers >= 0)
6863         {
6864                 int                     i;
6865
6866                 for (i = 0; i < nmembers; i++)
6867                 {
6868                         TransactionId memxid = members[i].xid;
6869                         MultiXactStatus memstatus = members[i].status;
6870
6871                         if (TransactionIdIsCurrentTransactionId(memxid))
6872                         {
6873                                 remain++;
6874                                 continue;
6875                         }
6876
6877                         if (!DoLockModesConflict(LOCKMODE_from_mxstatus(memstatus),
6878                                                                          LOCKMODE_from_mxstatus(status)))
6879                         {
6880                                 if (remaining && TransactionIdIsInProgress(memxid))
6881                                         remain++;
6882                                 continue;
6883                         }
6884
6885                         /*
6886                          * This member conflicts with our multi, so we have to sleep (or
6887                          * return failure, if asked to avoid waiting.)
6888                          *
6889                          * Note that we don't set up an error context callback ourselves,
6890                          * but instead we pass the info down to XactLockTableWait.  This
6891                          * might seem a bit wasteful because the context is set up and
6892                          * tore down for each member of the multixact, but in reality it
6893                          * should be barely noticeable, and it avoids duplicate code.
6894                          */
6895                         if (nowait)
6896                         {
6897                                 result = ConditionalXactLockTableWait(memxid);
6898                                 if (!result)
6899                                         break;
6900                         }
6901                         else
6902                                 XactLockTableWait(memxid, rel, ctid, oper);
6903                 }
6904
6905                 pfree(members);
6906         }
6907
6908         if (remaining)
6909                 *remaining = remain;
6910
6911         return result;
6912 }
6913
6914 /*
6915  * MultiXactIdWait
6916  *              Sleep on a MultiXactId.
6917  *
6918  * By the time we finish sleeping, someone else may have changed the Xmax
6919  * of the containing tuple, so the caller needs to iterate on us somehow.
6920  *
6921  * We return (in *remaining, if not NULL) the number of members that are still
6922  * running, including any (non-aborted) subtransactions of our own transaction.
6923  */
6924 static void
6925 MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
6926                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
6927                                 int *remaining)
6928 {
6929         (void) Do_MultiXactIdWait(multi, status, infomask, false,
6930                                                           rel, ctid, oper, remaining);
6931 }
6932
6933 /*
6934  * ConditionalMultiXactIdWait
6935  *              As above, but only lock if we can get the lock without blocking.
6936  *
6937  * By the time we finish sleeping, someone else may have changed the Xmax
6938  * of the containing tuple, so the caller needs to iterate on us somehow.
6939  *
6940  * If the multixact is now all gone, return true.  Returns false if some
6941  * transactions might still be running.
6942  *
6943  * We return (in *remaining, if not NULL) the number of members that are still
6944  * running, including any (non-aborted) subtransactions of our own transaction.
6945  */
6946 static bool
6947 ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6948                                                    uint16 infomask, Relation rel, int *remaining)
6949 {
6950         return Do_MultiXactIdWait(multi, status, infomask, true,
6951                                                           rel, NULL, XLTW_None, remaining);
6952 }
6953
6954 /*
6955  * heap_tuple_needs_eventual_freeze
6956  *
6957  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6958  * will eventually require freezing.  Similar to heap_tuple_needs_freeze,
6959  * but there's no cutoff, since we're trying to figure out whether freezing
6960  * will ever be needed, not whether it's needed now.
6961  */
6962 bool
6963 heap_tuple_needs_eventual_freeze(HeapTupleHeader tuple)
6964 {
6965         TransactionId xid;
6966
6967         /*
6968          * If xmin is a normal transaction ID, this tuple is definitely not
6969          * frozen.
6970          */
6971         xid = HeapTupleHeaderGetXmin(tuple);
6972         if (TransactionIdIsNormal(xid))
6973                 return true;
6974
6975         /*
6976          * If xmax is a valid xact or multixact, this tuple is also not frozen.
6977          */
6978         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
6979         {
6980                 MultiXactId multi;
6981
6982                 multi = HeapTupleHeaderGetRawXmax(tuple);
6983                 if (MultiXactIdIsValid(multi))
6984                         return true;
6985         }
6986         else
6987         {
6988                 xid = HeapTupleHeaderGetRawXmax(tuple);
6989                 if (TransactionIdIsNormal(xid))
6990                         return true;
6991         }
6992
6993         if (tuple->t_infomask & HEAP_MOVED)
6994         {
6995                 xid = HeapTupleHeaderGetXvac(tuple);
6996                 if (TransactionIdIsNormal(xid))
6997                         return true;
6998         }
6999
7000         return false;
7001 }
7002
7003 /*
7004  * heap_tuple_needs_freeze
7005  *
7006  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
7007  * are older than the specified cutoff XID or MultiXactId.  If so, return TRUE.
7008  *
7009  * It doesn't matter whether the tuple is alive or dead, we are checking
7010  * to see if a tuple needs to be removed or frozen to avoid wraparound.
7011  *
7012  * NB: Cannot rely on hint bits here, they might not be set after a crash or
7013  * on a standby.
7014  */
7015 bool
7016 heap_tuple_needs_freeze(HeapTupleHeader tuple, TransactionId cutoff_xid,
7017                                                 MultiXactId cutoff_multi, Buffer buf)
7018 {
7019         TransactionId xid;
7020
7021         xid = HeapTupleHeaderGetXmin(tuple);
7022         if (TransactionIdIsNormal(xid) &&
7023                 TransactionIdPrecedes(xid, cutoff_xid))
7024                 return true;
7025
7026         /*
7027          * The considerations for multixacts are complicated; look at
7028          * heap_freeze_tuple for justifications.  This routine had better be in
7029          * sync with that one!
7030          */
7031         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
7032         {
7033                 MultiXactId multi;
7034
7035                 multi = HeapTupleHeaderGetRawXmax(tuple);
7036                 if (!MultiXactIdIsValid(multi))
7037                 {
7038                         /* no xmax set, ignore */
7039                         ;
7040                 }
7041                 else if (MultiXactIdPrecedes(multi, cutoff_multi))
7042                         return true;
7043                 else
7044                 {
7045                         MultiXactMember *members;
7046                         int                     nmembers;
7047                         int                     i;
7048                         bool            allow_old;
7049
7050                         /* need to check whether any member of the mxact is too old */
7051
7052                         allow_old = !(tuple->t_infomask & HEAP_LOCK_MASK) &&
7053                                 HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask);
7054                         nmembers = GetMultiXactIdMembers(multi, &members, allow_old,
7055                                                                 HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask));
7056
7057                         for (i = 0; i < nmembers; i++)
7058                         {
7059                                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
7060                                 {
7061                                         pfree(members);
7062                                         return true;
7063                                 }
7064                         }
7065                         if (nmembers > 0)
7066                                 pfree(members);
7067                 }
7068         }
7069         else
7070         {
7071                 xid = HeapTupleHeaderGetRawXmax(tuple);
7072                 if (TransactionIdIsNormal(xid) &&
7073                         TransactionIdPrecedes(xid, cutoff_xid))
7074                         return true;
7075         }
7076
7077         if (tuple->t_infomask & HEAP_MOVED)
7078         {
7079                 xid = HeapTupleHeaderGetXvac(tuple);
7080                 if (TransactionIdIsNormal(xid) &&
7081                         TransactionIdPrecedes(xid, cutoff_xid))
7082                         return true;
7083         }
7084
7085         return false;
7086 }
7087
7088 /*
7089  * If 'tuple' contains any visible XID greater than latestRemovedXid,
7090  * ratchet forwards latestRemovedXid to the greatest one found.
7091  * This is used as the basis for generating Hot Standby conflicts, so
7092  * if a tuple was never visible then removing it should not conflict
7093  * with queries.
7094  */
7095 void
7096 HeapTupleHeaderAdvanceLatestRemovedXid(HeapTupleHeader tuple,
7097                                                                            TransactionId *latestRemovedXid)
7098 {
7099         TransactionId xmin = HeapTupleHeaderGetXmin(tuple);
7100         TransactionId xmax = HeapTupleHeaderGetUpdateXid(tuple);
7101         TransactionId xvac = HeapTupleHeaderGetXvac(tuple);
7102
7103         if (tuple->t_infomask & HEAP_MOVED)
7104         {
7105                 if (TransactionIdPrecedes(*latestRemovedXid, xvac))
7106                         *latestRemovedXid = xvac;
7107         }
7108
7109         /*
7110          * Ignore tuples inserted by an aborted transaction or if the tuple was
7111          * updated/deleted by the inserting transaction.
7112          *
7113          * Look for a committed hint bit, or if no xmin bit is set, check clog.
7114          * This needs to work on both master and standby, where it is used to
7115          * assess btree delete records.
7116          */
7117         if (HeapTupleHeaderXminCommitted(tuple) ||
7118                 (!HeapTupleHeaderXminInvalid(tuple) && TransactionIdDidCommit(xmin)))
7119         {
7120                 if (xmax != xmin &&
7121                         TransactionIdFollows(xmax, *latestRemovedXid))
7122                         *latestRemovedXid = xmax;
7123         }
7124
7125         /* *latestRemovedXid may still be invalid at end */
7126 }
7127
7128 /*
7129  * Perform XLogInsert to register a heap cleanup info message. These
7130  * messages are sent once per VACUUM and are required because
7131  * of the phasing of removal operations during a lazy VACUUM.
7132  * see comments for vacuum_log_cleanup_info().
