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) 1996-2019, 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  *              heap_beginscan  - begin relation scan
  16  *              heap_rescan             - restart a relation scan
  17  *              heap_endscan    - end relation scan
  18  *              heap_getnext    - retrieve next tuple in scan
  19  *              heap_fetch              - retrieve tuple with given tid
  20  *              heap_insert             - insert tuple into a relation
  21  *              heap_multi_insert - insert multiple tuples into a relation
  22  *              heap_delete             - delete a tuple from a relation
  23  *              heap_update             - replace a tuple in a relation with another tuple
  24  *              heap_sync               - sync heap, for when no WAL has been written
  25  *
  26  * NOTES
  27  *        This file contains the heap_ routines which implement
  28  *        the POSTGRES heap access method used for all POSTGRES
  29  *        relations.
  30  *
  31  *-------------------------------------------------------------------------
  32  */
  33 #include "postgres.h"
  34
  35 #include "access/bufmask.h"
  36 #include "access/genam.h"
  37 #include "access/heapam.h"
  38 #include "access/heapam_xlog.h"
  39 #include "access/hio.h"
  40 #include "access/multixact.h"
  41 #include "access/parallel.h"
  42 #include "access/relscan.h"
  43 #include "access/sysattr.h"
  44 #include "access/tableam.h"
  45 #include "access/transam.h"
  46 #include "access/tuptoaster.h"
  47 #include "access/valid.h"
  48 #include "access/visibilitymap.h"
  49 #include "access/xact.h"
  50 #include "access/xlog.h"
  51 #include "access/xloginsert.h"
  52 #include "access/xlogutils.h"
  53 #include "catalog/catalog.h"
  54 #include "miscadmin.h"
  55 #include "pgstat.h"
  56 #include "port/atomics.h"
  57 #include "storage/bufmgr.h"
  58 #include "storage/freespace.h"
  59 #include "storage/lmgr.h"
  60 #include "storage/predicate.h"
  61 #include "storage/procarray.h"
  62 #include "storage/smgr.h"
  63 #include "storage/spin.h"
  64 #include "storage/standby.h"
  65 #include "utils/datum.h"
  66 #include "utils/inval.h"
  67 #include "utils/lsyscache.h"
  68 #include "utils/relcache.h"
  69 #include "utils/snapmgr.h"
  70 #include "utils/spccache.h"
  71
  72
  73 static HeapTuple heap_prepare_insert(Relation relation, HeapTuple tup,
  74                                         TransactionId xid, CommandId cid, int options);
  75 static XLogRecPtr log_heap_update(Relation reln, Buffer oldbuf,
  76                                 Buffer newbuf, HeapTuple oldtup,
  77                                 HeapTuple newtup, HeapTuple old_key_tup,
  78                                 bool all_visible_cleared, bool new_all_visible_cleared);
  79 static Bitmapset *HeapDetermineModifiedColumns(Relation relation,
  80                                                          Bitmapset *interesting_cols,
  81                                                          HeapTuple oldtup, HeapTuple newtup);
  82 static bool heap_acquire_tuplock(Relation relation, ItemPointer tid,
  83                                          LockTupleMode mode, LockWaitPolicy wait_policy,
  84                                          bool *have_tuple_lock);
  85 static void compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
  86                                                   uint16 old_infomask2, TransactionId add_to_xmax,
  87                                                   LockTupleMode mode, bool is_update,
  88                                                   TransactionId *result_xmax, uint16 *result_infomask,
  89                                                   uint16 *result_infomask2);
  90 static TM_Result heap_lock_updated_tuple(Relation rel, HeapTuple tuple,
  91                                                 ItemPointer ctid, TransactionId xid,
  92                                                 LockTupleMode mode);
  93 static void GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
  94                                            uint16 *new_infomask2);
  95 static TransactionId MultiXactIdGetUpdateXid(TransactionId xmax,
  96                                                 uint16 t_infomask);
  97 static bool DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
  98                                                 LockTupleMode lockmode);
  99 static void MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
 100                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
 101                                 int *remaining);
 102 static bool ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
 103                                                    uint16 infomask, Relation rel, int *remaining);
 104 static XLogRecPtr log_heap_new_cid(Relation relation, HeapTuple tup);
 105 static HeapTuple ExtractReplicaIdentity(Relation rel, HeapTuple tup, bool key_modified,
 106                                            bool *copy);
 107
 108
 109 /*
 110  * Each tuple lock mode has a corresponding heavyweight lock, and one or two
 111  * corresponding MultiXactStatuses (one to merely lock tuples, another one to
 112  * update them).  This table (and the macros below) helps us determine the
 113  * heavyweight lock mode and MultiXactStatus values to use for any particular
 114  * tuple lock strength.
 115  *
 116  * Don't look at lockstatus/updstatus directly!  Use get_mxact_status_for_lock
 117  * instead.
 118  */
 119 static const struct
 120 {
 121         LOCKMODE        hwlock;
 122         int                     lockstatus;
 123         int                     updstatus;
 124 }
 125
 126                         tupleLockExtraInfo[MaxLockTupleMode + 1] =
 127 {
 128         {                                                       /* LockTupleKeyShare */
 129                 AccessShareLock,
 130                 MultiXactStatusForKeyShare,
 131                 -1                                              /* KeyShare does not allow updating tuples */
 132         },
 133         {                                                       /* LockTupleShare */
 134                 RowShareLock,
 135                 MultiXactStatusForShare,
 136                 -1                                              /* Share does not allow updating tuples */
 137         },
 138         {                                                       /* LockTupleNoKeyExclusive */
 139                 ExclusiveLock,
 140                 MultiXactStatusForNoKeyUpdate,
 141                 MultiXactStatusNoKeyUpdate
 142         },
 143         {                                                       /* LockTupleExclusive */
 144                 AccessExclusiveLock,
 145                 MultiXactStatusForUpdate,
 146                 MultiXactStatusUpdate
 147         }
 148 };
 149
 150 /* Get the LOCKMODE for a given MultiXactStatus */
 151 #define LOCKMODE_from_mxstatus(status) \
 152                         (tupleLockExtraInfo[TUPLOCK_from_mxstatus((status))].hwlock)
 153
 154 /*
 155  * Acquire heavyweight locks on tuples, using a LockTupleMode strength value.
 156  * This is more readable than having every caller translate it to lock.h's
 157  * LOCKMODE.
 158  */
 159 #define LockTupleTuplock(rel, tup, mode) \
 160         LockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 161 #define UnlockTupleTuplock(rel, tup, mode) \
 162         UnlockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 163 #define ConditionalLockTupleTuplock(rel, tup, mode) \
 164         ConditionalLockTuple((rel), (tup), tupleLockExtraInfo[mode].hwlock)
 165
 166 #ifdef USE_PREFETCH
 167 /*
 168  * heap_compute_xid_horizon_for_tuples and xid_horizon_prefetch_buffer use
 169  * this structure to coordinate prefetching activity.
 170  */
 171 typedef struct
 172 {
 173         BlockNumber cur_hblkno;
 174         int                     next_item;
 175         int                     nitems;
 176         ItemPointerData *tids;
 177 } XidHorizonPrefetchState;
 178 #endif
 179
 180 /*
 181  * This table maps tuple lock strength values for each particular
 182  * MultiXactStatus value.
 183  */
 184 static const int MultiXactStatusLock[MaxMultiXactStatus + 1] =
 185 {
 186         LockTupleKeyShare,                      /* ForKeyShare */
 187         LockTupleShare,                         /* ForShare */
 188         LockTupleNoKeyExclusive,        /* ForNoKeyUpdate */
 189         LockTupleExclusive,                     /* ForUpdate */
 190         LockTupleNoKeyExclusive,        /* NoKeyUpdate */
 191         LockTupleExclusive                      /* Update */
 192 };
 193
 194 /* Get the LockTupleMode for a given MultiXactStatus */
 195 #define TUPLOCK_from_mxstatus(status) \
 196                         (MultiXactStatusLock[(status)])
 197
 198 /* ----------------------------------------------------------------
 199  *                                               heap support routines
 200  * ----------------------------------------------------------------
 201  */
 202
 203 /* ----------------
 204  *              initscan - scan code common to heap_beginscan and heap_rescan
 205  * ----------------
 206  */
 207 static void
 208 initscan(HeapScanDesc scan, ScanKey key, bool keep_startblock)
 209 {
 210         ParallelBlockTableScanDesc bpscan = NULL;
 211         bool            allow_strat;
 212         bool            allow_sync;
 213
 214         /*
 215          * Determine the number of blocks we have to scan.
 216          *
 217          * It is sufficient to do this once at scan start, since any tuples added
 218          * while the scan is in progress will be invisible to my snapshot anyway.
 219          * (That is not true when using a non-MVCC snapshot.  However, we couldn't
 220          * guarantee to return tuples added after scan start anyway, since they
 221          * might go into pages we already scanned.  To guarantee consistent
 222          * results for a non-MVCC snapshot, the caller must hold some higher-level
 223          * lock that ensures the interesting tuple(s) won't change.)