7133  */
7134 XLogRecPtr
7135 log_heap_cleanup_info(RelFileNode rnode, TransactionId latestRemovedXid)
7136 {
7137         xl_heap_cleanup_info xlrec;
7138         XLogRecPtr      recptr;
7139
7140         xlrec.node = rnode;
7141         xlrec.latestRemovedXid = latestRemovedXid;
7142
7143         XLogBeginInsert();
7144         XLogRegisterData((char *) &xlrec, SizeOfHeapCleanupInfo);
7145
7146         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_CLEANUP_INFO);
7147
7148         return recptr;
7149 }
7150
7151 /*
7152  * Perform XLogInsert for a heap-clean operation.  Caller must already
7153  * have modified the buffer and marked it dirty.
7154  *
7155  * Note: prior to Postgres 8.3, the entries in the nowunused[] array were
7156  * zero-based tuple indexes.  Now they are one-based like other uses
7157  * of OffsetNumber.
7158  *
7159  * We also include latestRemovedXid, which is the greatest XID present in
7160  * the removed tuples. That allows recovery processing to cancel or wait
7161  * for long standby queries that can still see these tuples.
7162  */
7163 XLogRecPtr
7164 log_heap_clean(Relation reln, Buffer buffer,
7165                            OffsetNumber *redirected, int nredirected,
7166                            OffsetNumber *nowdead, int ndead,
7167                            OffsetNumber *nowunused, int nunused,
7168                            TransactionId latestRemovedXid)
7169 {
7170         xl_heap_clean xlrec;
7171         XLogRecPtr      recptr;
7172
7173         /* Caller should not call me on a non-WAL-logged relation */
7174         Assert(RelationNeedsWAL(reln));
7175
7176         xlrec.latestRemovedXid = latestRemovedXid;
7177         xlrec.nredirected = nredirected;
7178         xlrec.ndead = ndead;
7179
7180         XLogBeginInsert();
7181         XLogRegisterData((char *) &xlrec, SizeOfHeapClean);
7182
7183         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
7184
7185         /*
7186          * The OffsetNumber arrays are not actually in the buffer, but we pretend
7187          * that they are.  When XLogInsert stores the whole buffer, the offset
7188          * arrays need not be stored too.  Note that even if all three arrays are
7189          * empty, we want to expose the buffer as a candidate for whole-page
7190          * storage, since this record type implies a defragmentation operation
7191          * even if no item pointers changed state.
7192          */
7193         if (nredirected > 0)
7194                 XLogRegisterBufData(0, (char *) redirected,
7195                                                         nredirected * sizeof(OffsetNumber) * 2);
7196
7197         if (ndead > 0)
7198                 XLogRegisterBufData(0, (char *) nowdead,
7199                                                         ndead * sizeof(OffsetNumber));
7200
7201         if (nunused > 0)
7202                 XLogRegisterBufData(0, (char *) nowunused,
7203                                                         nunused * sizeof(OffsetNumber));
7204
7205         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_CLEAN);
7206
7207         return recptr;
7208 }
7209
7210 /*
7211  * Perform XLogInsert for a heap-freeze operation.  Caller must have already
7212  * modified the buffer and marked it dirty.
7213  */
7214 XLogRecPtr
7215 log_heap_freeze(Relation reln, Buffer buffer, TransactionId cutoff_xid,
7216                                 xl_heap_freeze_tuple *tuples, int ntuples)
7217 {
7218         xl_heap_freeze_page xlrec;
7219         XLogRecPtr      recptr;
7220
7221         /* Caller should not call me on a non-WAL-logged relation */
7222         Assert(RelationNeedsWAL(reln));
7223         /* nor when there are no tuples to freeze */
7224         Assert(ntuples > 0);
7225
7226         xlrec.cutoff_xid = cutoff_xid;
7227         xlrec.ntuples = ntuples;
7228
7229         XLogBeginInsert();
7230         XLogRegisterData((char *) &xlrec, SizeOfHeapFreezePage);
7231
7232         /*
7233          * The freeze plan array is not actually in the buffer, but pretend that
7234          * it is.  When XLogInsert stores the whole buffer, the freeze plan need
7235          * not be stored too.
7236          */
7237         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
7238         XLogRegisterBufData(0, (char *) tuples,
7239                                                 ntuples * sizeof(xl_heap_freeze_tuple));
7240
7241         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_FREEZE_PAGE);
7242
7243         return recptr;
7244 }
7245
7246 /*
7247  * Perform XLogInsert for a heap-visible operation.  'block' is the block
7248  * being marked all-visible, and vm_buffer is the buffer containing the
7249  * corresponding visibility map block.  Both should have already been modified
7250  * and dirtied.
7251  *
7252  * If checksums are enabled, we also generate a full-page image of
7253  * heap_buffer, if necessary.
7254  */
7255 XLogRecPtr
7256 log_heap_visible(RelFileNode rnode, Buffer heap_buffer, Buffer vm_buffer,
7257                                  TransactionId cutoff_xid, uint8 vmflags)
7258 {
7259         xl_heap_visible xlrec;
7260         XLogRecPtr      recptr;
7261         uint8           flags;
7262
7263         Assert(BufferIsValid(heap_buffer));
7264         Assert(BufferIsValid(vm_buffer));
7265
7266         xlrec.cutoff_xid = cutoff_xid;
7267         xlrec.flags = vmflags;
7268         XLogBeginInsert();
7269         XLogRegisterData((char *) &xlrec, SizeOfHeapVisible);
7270
7271         XLogRegisterBuffer(0, vm_buffer, 0);
7272
7273         flags = REGBUF_STANDARD;
7274         if (!XLogHintBitIsNeeded())
7275                 flags |= REGBUF_NO_IMAGE;
7276         XLogRegisterBuffer(1, heap_buffer, flags);
7277
7278         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_VISIBLE);
7279
7280         return recptr;
7281 }
7282
7283 /*
7284  * Perform XLogInsert for a heap-update operation.  Caller must already
7285  * have modified the buffer(s) and marked them dirty.
7286  */
7287 static XLogRecPtr
7288 log_heap_update(Relation reln, Buffer oldbuf,
7289                                 Buffer newbuf, HeapTuple oldtup, HeapTuple newtup,
7290                                 HeapTuple old_key_tuple,
7291                                 bool all_visible_cleared, bool new_all_visible_cleared)
7292 {
7293         xl_heap_update xlrec;
7294         xl_heap_header xlhdr;
7295         xl_heap_header xlhdr_idx;
7296         uint8           info;
7297         uint16          prefix_suffix[2];
7298         uint16          prefixlen = 0,
7299                                 suffixlen = 0;
7300         XLogRecPtr      recptr;
7301         Page            page = BufferGetPage(newbuf);
7302         bool            need_tuple_data = RelationIsLogicallyLogged(reln);
7303         bool            init;
7304         int                     bufflags;
7305
7306         /* Caller should not call me on a non-WAL-logged relation */
7307         Assert(RelationNeedsWAL(reln));
7308
7309         XLogBeginInsert();
7310
7311         if (HeapTupleIsHeapOnly(newtup))
7312                 info = XLOG_HEAP_HOT_UPDATE;
7313         else
7314                 info = XLOG_HEAP_UPDATE;
7315
7316         /*
7317          * If the old and new tuple are on the same page, we only need to log the
7318          * parts of the new tuple that were changed.  That saves on the amount of
7319          * WAL we need to write.  Currently, we just count any unchanged bytes in
7320          * the beginning and end of the tuple.  That's quick to check, and
7321          * perfectly covers the common case that only one field is updated.
7322          *
7323          * We could do this even if the old and new tuple are on different pages,
7324          * but only if we don't make a full-page image of the old page, which is
7325          * difficult to know in advance.  Also, if the old tuple is corrupt for
7326          * some reason, it would allow the corruption to propagate the new page,
7327          * so it seems best to avoid.  Under the general assumption that most
7328          * updates tend to create the new tuple version on the same page, there
7329          * isn't much to be gained by doing this across pages anyway.
7330          *
7331          * Skip this if we're taking a full-page image of the new page, as we
7332          * don't include the new tuple in the WAL record in that case.  Also
7333          * disable if wal_level='logical', as logical decoding needs to be able to
7334          * read the new tuple in whole from the WAL record alone.
7335          */
7336         if (oldbuf == newbuf && !need_tuple_data &&
7337                 !XLogCheckBufferNeedsBackup(newbuf))
7338         {
7339                 char       *oldp = (char *) oldtup->t_data + oldtup->t_data->t_hoff;
7340                 char       *newp = (char *) newtup->t_data + newtup->t_data->t_hoff;
7341                 int                     oldlen = oldtup->t_len - oldtup->t_data->t_hoff;
7342                 int                     newlen = newtup->t_len - newtup->t_data->t_hoff;
7343
7344                 /* Check for common prefix between old and new tuple */
7345                 for (prefixlen = 0; prefixlen < Min(oldlen, newlen); prefixlen++)
7346                 {
7347                         if (newp[prefixlen] != oldp[prefixlen])
7348                                 break;
7349                 }
7350
7351                 /*
7352                  * Storing the length of the prefix takes 2 bytes, so we need to save
7353                  * at least 3 bytes or there's no point.
7354                  */
7355                 if (prefixlen < 3)
7356                         prefixlen = 0;
7357
7358                 /* Same for suffix */
7359                 for (suffixlen = 0; suffixlen < Min(oldlen, newlen) - prefixlen; suffixlen++)
7360                 {
7361                         if (newp[newlen - suffixlen - 1] != oldp[oldlen - suffixlen - 1])
7362                                 break;
7363                 }
7364                 if (suffixlen < 3)
7365                         suffixlen = 0;
7366         }
7367
7368         /* Prepare main WAL data chain */
7369         xlrec.flags = 0;
7370         if (all_visible_cleared)
7371                 xlrec.flags |= XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED;
7372         if (new_all_visible_cleared)
7373                 xlrec.flags |= XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED;
7374         if (prefixlen > 0)
7375                 xlrec.flags |= XLH_UPDATE_PREFIX_FROM_OLD;
7376         if (suffixlen > 0)
7377                 xlrec.flags |= XLH_UPDATE_SUFFIX_FROM_OLD;
7378         if (need_tuple_data)
7379         {
7380                 xlrec.flags |= XLH_UPDATE_CONTAINS_NEW_TUPLE;
7381                 if (old_key_tuple)
7382                 {
7383                         if (reln->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
7384                                 xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_TUPLE;
7385                         else
7386                                 xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_KEY;
7387                 }
7388         }
7389
7390         /* If new tuple is the single and first tuple on page... */
7391         if (ItemPointerGetOffsetNumber(&(newtup->t_self)) == FirstOffsetNumber &&
7392                 PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
7393         {
7394                 info |= XLOG_HEAP_INIT_PAGE;
7395                 init = true;
7396         }
7397         else
7398                 init = false;
7399
7400         /* Prepare WAL data for the old page */
7401         xlrec.old_offnum = ItemPointerGetOffsetNumber(&oldtup->t_self);
7402         xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data);
7403         xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask,
7404                                                                                           oldtup->t_data->t_infomask2);
7405
7406         /* Prepare WAL data for the new page */
7407         xlrec.new_offnum = ItemPointerGetOffsetNumber(&newtup->t_self);
7408         xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data);
7409
7410         bufflags = REGBUF_STANDARD;
7411         if (init)
7412                 bufflags |= REGBUF_WILL_INIT;
7413         if (need_tuple_data)
7414                 bufflags |= REGBUF_KEEP_DATA;
7415
7416         XLogRegisterBuffer(0, newbuf, bufflags);
7417         if (oldbuf != newbuf)
7418                 XLogRegisterBuffer(1, oldbuf, REGBUF_STANDARD);
7419
7420         XLogRegisterData((char *) &xlrec, SizeOfHeapUpdate);
7421
7422         /*
7423          * Prepare WAL data for the new tuple.
7424          */
7425         if (prefixlen > 0 || suffixlen > 0)
7426         {
7427                 if (prefixlen > 0 && suffixlen > 0)
7428                 {
7429                         prefix_suffix[0] = prefixlen;
7430                         prefix_suffix[1] = suffixlen;
7431                         XLogRegisterBufData(0, (char *) &prefix_suffix, sizeof(uint16) * 2);
7432                 }
7433                 else if (prefixlen > 0)
7434                 {
7435                         XLogRegisterBufData(0, (char *) &prefixlen, sizeof(uint16));
7436                 }
7437                 else
7438                 {
7439                         XLogRegisterBufData(0, (char *) &suffixlen, sizeof(uint16));
7440                 }
7441         }
7442
7443         xlhdr.t_infomask2 = newtup->t_data->t_infomask2;
7444         xlhdr.t_infomask = newtup->t_data->t_infomask;
7445         xlhdr.t_hoff = newtup->t_data->t_hoff;
7446         Assert(SizeofHeapTupleHeader + prefixlen + suffixlen <= newtup->t_len);
7447
7448         /*
7449          * PG73FORMAT: write bitmap [+ padding] [+ oid] + data
7450          *
7451          * The 'data' doesn't include the common prefix or suffix.
7452          */
7453         XLogRegisterBufData(0, (char *) &xlhdr, SizeOfHeapHeader);
7454         if (prefixlen == 0)
7455         {
7456                 XLogRegisterBufData(0,
7457                                                         ((char *) newtup->t_data) + SizeofHeapTupleHeader,
7458                                                   newtup->t_len - SizeofHeapTupleHeader - suffixlen);
7459         }
7460         else
7461         {
7462                 /*
7463                  * Have to write the null bitmap and data after the common prefix as
7464                  * two separate rdata entries.
7465                  */
7466                 /* bitmap [+ padding] [+ oid] */
7467                 if (newtup->t_data->t_hoff - SizeofHeapTupleHeader > 0)
7468                 {
7469                         XLogRegisterBufData(0,
7470                                                    ((char *) newtup->t_data) + SizeofHeapTupleHeader,
7471                                                          newtup->t_data->t_hoff - SizeofHeapTupleHeader);
7472                 }
7473
7474                 /* data after common prefix */
7475                 XLogRegisterBufData(0,
7476                           ((char *) newtup->t_data) + newtup->t_data->t_hoff + prefixlen,
7477                          newtup->t_len - newtup->t_data->t_hoff - prefixlen - suffixlen);
7478         }
7479
7480         /* We need to log a tuple identity */
7481         if (need_tuple_data && old_key_tuple)
7482         {
7483                 /* don't really need this, but its more comfy to decode */
7484                 xlhdr_idx.t_infomask2 = old_key_tuple->t_data->t_infomask2;
7485                 xlhdr_idx.t_infomask = old_key_tuple->t_data->t_infomask;
7486                 xlhdr_idx.t_hoff = old_key_tuple->t_data->t_hoff;
7487
7488                 XLogRegisterData((char *) &xlhdr_idx, SizeOfHeapHeader);
7489
7490                 /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
7491                 XLogRegisterData((char *) old_key_tuple->t_data + SizeofHeapTupleHeader,
7492                                                  old_key_tuple->t_len - SizeofHeapTupleHeader);
7493         }
7494
7495         /* filtering by origin on a row level is much more efficient */
7496         XLogIncludeOrigin();
7497
7498         recptr = XLogInsert(RM_HEAP_ID, info);
7499
7500         return recptr;
7501 }
7502
7503 /*
7504  * Perform XLogInsert of an XLOG_HEAP2_NEW_CID record
7505  *
7506  * This is only used in wal_level >= WAL_LEVEL_LOGICAL, and only for catalog
7507  * tuples.
7508  */
7509 static XLogRecPtr
7510 log_heap_new_cid(Relation relation, HeapTuple tup)
7511 {
7512         xl_heap_new_cid xlrec;
7513
7514         XLogRecPtr      recptr;
7515         HeapTupleHeader hdr = tup->t_data;
7516
7517         Assert(ItemPointerIsValid(&tup->t_self));
7518         Assert(tup->t_tableOid != InvalidOid);
7519
7520         xlrec.top_xid = GetTopTransactionId();
7521         xlrec.target_node = relation->rd_node;
7522         xlrec.target_tid = tup->t_self;
7523
7524         /*
7525          * If the tuple got inserted & deleted in the same TX we definitely have a
7526          * combocid, set cmin and cmax.
7527          */
7528         if (hdr->t_infomask & HEAP_COMBOCID)
7529         {
7530                 Assert(!(hdr->t_infomask & HEAP_XMAX_INVALID));
7531                 Assert(!HeapTupleHeaderXminInvalid(hdr));
7532                 xlrec.cmin = HeapTupleHeaderGetCmin(hdr);
7533                 xlrec.cmax = HeapTupleHeaderGetCmax(hdr);
7534                 xlrec.combocid = HeapTupleHeaderGetRawCommandId(hdr);
7535         }
7536         /* No combocid, so only cmin or cmax can be set by this TX */
7537         else
7538         {
7539                 /*
7540                  * Tuple inserted.
7541                  *
7542                  * We need to check for LOCK ONLY because multixacts might be
7543                  * transferred to the new tuple in case of FOR KEY SHARE updates in
7544                  * which case there will be an xmax, although the tuple just got
7545                  * inserted.
7546                  */
7547                 if (hdr->t_infomask & HEAP_XMAX_INVALID ||
7548                         HEAP_XMAX_IS_LOCKED_ONLY(hdr->t_infomask))
7549                 {
7550                         xlrec.cmin = HeapTupleHeaderGetRawCommandId(hdr);
7551                         xlrec.cmax = InvalidCommandId;
7552                 }
7553                 /* Tuple from a different tx updated or deleted. */
7554                 else
7555                 {
7556                         xlrec.cmin = InvalidCommandId;
7557                         xlrec.cmax = HeapTupleHeaderGetRawCommandId(hdr);
7558
7559                 }
7560                 xlrec.combocid = InvalidCommandId;
7561         }
7562
7563         /*
7564          * Note that we don't need to register the buffer here, because this
7565          * operation does not modify the page. The insert/update/delete that
7566          * called us certainly did, but that's WAL-logged separately.
7567          */
7568         XLogBeginInsert();
7569         XLogRegisterData((char *) &xlrec, SizeOfHeapNewCid);
7570
7571         /* will be looked at irrespective of origin */
7572
7573         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_NEW_CID);
7574
7575         return recptr;
7576 }
7577
7578 /*
7579  * Build a heap tuple representing the configured REPLICA IDENTITY to represent
7580  * the old tuple in a UPDATE or DELETE.
7581  *
7582  * Returns NULL if there's no need to log an identity or if there's no suitable
7583  * key in the Relation relation.
7584  */
7585 static HeapTuple
7586 ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_changed, bool *copy)
7587 {
7588         TupleDesc       desc = RelationGetDescr(relation);
7589         Oid                     replidindex;
7590         Relation        idx_rel;
7591         TupleDesc       idx_desc;
7592         char            replident = relation->rd_rel->relreplident;
7593         HeapTuple       key_tuple = NULL;
7594         bool            nulls[MaxHeapAttributeNumber];
7595         Datum           values[MaxHeapAttributeNumber];
7596         int                     natt;
7597
7598         *copy = false;
7599
7600         if (!RelationIsLogicallyLogged(relation))
7601                 return NULL;
7602
7603         if (replident == REPLICA_IDENTITY_NOTHING)
7604                 return NULL;
7605
7606         if (replident == REPLICA_IDENTITY_FULL)
7607         {
7608                 /*
7609                  * When logging the entire old tuple, it very well could contain
7610                  * toasted columns. If so, force them to be inlined.