 224          */
 225         if (scan->rs_base.rs_parallel != NULL)
 226         {
 227                 bpscan = (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
 228                 scan->rs_nblocks = bpscan->phs_nblocks;
 229         }
 230         else
 231                 scan->rs_nblocks = RelationGetNumberOfBlocks(scan->rs_base.rs_rd);
 232
 233         /*
 234          * If the table is large relative to NBuffers, use a bulk-read access
 235          * strategy and enable synchronized scanning (see syncscan.c).  Although
 236          * the thresholds for these features could be different, we make them the
 237          * same so that there are only two behaviors to tune rather than four.
 238          * (However, some callers need to be able to disable one or both of these
 239          * behaviors, independently of the size of the table; also there is a GUC
 240          * variable that can disable synchronized scanning.)
 241          *
 242          * Note that heap_parallelscan_initialize has a very similar test; if you
 243          * change this, consider changing that one, too.
 244          */
 245         if (!RelationUsesLocalBuffers(scan->rs_base.rs_rd) &&
 246                 scan->rs_nblocks > NBuffers / 4)
 247         {
 248                 allow_strat = scan->rs_base.rs_allow_strat;
 249                 allow_sync = scan->rs_base.rs_allow_sync;
 250         }
 251         else
 252                 allow_strat = allow_sync = false;
 253
 254         if (allow_strat)
 255         {
 256                 /* During a rescan, keep the previous strategy object. */
 257                 if (scan->rs_strategy == NULL)
 258                         scan->rs_strategy = GetAccessStrategy(BAS_BULKREAD);
 259         }
 260         else
 261         {
 262                 if (scan->rs_strategy != NULL)
 263                         FreeAccessStrategy(scan->rs_strategy);
 264                 scan->rs_strategy = NULL;
 265         }
 266
 267         if (scan->rs_base.rs_parallel != NULL)
 268         {
 269                 /* For parallel scan, believe whatever ParallelTableScanDesc says. */
 270                 scan->rs_base.rs_syncscan = scan->rs_base.rs_parallel->phs_syncscan;
 271         }
 272         else if (keep_startblock)
 273         {
 274                 /*
 275                  * When rescanning, we want to keep the previous startblock setting,
 276                  * so that rewinding a cursor doesn't generate surprising results.
 277                  * Reset the active syncscan setting, though.
 278                  */
 279                 scan->rs_base.rs_syncscan = (allow_sync && synchronize_seqscans);
 280         }
 281         else if (allow_sync && synchronize_seqscans)
 282         {
 283                 scan->rs_base.rs_syncscan = true;
 284                 scan->rs_startblock = ss_get_location(scan->rs_base.rs_rd, scan->rs_nblocks);
 285         }
 286         else
 287         {
 288                 scan->rs_base.rs_syncscan = false;
 289                 scan->rs_startblock = 0;
 290         }
 291
 292         scan->rs_numblocks = InvalidBlockNumber;
 293         scan->rs_inited = false;
 294         scan->rs_ctup.t_data = NULL;
 295         ItemPointerSetInvalid(&scan->rs_ctup.t_self);
 296         scan->rs_cbuf = InvalidBuffer;
 297         scan->rs_cblock = InvalidBlockNumber;
 298
 299         /* page-at-a-time fields are always invalid when not rs_inited */
 300
 301         /*
 302          * copy the scan key, if appropriate
 303          */
 304         if (key != NULL)
 305                 memcpy(scan->rs_base.rs_key, key, scan->rs_base.rs_nkeys * sizeof(ScanKeyData));
 306
 307         /*
 308          * Currently, we don't have a stats counter for bitmap heap scans (but the
 309          * underlying bitmap index scans will be counted) or sample scans (we only
 310          * update stats for tuple fetches there)
 311          */
 312         if (!scan->rs_base.rs_bitmapscan && !scan->rs_base.rs_samplescan)
 313                 pgstat_count_heap_scan(scan->rs_base.rs_rd);
 314 }
 315
 316 /*
 317  * heap_setscanlimits - restrict range of a heapscan
 318  *
 319  * startBlk is the page to start at
 320  * numBlks is number of pages to scan (InvalidBlockNumber means "all")
 321  */
 322 void
 323 heap_setscanlimits(TableScanDesc sscan, BlockNumber startBlk, BlockNumber numBlks)
 324 {
 325         HeapScanDesc scan = (HeapScanDesc) sscan;
 326
 327         Assert(!scan->rs_inited);       /* else too late to change */
 328         Assert(!scan->rs_base.rs_syncscan); /* else rs_startblock is significant */
 329
 330         /* Check startBlk is valid (but allow case of zero blocks...) */
 331         Assert(startBlk == 0 || startBlk < scan->rs_nblocks);
 332
 333         scan->rs_startblock = startBlk;
 334         scan->rs_numblocks = numBlks;
 335 }
 336
 337 /*
 338  * heapgetpage - subroutine for heapgettup()
 339  *
 340  * This routine reads and pins the specified page of the relation.