7611                  */
7612                 if (HeapTupleHasExternal(tp))
7613                 {
7614                         *copy = true;
7615                         tp = toast_flatten_tuple(tp, RelationGetDescr(relation));
7616                 }
7617                 return tp;
7618         }
7619
7620         /* if the key hasn't changed and we're only logging the key, we're done */
7621         if (!key_changed)
7622                 return NULL;
7623
7624         /* find the replica identity index */
7625         replidindex = RelationGetReplicaIndex(relation);
7626         if (!OidIsValid(replidindex))
7627         {
7628                 elog(DEBUG4, "could not find configured replica identity for table \"%s\"",
7629                          RelationGetRelationName(relation));
7630                 return NULL;
7631         }
7632
7633         idx_rel = RelationIdGetRelation(replidindex);
7634         idx_desc = RelationGetDescr(idx_rel);
7635
7636         /* deform tuple, so we have fast access to columns */
7637         heap_deform_tuple(tp, desc, values, nulls);
7638
7639         /* set all columns to NULL, regardless of whether they actually are */
7640         memset(nulls, 1, sizeof(nulls));
7641
7642         /*
7643          * Now set all columns contained in the index to NOT NULL, they cannot
7644          * currently be NULL.
7645          */
7646         for (natt = 0; natt < idx_desc->natts; natt++)
7647         {
7648                 int                     attno = idx_rel->rd_index->indkey.values[natt];
7649
7650                 if (attno < 0)
7651                 {
7652                         /*
7653                          * The OID column can appear in an index definition, but that's
7654                          * OK, because we always copy the OID if present (see below).
7655                          * Other system columns may not.
7656                          */
7657                         if (attno == ObjectIdAttributeNumber)
7658                                 continue;
7659                         elog(ERROR, "system column in index");
7660                 }
7661                 nulls[attno - 1] = false;
7662         }
7663
7664         key_tuple = heap_form_tuple(desc, values, nulls);
7665         *copy = true;
7666         RelationClose(idx_rel);
7667
7668         /*
7669          * Always copy oids if the table has them, even if not included in the
7670          * index. The space in the logged tuple is used anyway, so there's little
7671          * point in not including the information.
7672          */
7673         if (relation->rd_rel->relhasoids)
7674                 HeapTupleSetOid(key_tuple, HeapTupleGetOid(tp));
7675
7676         /*
7677          * If the tuple, which by here only contains indexed columns, still has
7678          * toasted columns, force them to be inlined. This is somewhat unlikely
7679          * since there's limits on the size of indexed columns, so we don't
7680          * duplicate toast_flatten_tuple()s functionality in the above loop over
7681          * the indexed columns, even if it would be more efficient.
7682          */
7683         if (HeapTupleHasExternal(key_tuple))
7684         {
7685                 HeapTuple       oldtup = key_tuple;
7686
7687                 key_tuple = toast_flatten_tuple(oldtup, RelationGetDescr(relation));
7688                 heap_freetuple(oldtup);
7689         }
7690
7691         return key_tuple;
7692 }
7693
7694 /*
7695  * Handles CLEANUP_INFO
7696  */
7697 static void
7698 heap_xlog_cleanup_info(XLogReaderState *record)
7699 {
7700         xl_heap_cleanup_info *xlrec = (xl_heap_cleanup_info *) XLogRecGetData(record);
7701
7702         if (InHotStandby)
7703                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, xlrec->node);
7704
7705         /*
7706          * Actual operation is a no-op. Record type exists to provide a means for
7707          * conflict processing to occur before we begin index vacuum actions. see
7708          * vacuumlazy.c and also comments in btvacuumpage()
7709          */
7710
7711         /* Backup blocks are not used in cleanup_info records */
7712         Assert(!XLogRecHasAnyBlockRefs(record));
7713 }
7714
7715 /*
7716  * Handles HEAP2_CLEAN record type
7717  */
7718 static void
7719 heap_xlog_clean(XLogReaderState *record)
7720 {
7721         XLogRecPtr      lsn = record->EndRecPtr;
7722         xl_heap_clean *xlrec = (xl_heap_clean *) XLogRecGetData(record);
7723         Buffer          buffer;
7724         Size            freespace = 0;
7725         RelFileNode rnode;
7726         BlockNumber blkno;
7727         XLogRedoAction action;
7728
7729         XLogRecGetBlockTag(record, 0, &rnode, NULL, &blkno);
7730
7731         /*
7732          * We're about to remove tuples. In Hot Standby mode, ensure that there's
7733          * no queries running for which the removed tuples are still visible.
7734          *
7735          * Not all HEAP2_CLEAN records remove tuples with xids, so we only want to
7736          * conflict on the records that cause MVCC failures for user queries. If
7737          * latestRemovedXid is invalid, skip conflict processing.
7738          */
7739         if (InHotStandby && TransactionIdIsValid(xlrec->latestRemovedXid))
7740                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, rnode);
7741
7742         /*
7743          * If we have a full-page image, restore it (using a cleanup lock) and
7744          * we're done.
7745          */
7746         action = XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true,
7747                                                                                    &buffer);
7748         if (action == BLK_NEEDS_REDO)
7749         {
7750                 Page            page = (Page) BufferGetPage(buffer);
7751                 OffsetNumber *end;
7752                 OffsetNumber *redirected;
7753                 OffsetNumber *nowdead;
7754                 OffsetNumber *nowunused;
7755                 int                     nredirected;
7756                 int                     ndead;
7757                 int                     nunused;
7758                 Size            datalen;
7759
7760                 redirected = (OffsetNumber *) XLogRecGetBlockData(record, 0, &datalen);
7761
7762                 nredirected = xlrec->nredirected;
7763                 ndead = xlrec->ndead;
7764                 end = (OffsetNumber *) ((char *) redirected + datalen);
7765                 nowdead = redirected + (nredirected * 2);
7766                 nowunused = nowdead + ndead;
7767                 nunused = (end - nowunused);
7768                 Assert(nunused >= 0);
7769
7770                 /* Update all item pointers per the record, and repair fragmentation */
7771                 heap_page_prune_execute(buffer,
7772                                                                 redirected, nredirected,
7773                                                                 nowdead, ndead,
7774                                                                 nowunused, nunused);
7775
7776                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
7777
7778                 /*
7779                  * Note: we don't worry about updating the page's prunability hints.
7780                  * At worst this will cause an extra prune cycle to occur soon.
7781                  */
7782
7783                 PageSetLSN(page, lsn);
7784                 MarkBufferDirty(buffer);
7785         }
7786         if (BufferIsValid(buffer))
7787                 UnlockReleaseBuffer(buffer);
7788
7789         /*
7790          * Update the FSM as well.
7791          *
7792          * XXX: Don't do this if the page was restored from full page image. We
7793          * don't bother to update the FSM in that case, it doesn't need to be
7794          * totally accurate anyway.
7795          */
7796         if (action == BLK_NEEDS_REDO)
7797                 XLogRecordPageWithFreeSpace(rnode, blkno, freespace);
7798 }
7799
7800 /*
7801  * Replay XLOG_HEAP2_VISIBLE record.
7802  *
7803  * The critical integrity requirement here is that we must never end up with
7804  * a situation where the visibility map bit is set, and the page-level
7805  * PD_ALL_VISIBLE bit is clear.  If that were to occur, then a subsequent
7806  * page modification would fail to clear the visibility map bit.
7807  */
7808 static void
7809 heap_xlog_visible(XLogReaderState *record)
7810 {
7811         XLogRecPtr      lsn = record->EndRecPtr;
7812         xl_heap_visible *xlrec = (xl_heap_visible *) XLogRecGetData(record);
7813         Buffer          vmbuffer = InvalidBuffer;
7814         Buffer          buffer;
7815         Page            page;
7816         RelFileNode rnode;
7817         BlockNumber blkno;
7818         XLogRedoAction action;
7819
7820         XLogRecGetBlockTag(record, 1, &rnode, NULL, &blkno);
7821
7822         /*
7823          * If there are any Hot Standby transactions running that have an xmin
7824          * horizon old enough that this page isn't all-visible for them, they
7825          * might incorrectly decide that an index-only scan can skip a heap fetch.
7826          *
7827          * NB: It might be better to throw some kind of "soft" conflict here that
7828          * forces any index-only scan that is in flight to perform heap fetches,
7829          * rather than killing the transaction outright.
7830          */
7831         if (InHotStandby)
7832                 ResolveRecoveryConflictWithSnapshot(xlrec->cutoff_xid, rnode);
7833
7834         /*
7835          * Read the heap page, if it still exists. If the heap file has dropped or
7836          * truncated later in recovery, we don't need to update the page, but we'd
7837          * better still update the visibility map.
7838          */
7839         action = XLogReadBufferForRedo(record, 1, &buffer);
7840         if (action == BLK_NEEDS_REDO)
7841         {
7842                 /*
7843                  * We don't bump the LSN of the heap page when setting the visibility
7844                  * map bit (unless checksums or wal_hint_bits is enabled, in which
7845                  * case we must), because that would generate an unworkable volume of
7846                  * full-page writes.  This exposes us to torn page hazards, but since
7847                  * we're not inspecting the existing page contents in any way, we
7848                  * don't care.
7849                  *
7850                  * However, all operations that clear the visibility map bit *do* bump
7851                  * the LSN, and those operations will only be replayed if the XLOG LSN
7852                  * follows the page LSN.  Thus, if the page LSN has advanced past our
7853                  * XLOG record's LSN, we mustn't mark the page all-visible, because
7854                  * the subsequent update won't be replayed to clear the flag.
7855                  */
7856                 page = BufferGetPage(buffer);
7857
7858                 if (xlrec->flags & VISIBILITYMAP_ALL_VISIBLE)
7859                         PageSetAllVisible(page);
7860                 if (xlrec->flags & VISIBILITYMAP_ALL_FROZEN)
7861                         PageSetAllFrozen(page);
7862
7863                 MarkBufferDirty(buffer);
7864         }
7865         else if (action == BLK_RESTORED)
7866         {
7867                 /*
7868                  * If heap block was backed up, we already restored it and there's
7869                  * nothing more to do. (This can only happen with checksums or
7870                  * wal_log_hints enabled.)
7871                  */
7872         }
7873         if (BufferIsValid(buffer))
7874                 UnlockReleaseBuffer(buffer);
7875
7876         /*
7877          * Even if we skipped the heap page update due to the LSN interlock, it's
7878          * still safe to update the visibility map.  Any WAL record that clears
7879          * the visibility map bit does so before checking the page LSN, so any
7880          * bits that need to be cleared will still be cleared.
7881          */
7882         if (XLogReadBufferForRedoExtended(record, 0, RBM_ZERO_ON_ERROR, false,
7883                                                                           &vmbuffer) == BLK_NEEDS_REDO)
7884         {
7885                 Page            vmpage = BufferGetPage(vmbuffer);
7886                 Relation        reln;
7887
7888                 /* initialize the page if it was read as zeros */
7889                 if (PageIsNew(vmpage))
7890                         PageInit(vmpage, BLCKSZ, 0);
7891
7892                 /*
7893                  * XLogReplayBufferExtended locked the buffer. But visibilitymap_set
7894                  * will handle locking itself.
7895                  */
7896                 LockBuffer(vmbuffer, BUFFER_LOCK_UNLOCK);
7897
7898                 reln = CreateFakeRelcacheEntry(rnode);
7899                 visibilitymap_pin(reln, blkno, &vmbuffer);
7900
7901                 /*
7902                  * Don't set the bit if replay has already passed this point.
7903                  *
7904                  * It might be safe to do this unconditionally; if replay has passed
7905                  * this point, we'll replay at least as far this time as we did
7906                  * before, and if this bit needs to be cleared, the record responsible
7907                  * for doing so should be again replayed, and clear it.  For right
7908                  * now, out of an abundance of conservatism, we use the same test here
7909                  * we did for the heap page.  If this results in a dropped bit, no
7910                  * real harm is done; and the next VACUUM will fix it.
7911                  */
7912                 if (lsn > PageGetLSN(vmpage))
7913                         visibilitymap_set(reln, blkno, InvalidBuffer, lsn, vmbuffer,
7914                                                           xlrec->cutoff_xid, xlrec->flags);
7915
7916                 ReleaseBuffer(vmbuffer);
7917                 FreeFakeRelcacheEntry(reln);
7918         }
7919         else if (BufferIsValid(vmbuffer))
7920                 UnlockReleaseBuffer(vmbuffer);
7921 }
7922
7923 /*
7924  * Replay XLOG_HEAP2_FREEZE_PAGE records
7925  */
7926 static void
7927 heap_xlog_freeze_page(XLogReaderState *record)
7928 {
7929         XLogRecPtr      lsn = record->EndRecPtr;
7930         xl_heap_freeze_page *xlrec = (xl_heap_freeze_page *) XLogRecGetData(record);
7931         TransactionId cutoff_xid = xlrec->cutoff_xid;
7932         Buffer          buffer;
7933         int                     ntup;
7934
7935         /*
7936          * In Hot Standby mode, ensure that there's no queries running which still
7937          * consider the frozen xids as running.
7938          */
7939         if (InHotStandby)
7940         {
7941                 RelFileNode rnode;
7942                 TransactionId latestRemovedXid = cutoff_xid;
7943
7944                 TransactionIdRetreat(latestRemovedXid);
7945
7946                 XLogRecGetBlockTag(record, 0, &rnode, NULL, NULL);
7947                 ResolveRecoveryConflictWithSnapshot(latestRemovedXid, rnode);
7948         }
7949
7950         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
7951         {
7952                 Page            page = BufferGetPage(buffer);
7953                 xl_heap_freeze_tuple *tuples;
7954
7955                 tuples = (xl_heap_freeze_tuple *) XLogRecGetBlockData(record, 0, NULL);
7956
7957                 /* now execute freeze plan for each frozen tuple */
7958                 for (ntup = 0; ntup < xlrec->ntuples; ntup++)
7959                 {
7960                         xl_heap_freeze_tuple *xlrec_tp;
7961                         ItemId          lp;
7962                         HeapTupleHeader tuple;
7963
7964                         xlrec_tp = &tuples[ntup];
7965                         lp = PageGetItemId(page, xlrec_tp->offset); /* offsets are one-based */
7966                         tuple = (HeapTupleHeader) PageGetItem(page, lp);
7967
7968                         heap_execute_freeze_tuple(tuple, xlrec_tp);
7969                 }
7970
7971                 PageSetLSN(page, lsn);
7972                 MarkBufferDirty(buffer);
7973         }
7974         if (BufferIsValid(buffer))
7975                 UnlockReleaseBuffer(buffer);
7976 }
7977
7978 /*
7979  * Given an "infobits" field from an XLog record, set the correct bits in the
7980  * given infomask and infomask2 for the tuple touched by the record.
7981  *
7982  * (This is the reverse of compute_infobits).
7983  */
7984 static void
7985 fix_infomask_from_infobits(uint8 infobits, uint16 *infomask, uint16 *infomask2)
7986 {
7987         *infomask &= ~(HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY |
7988                                    HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_EXCL_LOCK);
7989         *infomask2 &= ~HEAP_KEYS_UPDATED;
7990
7991         if (infobits & XLHL_XMAX_IS_MULTI)
7992                 *infomask |= HEAP_XMAX_IS_MULTI;
7993         if (infobits & XLHL_XMAX_LOCK_ONLY)
7994                 *infomask |= HEAP_XMAX_LOCK_ONLY;
7995         if (infobits & XLHL_XMAX_EXCL_LOCK)
7996                 *infomask |= HEAP_XMAX_EXCL_LOCK;
7997         /* note HEAP_XMAX_SHR_LOCK isn't considered here */
7998         if (infobits & XLHL_XMAX_KEYSHR_LOCK)
7999                 *infomask |= HEAP_XMAX_KEYSHR_LOCK;
8000
8001         if (infobits & XLHL_KEYS_UPDATED)
8002                 *infomask2 |= HEAP_KEYS_UPDATED;
8003 }
8004
8005 static void
8006 heap_xlog_delete(XLogReaderState *record)
8007 {
8008         XLogRecPtr      lsn = record->EndRecPtr;
8009         xl_heap_delete *xlrec = (xl_heap_delete *) XLogRecGetData(record);
8010         Buffer          buffer;
8011         Page            page;
8012         ItemId          lp = NULL;
8013         HeapTupleHeader htup;
8014         BlockNumber blkno;
8015         RelFileNode target_node;
8016         ItemPointerData target_tid;
8017
8018         XLogRecGetBlockTag(record, 0, &target_node, NULL, &blkno);
8019         ItemPointerSetBlockNumber(&target_tid, blkno);
8020         ItemPointerSetOffsetNumber(&target_tid, xlrec->offnum);
8021
8022         /*
8023          * The visibility map may need to be fixed even if the heap page is
8024          * already up-to-date.
8025          */
8026         if (xlrec->flags & XLH_DELETE_ALL_VISIBLE_CLEARED)
8027         {
8028                 Relation        reln = CreateFakeRelcacheEntry(target_node);
8029                 Buffer          vmbuffer = InvalidBuffer;
8030
8031                 visibilitymap_pin(reln, blkno, &vmbuffer);
8032                 visibilitymap_clear(reln, blkno, vmbuffer);
8033                 ReleaseBuffer(vmbuffer);
8034                 FreeFakeRelcacheEntry(reln);
8035         }
8036
8037         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8038         {
8039                 page = BufferGetPage(buffer);
8040
8041                 if (PageGetMaxOffsetNumber(page) >= xlrec->offnum)
8042                         lp = PageGetItemId(page, xlrec->offnum);
8043
8044                 if (PageGetMaxOffsetNumber(page) < xlrec->offnum || !ItemIdIsNormal(lp))
8045                         elog(PANIC, "invalid lp");
8046
8047                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8048
8049                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
8050                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
8051                 HeapTupleHeaderClearHotUpdated(htup);
8052                 fix_infomask_from_infobits(xlrec->infobits_set,
8053                                                                    &htup->t_infomask, &htup->t_infomask2);
8054                 if (!(xlrec->flags & XLH_DELETE_IS_SUPER))
8055                         HeapTupleHeaderSetXmax(htup, xlrec->xmax);
8056                 else
8057                         HeapTupleHeaderSetXmin(htup, InvalidTransactionId);
8058                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
8059
8060                 /* Mark the page as a candidate for pruning */
8061                 PageSetPrunable(page, XLogRecGetXid(record));
8062
8063                 if (xlrec->flags & XLH_DELETE_ALL_VISIBLE_CLEARED)
8064                         PageClearAllVisible(page);
8065
8066                 /* Make sure there is no forward chain link in t_ctid */
8067                 htup->t_ctid = target_tid;
8068                 PageSetLSN(page, lsn);
8069                 MarkBufferDirty(buffer);
8070         }
8071         if (BufferIsValid(buffer))
8072                 UnlockReleaseBuffer(buffer);
8073 }
8074
8075 static void
8076 heap_xlog_insert(XLogReaderState *record)
8077 {
8078         XLogRecPtr      lsn = record->EndRecPtr;
8079         xl_heap_insert *xlrec = (xl_heap_insert *) XLogRecGetData(record);
8080         Buffer          buffer;
8081         Page            page;
8082         union
8083         {
8084                 HeapTupleHeaderData hdr;
8085                 char            data[MaxHeapTupleSize];
8086         }                       tbuf;
8087         HeapTupleHeader htup;
8088         xl_heap_header xlhdr;
8089         uint32          newlen;
8090         Size            freespace = 0;
8091         RelFileNode target_node;
8092         BlockNumber blkno;
8093         ItemPointerData target_tid;
8094         XLogRedoAction action;
8095
8096         XLogRecGetBlockTag(record, 0, &target_node, NULL, &blkno);
8097         ItemPointerSetBlockNumber(&target_tid, blkno);
8098         ItemPointerSetOffsetNumber(&target_tid, xlrec->offnum);
8099
8100         /*
8101          * The visibility map may need to be fixed even if the heap page is
8102          * already up-to-date.