 341  * In page-at-a-time mode it performs additional work, namely determining
 342  * which tuples on the page are visible.
 343  */
 344 void
 345 heapgetpage(TableScanDesc sscan, BlockNumber page)
 346 {
 347         HeapScanDesc scan = (HeapScanDesc) sscan;
 348         Buffer          buffer;
 349         Snapshot        snapshot;
 350         Page            dp;
 351         int                     lines;
 352         int                     ntup;
 353         OffsetNumber lineoff;
 354         ItemId          lpp;
 355         bool            all_visible;
 356
 357         Assert(page < scan->rs_nblocks);
 358
 359         /* release previous scan buffer, if any */
 360         if (BufferIsValid(scan->rs_cbuf))
 361         {
 362                 ReleaseBuffer(scan->rs_cbuf);
 363                 scan->rs_cbuf = InvalidBuffer;
 364         }
 365
 366         /*
 367          * Be sure to check for interrupts at least once per page.  Checks at
 368          * higher code levels won't be able to stop a seqscan that encounters many
 369          * pages' worth of consecutive dead tuples.
 370          */
 371         CHECK_FOR_INTERRUPTS();
 372
 373         /* read page using selected strategy */
 374         scan->rs_cbuf = ReadBufferExtended(scan->rs_base.rs_rd, MAIN_FORKNUM, page,
 375                                                                            RBM_NORMAL, scan->rs_strategy);
 376         scan->rs_cblock = page;
 377
 378         if (!scan->rs_base.rs_pageatatime)
 379                 return;
 380
 381         buffer = scan->rs_cbuf;
 382         snapshot = scan->rs_base.rs_snapshot;
 383
 384         /*
 385          * Prune and repair fragmentation for the whole page, if possible.
 386          */
 387         heap_page_prune_opt(scan->rs_base.rs_rd, buffer);
 388
 389         /*
 390          * We must hold share lock on the buffer content while examining tuple
 391          * visibility.  Afterwards, however, the tuples we have found to be
 392          * visible are guaranteed good as long as we hold the buffer pin.
 393          */
 394         LockBuffer(buffer, BUFFER_LOCK_SHARE);
 395
 396         dp = BufferGetPage(buffer);
 397         TestForOldSnapshot(snapshot, scan->rs_base.rs_rd, dp);
 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_base.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_base.rs_rd,
 443                                                                                         &loctup, 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_base.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_base.rs_parallel != NULL)
 513                         {
 514                                 ParallelBlockTableScanDesc pbscan =
 515                                 (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
 516
 517                                 table_block_parallelscan_startblock_init(scan->rs_base.rs_rd,
 518                                                                                                                  pbscan);
 519
 520                                 page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
 521                                                                                                                  pbscan);
 522
 523                                 /* Other processes might have already finished the scan. */
 524                                 if (page == InvalidBlockNumber)
 525                                 {
 526                                         Assert(!BufferIsValid(scan->rs_cbuf));
 527                                         tuple->t_data = NULL;
 528                                         return;
 529                                 }
 530                         }
 531                         else
 532                                 page = scan->rs_startblock; /* first page */
 533                         heapgetpage((TableScanDesc) scan, page);
 534                         lineoff = FirstOffsetNumber;    /* first offnum */
 535                         scan->rs_inited = true;
 536                 }
 537                 else
 538                 {
 539                         /* continue from previously returned page/tuple */
 540                         page = scan->rs_cblock; /* current page */
 541                         lineoff =                       /* next offnum */
 542                                 OffsetNumberNext(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 543                 }
 544
 545                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 546
 547                 dp = BufferGetPage(scan->rs_cbuf);
 548                 TestForOldSnapshot(snapshot, scan->rs_base.rs_rd, dp);
 549                 lines = PageGetMaxOffsetNumber(dp);
 550                 /* page and lineoff now reference the physically next tid */
 551
 552                 linesleft = lines - lineoff + 1;
 553         }
 554         else if (backward)
 555         {
 556                 /* backward parallel scan not supported */
 557                 Assert(scan->rs_base.rs_parallel == NULL);
 558
 559                 if (!scan->rs_inited)
 560                 {
 561                         /*
 562                          * return null immediately if relation is empty
 563                          */
 564                         if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
 565                         {
 566                                 Assert(!BufferIsValid(scan->rs_cbuf));
 567                                 tuple->t_data = NULL;
 568                                 return;
 569                         }
 570
 571                         /*
 572                          * Disable reporting to syncscan logic in a backwards scan; it's
 573                          * not very likely anyone else is doing the same thing at the same
 574                          * time, and much more likely that we'll just bollix things for
 575                          * forward scanners.