8103          */
8104         if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8105         {
8106                 Relation        reln = CreateFakeRelcacheEntry(target_node);
8107                 Buffer          vmbuffer = InvalidBuffer;
8108
8109                 visibilitymap_pin(reln, blkno, &vmbuffer);
8110                 visibilitymap_clear(reln, blkno, vmbuffer);
8111                 ReleaseBuffer(vmbuffer);
8112                 FreeFakeRelcacheEntry(reln);
8113         }
8114
8115         /*
8116          * If we inserted the first and only tuple on the page, re-initialize the
8117          * page from scratch.
8118          */
8119         if (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE)
8120         {
8121                 buffer = XLogInitBufferForRedo(record, 0);
8122                 page = BufferGetPage(buffer);
8123                 PageInit(page, BufferGetPageSize(buffer), 0);
8124                 action = BLK_NEEDS_REDO;
8125         }
8126         else
8127                 action = XLogReadBufferForRedo(record, 0, &buffer);
8128         if (action == BLK_NEEDS_REDO)
8129         {
8130                 Size            datalen;
8131                 char       *data;
8132
8133                 page = BufferGetPage(buffer);
8134
8135                 if (PageGetMaxOffsetNumber(page) + 1 < xlrec->offnum)
8136                         elog(PANIC, "invalid max offset number");
8137
8138                 data = XLogRecGetBlockData(record, 0, &datalen);
8139
8140                 newlen = datalen - SizeOfHeapHeader;
8141                 Assert(datalen > SizeOfHeapHeader && newlen <= MaxHeapTupleSize);
8142                 memcpy((char *) &xlhdr, data, SizeOfHeapHeader);
8143                 data += SizeOfHeapHeader;
8144
8145                 htup = &tbuf.hdr;
8146                 MemSet((char *) htup, 0, SizeofHeapTupleHeader);
8147                 /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
8148                 memcpy((char *) htup + SizeofHeapTupleHeader,
8149                            data,
8150                            newlen);
8151                 newlen += SizeofHeapTupleHeader;
8152                 htup->t_infomask2 = xlhdr.t_infomask2;
8153                 htup->t_infomask = xlhdr.t_infomask;
8154                 htup->t_hoff = xlhdr.t_hoff;
8155                 HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
8156                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
8157                 htup->t_ctid = target_tid;
8158
8159                 if (PageAddItem(page, (Item) htup, newlen, xlrec->offnum,
8160                                                 true, true) == InvalidOffsetNumber)
8161                         elog(PANIC, "failed to add tuple");
8162
8163                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
8164
8165                 PageSetLSN(page, lsn);
8166
8167                 if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8168                         PageClearAllVisible(page);
8169
8170                 MarkBufferDirty(buffer);
8171         }
8172         if (BufferIsValid(buffer))
8173                 UnlockReleaseBuffer(buffer);
8174
8175         /*
8176          * If the page is running low on free space, update the FSM as well.
8177          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
8178          * better than that without knowing the fill-factor for the table.
8179          *
8180          * XXX: Don't do this if the page was restored from full page image. We
8181          * don't bother to update the FSM in that case, it doesn't need to be
8182          * totally accurate anyway.
8183          */
8184         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
8185                 XLogRecordPageWithFreeSpace(target_node, blkno, freespace);
8186 }
8187
8188 /*
8189  * Handles MULTI_INSERT record type.
8190  */
8191 static void
8192 heap_xlog_multi_insert(XLogReaderState *record)
8193 {
8194         XLogRecPtr      lsn = record->EndRecPtr;
8195         xl_heap_multi_insert *xlrec;
8196         RelFileNode rnode;
8197         BlockNumber blkno;
8198         Buffer          buffer;
8199         Page            page;
8200         union
8201         {
8202                 HeapTupleHeaderData hdr;
8203                 char            data[MaxHeapTupleSize];
8204         }                       tbuf;
8205         HeapTupleHeader htup;
8206         uint32          newlen;
8207         Size            freespace = 0;
8208         int                     i;
8209         bool            isinit = (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE) != 0;
8210         XLogRedoAction action;
8211
8212         /*
8213          * Insertion doesn't overwrite MVCC data, so no conflict processing is
8214          * required.
8215          */
8216         xlrec = (xl_heap_multi_insert *) XLogRecGetData(record);
8217
8218         XLogRecGetBlockTag(record, 0, &rnode, NULL, &blkno);
8219
8220         /*
8221          * The visibility map may need to be fixed even if the heap page is
8222          * already up-to-date.
8223          */
8224         if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8225         {
8226                 Relation        reln = CreateFakeRelcacheEntry(rnode);
8227                 Buffer          vmbuffer = InvalidBuffer;
8228
8229                 visibilitymap_pin(reln, blkno, &vmbuffer);
8230                 visibilitymap_clear(reln, blkno, vmbuffer);
8231                 ReleaseBuffer(vmbuffer);
8232                 FreeFakeRelcacheEntry(reln);
8233         }
8234
8235         if (isinit)
8236         {
8237                 buffer = XLogInitBufferForRedo(record, 0);
8238                 page = BufferGetPage(buffer);
8239                 PageInit(page, BufferGetPageSize(buffer), 0);
8240                 action = BLK_NEEDS_REDO;
8241         }
8242         else
8243                 action = XLogReadBufferForRedo(record, 0, &buffer);
8244         if (action == BLK_NEEDS_REDO)
8245         {
8246                 char       *tupdata;
8247                 char       *endptr;
8248                 Size            len;
8249
8250                 /* Tuples are stored as block data */
8251                 tupdata = XLogRecGetBlockData(record, 0, &len);
8252                 endptr = tupdata + len;
8253
8254                 page = (Page) BufferGetPage(buffer);
8255
8256                 for (i = 0; i < xlrec->ntuples; i++)
8257                 {
8258                         OffsetNumber offnum;
8259                         xl_multi_insert_tuple *xlhdr;
8260
8261                         /*
8262                          * If we're reinitializing the page, the tuples are stored in
8263                          * order from FirstOffsetNumber. Otherwise there's an array of
8264                          * offsets in the WAL record, and the tuples come after that.
8265                          */
8266                         if (isinit)
8267                                 offnum = FirstOffsetNumber + i;
8268                         else
8269                                 offnum = xlrec->offsets[i];
8270                         if (PageGetMaxOffsetNumber(page) + 1 < offnum)
8271                                 elog(PANIC, "invalid max offset number");
8272
8273                         xlhdr = (xl_multi_insert_tuple *) SHORTALIGN(tupdata);
8274                         tupdata = ((char *) xlhdr) + SizeOfMultiInsertTuple;
8275
8276                         newlen = xlhdr->datalen;
8277                         Assert(newlen <= MaxHeapTupleSize);
8278                         htup = &tbuf.hdr;
8279                         MemSet((char *) htup, 0, SizeofHeapTupleHeader);
8280                         /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
8281                         memcpy((char *) htup + SizeofHeapTupleHeader,
8282                                    (char *) tupdata,
8283                                    newlen);
8284                         tupdata += newlen;
8285
8286                         newlen += SizeofHeapTupleHeader;
8287                         htup->t_infomask2 = xlhdr->t_infomask2;
8288                         htup->t_infomask = xlhdr->t_infomask;
8289                         htup->t_hoff = xlhdr->t_hoff;
8290                         HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
8291                         HeapTupleHeaderSetCmin(htup, FirstCommandId);
8292                         ItemPointerSetBlockNumber(&htup->t_ctid, blkno);
8293                         ItemPointerSetOffsetNumber(&htup->t_ctid, offnum);
8294
8295                         offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
8296                         if (offnum == InvalidOffsetNumber)
8297                                 elog(PANIC, "failed to add tuple");
8298                 }
8299                 if (tupdata != endptr)
8300                         elog(PANIC, "total tuple length mismatch");
8301
8302                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
8303
8304                 PageSetLSN(page, lsn);
8305
8306                 if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8307                         PageClearAllVisible(page);
8308
8309                 MarkBufferDirty(buffer);
8310         }
8311         if (BufferIsValid(buffer))
8312                 UnlockReleaseBuffer(buffer);
8313
8314         /*
8315          * If the page is running low on free space, update the FSM as well.
8316          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
8317          * better than that without knowing the fill-factor for the table.
8318          *
8319          * XXX: Don't do this if the page was restored from full page image. We
8320          * don't bother to update the FSM in that case, it doesn't need to be
8321          * totally accurate anyway.
8322          */
8323         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
8324                 XLogRecordPageWithFreeSpace(rnode, blkno, freespace);
8325 }
8326
8327 /*
8328  * Handles UPDATE and HOT_UPDATE
8329  */
8330 static void
8331 heap_xlog_update(XLogReaderState *record, bool hot_update)
8332 {
8333         XLogRecPtr      lsn = record->EndRecPtr;
8334         xl_heap_update *xlrec = (xl_heap_update *) XLogRecGetData(record);
8335         RelFileNode rnode;
8336         BlockNumber oldblk;
8337         BlockNumber newblk;
8338         ItemPointerData newtid;
8339         Buffer          obuffer,
8340                                 nbuffer;
8341         Page            page;
8342         OffsetNumber offnum;
8343         ItemId          lp = NULL;
8344         HeapTupleData oldtup;
8345         HeapTupleHeader htup;
8346         uint16          prefixlen = 0,
8347                                 suffixlen = 0;
8348         char       *newp;
8349         union
8350         {
8351                 HeapTupleHeaderData hdr;
8352                 char            data[MaxHeapTupleSize];
8353         }                       tbuf;
8354         xl_heap_header xlhdr;
8355         uint32          newlen;
8356         Size            freespace = 0;
8357         XLogRedoAction oldaction;
8358         XLogRedoAction newaction;
8359
8360         /* initialize to keep the compiler quiet */
8361         oldtup.t_data = NULL;
8362         oldtup.t_len = 0;
8363
8364         XLogRecGetBlockTag(record, 0, &rnode, NULL, &newblk);
8365         if (XLogRecGetBlockTag(record, 1, NULL, NULL, &oldblk))
8366         {
8367                 /* HOT updates are never done across pages */
8368                 Assert(!hot_update);
8369         }
8370         else
8371                 oldblk = newblk;
8372
8373         ItemPointerSet(&newtid, newblk, xlrec->new_offnum);
8374
8375         /*
8376          * The visibility map may need to be fixed even if the heap page is
8377          * already up-to-date.
8378          */
8379         if (xlrec->flags & XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED)
8380         {
8381                 Relation        reln = CreateFakeRelcacheEntry(rnode);
8382                 Buffer          vmbuffer = InvalidBuffer;
8383
8384                 visibilitymap_pin(reln, oldblk, &vmbuffer);
8385                 visibilitymap_clear(reln, oldblk, vmbuffer);
8386                 ReleaseBuffer(vmbuffer);
8387                 FreeFakeRelcacheEntry(reln);
8388         }
8389
8390         /*
8391          * In normal operation, it is important to lock the two pages in
8392          * page-number order, to avoid possible deadlocks against other update
8393          * operations going the other way.  However, during WAL replay there can
8394          * be no other update happening, so we don't need to worry about that. But
8395          * we *do* need to worry that we don't expose an inconsistent state to Hot
8396          * Standby queries --- so the original page can't be unlocked before we've
8397          * added the new tuple to the new page.
8398          */
8399
8400         /* Deal with old tuple version */
8401         oldaction = XLogReadBufferForRedo(record, (oldblk == newblk) ? 0 : 1,
8402                                                                           &obuffer);
8403         if (oldaction == BLK_NEEDS_REDO)
8404         {
8405                 page = BufferGetPage(obuffer);
8406                 offnum = xlrec->old_offnum;
8407                 if (PageGetMaxOffsetNumber(page) >= offnum)
8408                         lp = PageGetItemId(page, offnum);
8409
8410                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8411                         elog(PANIC, "invalid lp");
8412
8413                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8414
8415                 oldtup.t_data = htup;
8416                 oldtup.t_len = ItemIdGetLength(lp);
8417
8418                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
8419                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
8420                 if (hot_update)
8421                         HeapTupleHeaderSetHotUpdated(htup);
8422                 else
8423                         HeapTupleHeaderClearHotUpdated(htup);
8424                 fix_infomask_from_infobits(xlrec->old_infobits_set, &htup->t_infomask,
8425                                                                    &htup->t_infomask2);
8426                 HeapTupleHeaderSetXmax(htup, xlrec->old_xmax);
8427                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
8428                 /* Set forward chain link in t_ctid */
8429                 htup->t_ctid = newtid;
8430
8431                 /* Mark the page as a candidate for pruning */
8432                 PageSetPrunable(page, XLogRecGetXid(record));
8433
8434                 if (xlrec->flags & XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED)
8435                         PageClearAllVisible(page);
8436
8437                 PageSetLSN(page, lsn);
8438                 MarkBufferDirty(obuffer);
8439         }
8440
8441         /*
8442          * Read the page the new tuple goes into, if different from old.
8443          */
8444         if (oldblk == newblk)
8445         {
8446                 nbuffer = obuffer;
8447                 newaction = oldaction;
8448         }
8449         else if (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE)
8450         {
8451                 nbuffer = XLogInitBufferForRedo(record, 0);
8452                 page = (Page) BufferGetPage(nbuffer);
8453                 PageInit(page, BufferGetPageSize(nbuffer), 0);
8454                 newaction = BLK_NEEDS_REDO;
8455         }
8456         else
8457                 newaction = XLogReadBufferForRedo(record, 0, &nbuffer);
8458
8459         /*
8460          * The visibility map may need to be fixed even if the heap page is
8461          * already up-to-date.
8462          */
8463         if (xlrec->flags & XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED)
8464         {
8465                 Relation        reln = CreateFakeRelcacheEntry(rnode);
8466                 Buffer          vmbuffer = InvalidBuffer;
8467
8468                 visibilitymap_pin(reln, newblk, &vmbuffer);
8469                 visibilitymap_clear(reln, newblk, vmbuffer);
8470                 ReleaseBuffer(vmbuffer);
8471                 FreeFakeRelcacheEntry(reln);
8472         }
8473
8474         /* Deal with new tuple */
8475         if (newaction == BLK_NEEDS_REDO)
8476         {
8477                 char       *recdata;
8478                 char       *recdata_end;
8479                 Size            datalen;
8480                 Size            tuplen;
8481
8482                 recdata = XLogRecGetBlockData(record, 0, &datalen);
8483                 recdata_end = recdata + datalen;
8484
8485                 page = BufferGetPage(nbuffer);
8486
8487                 offnum = xlrec->new_offnum;
8488                 if (PageGetMaxOffsetNumber(page) + 1 < offnum)
8489                         elog(PANIC, "invalid max offset number");
8490
8491                 if (xlrec->flags & XLH_UPDATE_PREFIX_FROM_OLD)
8492                 {
8493                         Assert(newblk == oldblk);
8494                         memcpy(&prefixlen, recdata, sizeof(uint16));
8495                         recdata += sizeof(uint16);
8496                 }
8497                 if (xlrec->flags & XLH_UPDATE_SUFFIX_FROM_OLD)
8498                 {
8499                         Assert(newblk == oldblk);
8500                         memcpy(&suffixlen, recdata, sizeof(uint16));
8501                         recdata += sizeof(uint16);
8502                 }
8503
8504                 memcpy((char *) &xlhdr, recdata, SizeOfHeapHeader);
8505                 recdata += SizeOfHeapHeader;
8506
8507                 tuplen = recdata_end - recdata;
8508                 Assert(tuplen <= MaxHeapTupleSize);
8509
8510                 htup = &tbuf.hdr;
8511                 MemSet((char *) htup, 0, SizeofHeapTupleHeader);
8512
8513                 /*
8514                  * Reconstruct the new tuple using the prefix and/or suffix from the
8515                  * old tuple, and the data stored in the WAL record.
8516                  */
8517                 newp = (char *) htup + SizeofHeapTupleHeader;
8518                 if (prefixlen > 0)
8519                 {
8520                         int                     len;
8521
8522                         /* copy bitmap [+ padding] [+ oid] from WAL record */
8523                         len = xlhdr.t_hoff - SizeofHeapTupleHeader;
8524                         memcpy(newp, recdata, len);
8525                         recdata += len;
8526                         newp += len;
8527
8528                         /* copy prefix from old tuple */
8529                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_data->t_hoff, prefixlen);
8530                         newp += prefixlen;
8531
8532                         /* copy new tuple data from WAL record */
8533                         len = tuplen - (xlhdr.t_hoff - SizeofHeapTupleHeader);
8534                         memcpy(newp, recdata, len);
8535                         recdata += len;
8536                         newp += len;
8537                 }
8538                 else
8539                 {
8540                         /*
8541                          * copy bitmap [+ padding] [+ oid] + data from record, all in one
8542                          * go
8543                          */
8544                         memcpy(newp, recdata, tuplen);
8545                         recdata += tuplen;
8546                         newp += tuplen;
8547                 }
8548                 Assert(recdata == recdata_end);
8549
8550                 /* copy suffix from old tuple */
8551                 if (suffixlen > 0)
8552                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_len - suffixlen, suffixlen);
8553
8554                 newlen = SizeofHeapTupleHeader + tuplen + prefixlen + suffixlen;
8555                 htup->t_infomask2 = xlhdr.t_infomask2;
8556                 htup->t_infomask = xlhdr.t_infomask;
8557                 htup->t_hoff = xlhdr.t_hoff;
8558
8559                 HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
8560                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
8561                 HeapTupleHeaderSetXmax(htup, xlrec->new_xmax);
8562                 /* Make sure there is no forward chain link in t_ctid */
8563                 htup->t_ctid = newtid;
8564
8565                 offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
8566                 if (offnum == InvalidOffsetNumber)
8567                         elog(PANIC, "failed to add tuple");
8568
8569                 if (xlrec->flags & XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED)
8570                         PageClearAllVisible(page);
8571
8572                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
8573
8574                 PageSetLSN(page, lsn);
8575                 MarkBufferDirty(nbuffer);
8576         }
8577
8578         if (BufferIsValid(nbuffer) && nbuffer != obuffer)
8579                 UnlockReleaseBuffer(nbuffer);
8580         if (BufferIsValid(obuffer))
8581                 UnlockReleaseBuffer(obuffer);
8582
8583         /*
8584          * If the new page is running low on free space, update the FSM as well.
8585          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
8586          * better than that without knowing the fill-factor for the table.
8587          *
8588          * However, don't update the FSM on HOT updates, because after crash
8589          * recovery, either the old or the new tuple will certainly be dead and
8590          * prunable. After pruning, the page will have roughly as much free space
8591          * as it did before the update, assuming the new tuple is about the same
8592          * size as the old one.
8593          *
8594          * XXX: Don't do this if the page was restored from full page image. We
8595          * don't bother to update the FSM in that case, it doesn't need to be
8596          * totally accurate anyway.