 576                          */
 577                         scan->rs_base.rs_syncscan = false;
 578                         /* start from last page of the scan */
 579                         if (scan->rs_startblock > 0)
 580                                 page = scan->rs_startblock - 1;
 581                         else
 582                                 page = scan->rs_nblocks - 1;
 583                         heapgetpage((TableScanDesc) scan, page);
 584                 }
 585                 else
 586                 {
 587                         /* continue from previously returned page/tuple */
 588                         page = scan->rs_cblock; /* current page */
 589                 }
 590
 591                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 592
 593                 dp = BufferGetPage(scan->rs_cbuf);
 594                 TestForOldSnapshot(snapshot, scan->rs_base.rs_rd, dp);
 595                 lines = PageGetMaxOffsetNumber(dp);
 596
 597                 if (!scan->rs_inited)
 598                 {
 599                         lineoff = lines;        /* final offnum */
 600                         scan->rs_inited = true;
 601                 }
 602                 else
 603                 {
 604                         lineoff =                       /* previous offnum */
 605                                 OffsetNumberPrev(ItemPointerGetOffsetNumber(&(tuple->t_self)));
 606                 }
 607                 /* page and lineoff now reference the physically previous tid */
 608
 609                 linesleft = lineoff;
 610         }
 611         else
 612         {
 613                 /*
 614                  * ``no movement'' scan direction: refetch prior tuple
 615                  */
 616                 if (!scan->rs_inited)
 617                 {
 618                         Assert(!BufferIsValid(scan->rs_cbuf));
 619                         tuple->t_data = NULL;
 620                         return;
 621                 }
 622
 623                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 624                 if (page != scan->rs_cblock)
 625                         heapgetpage((TableScanDesc) scan, page);
 626
 627                 /* Since the tuple was previously fetched, needn't lock page here */
 628                 dp = BufferGetPage(scan->rs_cbuf);
 629                 TestForOldSnapshot(snapshot, scan->rs_base.rs_rd, dp);
 630                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 631                 lpp = PageGetItemId(dp, lineoff);
 632                 Assert(ItemIdIsNormal(lpp));
 633
 634                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 635                 tuple->t_len = ItemIdGetLength(lpp);
 636
 637                 return;
 638         }
 639
 640         /*
 641          * advance the scan until we find a qualifying tuple or run out of stuff
 642          * to scan
 643          */
 644         lpp = PageGetItemId(dp, lineoff);
 645         for (;;)
 646         {
 647                 while (linesleft > 0)
 648                 {
 649                         if (ItemIdIsNormal(lpp))
 650                         {
 651                                 bool            valid;
 652
 653                                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 654                                 tuple->t_len = ItemIdGetLength(lpp);
 655                                 ItemPointerSet(&(tuple->t_self), page, lineoff);
 656
 657                                 /*
 658                                  * if current tuple qualifies, return it.
 659                                  */
 660                                 valid = HeapTupleSatisfiesVisibility(tuple,
 661                                                                                                          snapshot,
 662                                                                                                          scan->rs_cbuf);
 663
 664                                 CheckForSerializableConflictOut(valid, scan->rs_base.rs_rd,
 665                                                                                                 tuple, scan->rs_cbuf,
 666                                                                                                 snapshot);
 667
 668                                 if (valid && key != NULL)
 669                                         HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
 670                                                                 nkeys, key, valid);
 671
 672                                 if (valid)
 673                                 {
 674                                         LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 675                                         return;
 676                                 }
 677                         }
 678
 679                         /*
 680                          * otherwise move to the next item on the page
 681                          */
 682                         --linesleft;
 683                         if (backward)
 684                         {
 685                                 --lpp;                  /* move back in this page's ItemId array */
 686                                 --lineoff;
 687                         }
 688                         else
 689                         {
 690                                 ++lpp;                  /* move forward in this page's ItemId array */
 691                                 ++lineoff;
 692                         }
 693                 }
 694
 695                 /*
 696                  * if we get here, it means we've exhausted the items on this page and
 697                  * it's time to move to the next.
 698                  */
 699                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_UNLOCK);
 700
 701                 /*
 702                  * advance to next/prior page and detect end of scan
 703                  */
 704                 if (backward)
 705                 {
 706                         finished = (page == scan->rs_startblock) ||
 707                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
 708                         if (page == 0)
 709                                 page = scan->rs_nblocks;
 710                         page--;
 711                 }
 712                 else if (scan->rs_base.rs_parallel != NULL)
 713                 {
 714                         ParallelBlockTableScanDesc pbscan =
 715                         (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
 716
 717                         page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
 718                                                                                                          pbscan);
 719                         finished = (page == InvalidBlockNumber);
 720                 }
 721                 else
 722                 {
 723                         page++;
 724                         if (page >= scan->rs_nblocks)
 725                                 page = 0;
 726                         finished = (page == scan->rs_startblock) ||
 727                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
 728
 729                         /*
 730                          * Report our new scan position for synchronization purposes. We
 731                          * don't do that when moving backwards, however. That would just
 732                          * mess up any other forward-moving scanners.