8597          */
8598         if (newaction == BLK_NEEDS_REDO && !hot_update && freespace < BLCKSZ / 5)
8599                 XLogRecordPageWithFreeSpace(rnode, newblk, freespace);
8600 }
8601
8602 static void
8603 heap_xlog_confirm(XLogReaderState *record)
8604 {
8605         XLogRecPtr      lsn = record->EndRecPtr;
8606         xl_heap_confirm *xlrec = (xl_heap_confirm *) XLogRecGetData(record);
8607         Buffer          buffer;
8608         Page            page;
8609         OffsetNumber offnum;
8610         ItemId          lp = NULL;
8611         HeapTupleHeader htup;
8612
8613         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8614         {
8615                 page = BufferGetPage(buffer);
8616
8617                 offnum = xlrec->offnum;
8618                 if (PageGetMaxOffsetNumber(page) >= offnum)
8619                         lp = PageGetItemId(page, offnum);
8620
8621                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8622                         elog(PANIC, "invalid lp");
8623
8624                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8625
8626                 /*
8627                  * Confirm tuple as actually inserted
8628                  */
8629                 ItemPointerSet(&htup->t_ctid, BufferGetBlockNumber(buffer), offnum);
8630
8631                 PageSetLSN(page, lsn);
8632                 MarkBufferDirty(buffer);
8633         }
8634         if (BufferIsValid(buffer))
8635                 UnlockReleaseBuffer(buffer);
8636 }
8637
8638 static void
8639 heap_xlog_lock(XLogReaderState *record)
8640 {
8641         XLogRecPtr      lsn = record->EndRecPtr;
8642         xl_heap_lock *xlrec = (xl_heap_lock *) XLogRecGetData(record);
8643         Buffer          buffer;
8644         Page            page;
8645         OffsetNumber offnum;
8646         ItemId          lp = NULL;
8647         HeapTupleHeader htup;
8648
8649         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8650         {
8651                 page = (Page) BufferGetPage(buffer);
8652
8653                 offnum = xlrec->offnum;
8654                 if (PageGetMaxOffsetNumber(page) >= offnum)
8655                         lp = PageGetItemId(page, offnum);
8656
8657                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8658                         elog(PANIC, "invalid lp");
8659
8660                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8661
8662                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
8663                                                                    &htup->t_infomask2);
8664
8665                 /*
8666                  * Clear relevant update flags, but only if the modified infomask says
8667                  * there's no update.
8668                  */
8669                 if (HEAP_XMAX_IS_LOCKED_ONLY(htup->t_infomask))
8670                 {
8671                         HeapTupleHeaderClearHotUpdated(htup);
8672                         /* Make sure there is no forward chain link in t_ctid */
8673                         ItemPointerSet(&htup->t_ctid,
8674                                                    BufferGetBlockNumber(buffer),
8675                                                    offnum);
8676                 }
8677                 HeapTupleHeaderSetXmax(htup, xlrec->locking_xid);
8678                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
8679                 PageSetLSN(page, lsn);
8680                 MarkBufferDirty(buffer);
8681         }
8682         if (BufferIsValid(buffer))
8683                 UnlockReleaseBuffer(buffer);
8684 }
8685
8686 static void
8687 heap_xlog_lock_updated(XLogReaderState *record)
8688 {
8689         XLogRecPtr      lsn = record->EndRecPtr;
8690         xl_heap_lock_updated *xlrec;
8691         Buffer          buffer;
8692         Page            page;
8693         OffsetNumber offnum;
8694         ItemId          lp = NULL;
8695         HeapTupleHeader htup;
8696
8697         xlrec = (xl_heap_lock_updated *) XLogRecGetData(record);
8698
8699         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8700         {
8701                 page = BufferGetPage(buffer);
8702
8703                 offnum = xlrec->offnum;
8704                 if (PageGetMaxOffsetNumber(page) >= offnum)
8705                         lp = PageGetItemId(page, offnum);
8706
8707                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8708                         elog(PANIC, "invalid lp");
8709
8710                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8711
8712                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
8713                                                                    &htup->t_infomask2);
8714                 HeapTupleHeaderSetXmax(htup, xlrec->xmax);
8715
8716                 PageSetLSN(page, lsn);
8717                 MarkBufferDirty(buffer);
8718         }
8719         if (BufferIsValid(buffer))
8720                 UnlockReleaseBuffer(buffer);
8721 }
8722
8723 static void
8724 heap_xlog_inplace(XLogReaderState *record)
8725 {
8726         XLogRecPtr      lsn = record->EndRecPtr;
8727         xl_heap_inplace *xlrec = (xl_heap_inplace *) XLogRecGetData(record);
8728         Buffer          buffer;
8729         Page            page;
8730         OffsetNumber offnum;
8731         ItemId          lp = NULL;
8732         HeapTupleHeader htup;
8733         uint32          oldlen;
8734         Size            newlen;
8735
8736         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8737         {
8738                 char       *newtup = XLogRecGetBlockData(record, 0, &newlen);
8739
8740                 page = BufferGetPage(buffer);
8741
8742                 offnum = xlrec->offnum;
8743                 if (PageGetMaxOffsetNumber(page) >= offnum)
8744                         lp = PageGetItemId(page, offnum);
8745
8746                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8747                         elog(PANIC, "invalid lp");
8748
8749                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8750
8751                 oldlen = ItemIdGetLength(lp) - htup->t_hoff;
8752                 if (oldlen != newlen)
8753                         elog(PANIC, "wrong tuple length");
8754
8755                 memcpy((char *) htup + htup->t_hoff, newtup, newlen);
8756
8757                 PageSetLSN(page, lsn);
8758                 MarkBufferDirty(buffer);
8759         }
8760         if (BufferIsValid(buffer))
8761                 UnlockReleaseBuffer(buffer);
8762 }
8763
8764 void
8765 heap_redo(XLogReaderState *record)
8766 {
8767         uint8           info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8768
8769         /*
8770          * These operations don't overwrite MVCC data so no conflict processing is
8771          * required. The ones in heap2 rmgr do.
8772          */
8773
8774         switch (info & XLOG_HEAP_OPMASK)
8775         {
8776                 case XLOG_HEAP_INSERT:
8777                         heap_xlog_insert(record);
8778                         break;
8779                 case XLOG_HEAP_DELETE:
8780                         heap_xlog_delete(record);
8781                         break;
8782                 case XLOG_HEAP_UPDATE:
8783                         heap_xlog_update(record, false);
8784                         break;
8785                 case XLOG_HEAP_HOT_UPDATE:
8786                         heap_xlog_update(record, true);
8787                         break;
8788                 case XLOG_HEAP_CONFIRM:
8789                         heap_xlog_confirm(record);
8790                         break;
8791                 case XLOG_HEAP_LOCK:
8792                         heap_xlog_lock(record);
8793                         break;
8794                 case XLOG_HEAP_INPLACE:
8795                         heap_xlog_inplace(record);
8796                         break;
8797                 default:
8798                         elog(PANIC, "heap_redo: unknown op code %u", info);
8799         }
8800 }
8801
8802 void
8803 heap2_redo(XLogReaderState *record)
8804 {
8805         uint8           info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8806
8807         switch (info & XLOG_HEAP_OPMASK)
8808         {
8809                 case XLOG_HEAP2_CLEAN:
8810                         heap_xlog_clean(record);
8811                         break;
8812                 case XLOG_HEAP2_FREEZE_PAGE:
8813                         heap_xlog_freeze_page(record);
8814                         break;
8815                 case XLOG_HEAP2_CLEANUP_INFO:
8816                         heap_xlog_cleanup_info(record);
8817                         break;
8818                 case XLOG_HEAP2_VISIBLE:
8819                         heap_xlog_visible(record);
8820                         break;
8821                 case XLOG_HEAP2_MULTI_INSERT:
8822                         heap_xlog_multi_insert(record);
8823                         break;
8824                 case XLOG_HEAP2_LOCK_UPDATED:
8825                         heap_xlog_lock_updated(record);
8826                         break;
8827                 case XLOG_HEAP2_NEW_CID:
8828
8829                         /*
8830                          * Nothing to do on a real replay, only used during logical
8831                          * decoding.
8832                          */
8833                         break;
8834                 case XLOG_HEAP2_REWRITE:
8835                         heap_xlog_logical_rewrite(record);
8836                         break;
8837                 default:
8838                         elog(PANIC, "heap2_redo: unknown op code %u", info);
8839         }
8840 }
8841
8842 /*
8843  *      heap_sync               - sync a heap, for use when no WAL has been written
8844  *
8845  * This forces the heap contents (including TOAST heap if any) down to disk.
8846  * If we skipped using WAL, and WAL is otherwise needed, we must force the
8847  * relation down to disk before it's safe to commit the transaction.  This
8848  * requires writing out any dirty buffers and then doing a forced fsync.
8849  *
8850  * Indexes are not touched.  (Currently, index operations associated with
8851  * the commands that use this are WAL-logged and so do not need fsync.
8852  * That behavior might change someday, but in any case it's likely that
8853  * any fsync decisions required would be per-index and hence not appropriate
8854  * to be done here.)
8855  */
8856 void
8857 heap_sync(Relation rel)
8858 {
8859         /* non-WAL-logged tables never need fsync */
8860         if (!RelationNeedsWAL(rel))
8861                 return;
8862
8863         /* main heap */
8864         FlushRelationBuffers(rel);
8865         /* FlushRelationBuffers will have opened rd_smgr */
8866         smgrimmedsync(rel->rd_smgr, MAIN_FORKNUM);
8867
8868         /* FSM is not critical, don't bother syncing it */
8869
8870         /* toast heap, if any */
8871         if (OidIsValid(rel->rd_rel->reltoastrelid))
8872         {
8873                 Relation        toastrel;
8874
8875                 toastrel = heap_open(rel->rd_rel->reltoastrelid, AccessShareLock);
8876                 FlushRelationBuffers(toastrel);
8877                 smgrimmedsync(toastrel->rd_smgr, MAIN_FORKNUM);
8878                 heap_close(toastrel, AccessShareLock);
8879         }
8880 }