 733                          *
 734                          * Note: we do this before checking for end of scan so that the
 735                          * final state of the position hint is back at the start of the
 736                          * rel.  That's not strictly necessary, but otherwise when you run
 737                          * the same query multiple times the starting position would shift
 738                          * a little bit backwards on every invocation, which is confusing.
 739                          * We don't guarantee any specific ordering in general, though.
 740                          */
 741                         if (scan->rs_base.rs_syncscan)
 742                                 ss_report_location(scan->rs_base.rs_rd, page);
 743                 }
 744
 745                 /*
 746                  * return NULL if we've exhausted all the pages
 747                  */
 748                 if (finished)
 749                 {
 750                         if (BufferIsValid(scan->rs_cbuf))
 751                                 ReleaseBuffer(scan->rs_cbuf);
 752                         scan->rs_cbuf = InvalidBuffer;
 753                         scan->rs_cblock = InvalidBlockNumber;
 754                         tuple->t_data = NULL;
 755                         scan->rs_inited = false;
 756                         return;
 757                 }
 758
 759                 heapgetpage((TableScanDesc) scan, page);
 760
 761                 LockBuffer(scan->rs_cbuf, BUFFER_LOCK_SHARE);
 762
 763                 dp = BufferGetPage(scan->rs_cbuf);
 764                 TestForOldSnapshot(snapshot, scan->rs_base.rs_rd, dp);
 765                 lines = PageGetMaxOffsetNumber((Page) dp);
 766                 linesleft = lines;
 767                 if (backward)
 768                 {
 769                         lineoff = lines;
 770                         lpp = PageGetItemId(dp, lines);
 771                 }
 772                 else
 773                 {
 774                         lineoff = FirstOffsetNumber;
 775                         lpp = PageGetItemId(dp, FirstOffsetNumber);
 776                 }
 777         }
 778 }
 779
 780 /* ----------------
 781  *              heapgettup_pagemode - fetch next heap tuple in page-at-a-time mode
 782  *
 783  *              Same API as heapgettup, but used in page-at-a-time mode
 784  *
 785  * The internal logic is much the same as heapgettup's too, but there are some
 786  * differences: we do not take the buffer content lock (that only needs to
 787  * happen inside heapgetpage), and we iterate through just the tuples listed
 788  * in rs_vistuples[] rather than all tuples on the page.  Notice that
 789  * lineindex is 0-based, where the corresponding loop variable lineoff in
 790  * heapgettup is 1-based.
 791  * ----------------
 792  */
 793 static void
 794 heapgettup_pagemode(HeapScanDesc scan,
 795                                         ScanDirection dir,
 796                                         int nkeys,
 797                                         ScanKey key)
 798 {
 799         HeapTuple       tuple = &(scan->rs_ctup);
 800         bool            backward = ScanDirectionIsBackward(dir);
 801         BlockNumber page;
 802         bool            finished;
 803         Page            dp;
 804         int                     lines;
 805         int                     lineindex;
 806         OffsetNumber lineoff;
 807         int                     linesleft;
 808         ItemId          lpp;
 809
 810         /*
 811          * calculate next starting lineindex, given scan direction
 812          */
 813         if (ScanDirectionIsForward(dir))
 814         {
 815                 if (!scan->rs_inited)
 816                 {
 817                         /*
 818                          * return null immediately if relation is empty
 819                          */
 820                         if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
 821                         {
 822                                 Assert(!BufferIsValid(scan->rs_cbuf));
 823                                 tuple->t_data = NULL;
 824                                 return;
 825                         }
 826                         if (scan->rs_base.rs_parallel != NULL)
 827                         {
 828                                 ParallelBlockTableScanDesc pbscan =
 829                                 (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
 830
 831                                 table_block_parallelscan_startblock_init(scan->rs_base.rs_rd,
 832                                                                                                                  pbscan);
 833
 834                                 page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
 835                                                                                                                  pbscan);
 836
 837                                 /* Other processes might have already finished the scan. */
 838                                 if (page == InvalidBlockNumber)
 839                                 {
 840                                         Assert(!BufferIsValid(scan->rs_cbuf));
 841                                         tuple->t_data = NULL;
 842                                         return;
 843                                 }
 844                         }
 845                         else
 846                                 page = scan->rs_startblock; /* first page */
 847                         heapgetpage((TableScanDesc) scan, page);
 848                         lineindex = 0;
 849                         scan->rs_inited = true;
 850                 }
 851                 else
 852                 {
 853                         /* continue from previously returned page/tuple */
 854                         page = scan->rs_cblock; /* current page */
 855                         lineindex = scan->rs_cindex + 1;
 856                 }
 857
 858                 dp = BufferGetPage(scan->rs_cbuf);
 859                 TestForOldSnapshot(scan->rs_base.rs_snapshot, scan->rs_base.rs_rd, dp);
 860                 lines = scan->rs_ntuples;
 861                 /* page and lineindex now reference the next visible tid */
 862
 863                 linesleft = lines - lineindex;
 864         }
 865         else if (backward)
 866         {
 867                 /* backward parallel scan not supported */
 868                 Assert(scan->rs_base.rs_parallel == NULL);
 869
 870                 if (!scan->rs_inited)
 871                 {
 872                         /*
 873                          * return null immediately if relation is empty
 874                          */
 875                         if (scan->rs_nblocks == 0 || scan->rs_numblocks == 0)
 876                         {
 877                                 Assert(!BufferIsValid(scan->rs_cbuf));
 878                                 tuple->t_data = NULL;
 879                                 return;
 880                         }
 881
 882                         /*
 883                          * Disable reporting to syncscan logic in a backwards scan; it's
 884                          * not very likely anyone else is doing the same thing at the same
 885                          * time, and much more likely that we'll just bollix things for
 886                          * forward scanners.
 887                          */
 888                         scan->rs_base.rs_syncscan = false;
 889                         /* start from last page of the scan */
 890                         if (scan->rs_startblock > 0)
 891                                 page = scan->rs_startblock - 1;
 892                         else
 893                                 page = scan->rs_nblocks - 1;
 894                         heapgetpage((TableScanDesc) scan, page);
 895                 }
 896                 else
 897                 {
 898                         /* continue from previously returned page/tuple */
 899                         page = scan->rs_cblock; /* current page */
 900                 }
 901
 902                 dp = BufferGetPage(scan->rs_cbuf);
 903                 TestForOldSnapshot(scan->rs_base.rs_snapshot, scan->rs_base.rs_rd, dp);
 904                 lines = scan->rs_ntuples;
 905
 906                 if (!scan->rs_inited)
 907                 {
 908                         lineindex = lines - 1;
 909                         scan->rs_inited = true;
 910                 }
 911                 else
 912                 {
 913                         lineindex = scan->rs_cindex - 1;
 914                 }
 915                 /* page and lineindex now reference the previous visible tid */
 916
 917                 linesleft = lineindex + 1;
 918         }
 919         else
 920         {
 921                 /*
 922                  * ``no movement'' scan direction: refetch prior tuple
 923                  */
 924                 if (!scan->rs_inited)
 925                 {
 926                         Assert(!BufferIsValid(scan->rs_cbuf));
 927                         tuple->t_data = NULL;
 928                         return;
 929                 }
 930
 931                 page = ItemPointerGetBlockNumber(&(tuple->t_self));
 932                 if (page != scan->rs_cblock)
 933                         heapgetpage((TableScanDesc) scan, page);
 934
 935                 /* Since the tuple was previously fetched, needn't lock page here */
 936                 dp = BufferGetPage(scan->rs_cbuf);
 937                 TestForOldSnapshot(scan->rs_base.rs_snapshot, scan->rs_base.rs_rd, dp);
 938                 lineoff = ItemPointerGetOffsetNumber(&(tuple->t_self));
 939                 lpp = PageGetItemId(dp, lineoff);
 940                 Assert(ItemIdIsNormal(lpp));
 941
 942                 tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 943                 tuple->t_len = ItemIdGetLength(lpp);
 944
 945                 /* check that rs_cindex is in sync */
 946                 Assert(scan->rs_cindex < scan->rs_ntuples);
 947                 Assert(lineoff == scan->rs_vistuples[scan->rs_cindex]);
 948
 949                 return;
 950         }
 951
 952         /*
 953          * advance the scan until we find a qualifying tuple or run out of stuff
 954          * to scan
 955          */
 956         for (;;)
 957         {
 958                 while (linesleft > 0)
 959                 {
 960                         lineoff = scan->rs_vistuples[lineindex];
 961                         lpp = PageGetItemId(dp, lineoff);
 962                         Assert(ItemIdIsNormal(lpp));
 963
 964                         tuple->t_data = (HeapTupleHeader) PageGetItem((Page) dp, lpp);
 965                         tuple->t_len = ItemIdGetLength(lpp);
 966                         ItemPointerSet(&(tuple->t_self), page, lineoff);
 967
 968                         /*
 969                          * if current tuple qualifies, return it.
 970                          */
 971                         if (key != NULL)
 972                         {
 973                                 bool            valid;
 974
 975                                 HeapKeyTest(tuple, RelationGetDescr(scan->rs_base.rs_rd),
 976                                                         nkeys, key, valid);
 977                                 if (valid)
 978                                 {
 979                                         scan->rs_cindex = lineindex;
 980                                         return;
 981                                 }
 982                         }
 983                         else
 984                         {
 985                                 scan->rs_cindex = lineindex;
 986                                 return;
 987                         }
 988
 989                         /*
 990                          * otherwise move to the next item on the page
 991                          */
 992                         --linesleft;
 993                         if (backward)
 994                                 --lineindex;
 995                         else
 996                                 ++lineindex;
 997                 }
 998
 999                 /*
1000                  * if we get here, it means we've exhausted the items on this page and
1001                  * it's time to move to the next.
1002                  */
1003                 if (backward)
1004                 {
1005                         finished = (page == scan->rs_startblock) ||
1006                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
1007                         if (page == 0)
1008                                 page = scan->rs_nblocks;
1009                         page--;
1010                 }
1011                 else if (scan->rs_base.rs_parallel != NULL)
1012                 {
1013                         ParallelBlockTableScanDesc pbscan =
1014                         (ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
1015
1016                         page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
1017                                                                                                          pbscan);
1018                         finished = (page == InvalidBlockNumber);
1019                 }
1020                 else
1021                 {
1022                         page++;
1023                         if (page >= scan->rs_nblocks)
1024                                 page = 0;
1025                         finished = (page == scan->rs_startblock) ||
1026                                 (scan->rs_numblocks != InvalidBlockNumber ? --scan->rs_numblocks == 0 : false);
1027
1028                         /*
1029                          * Report our new scan position for synchronization purposes. We
1030                          * don't do that when moving backwards, however. That would just
1031                          * mess up any other forward-moving scanners.
1032                          *
1033                          * Note: we do this before checking for end of scan so that the
1034                          * final state of the position hint is back at the start of the
1035                          * rel.  That's not strictly necessary, but otherwise when you run
1036                          * the same query multiple times the starting position would shift
1037                          * a little bit backwards on every invocation, which is confusing.
1038                          * We don't guarantee any specific ordering in general, though.
1039                          */
1040                         if (scan->rs_base.rs_syncscan)
1041                                 ss_report_location(scan->rs_base.rs_rd, page);
1042                 }
1043
1044                 /*
1045                  * return NULL if we've exhausted all the pages
1046                  */
1047                 if (finished)
1048                 {
1049                         if (BufferIsValid(scan->rs_cbuf))
1050                                 ReleaseBuffer(scan->rs_cbuf);
1051                         scan->rs_cbuf = InvalidBuffer;
1052                         scan->rs_cblock = InvalidBlockNumber;
1053                         tuple->t_data = NULL;
1054                         scan->rs_inited = false;
1055                         return;
1056                 }
1057
1058                 heapgetpage((TableScanDesc) scan, page);
1059
1060                 dp = BufferGetPage(scan->rs_cbuf);
1061                 TestForOldSnapshot(scan->rs_base.rs_snapshot, scan->rs_base.rs_rd, dp);
1062                 lines = scan->rs_ntuples;
1063                 linesleft = lines;
1064                 if (backward)
1065                         lineindex = lines - 1;
1066                 else
1067                         lineindex = 0;
1068         }
1069 }
1070
1071
1072 #if defined(DISABLE_COMPLEX_MACRO)
1073 /*
1074  * This is formatted so oddly so that the correspondence to the macro
1075  * definition in access/htup_details.h is maintained.
1076  */
1077 Datum
1078 fastgetattr(HeapTuple tup, int attnum, TupleDesc tupleDesc,
1079                         bool *isnull)
1080 {
1081         return (
1082                         (attnum) > 0 ?
1083                         (
1084                          (*(isnull) = false),
1085                          HeapTupleNoNulls(tup) ?
1086                          (
1087                           TupleDescAttr((tupleDesc), (attnum) - 1)->attcacheoff >= 0 ?
1088                           (
1089                            fetchatt(TupleDescAttr((tupleDesc), (attnum) - 1),
1090                                                 (char *) (tup)->t_data + (tup)->t_data->t_hoff +
1091                                                 TupleDescAttr((tupleDesc), (attnum) - 1)->attcacheoff)
1092                            )
1093                           :
1094                           nocachegetattr((tup), (attnum), (tupleDesc))
1095                           )
1096                          :
1097                          (
1098                           att_isnull((attnum) - 1, (tup)->t_data->t_bits) ?
1099                           (
1100                            (*(isnull) = true),
1101                            (Datum) NULL
1102                            )
1103                           :
1104                           (
1105                            nocachegetattr((tup), (attnum), (tupleDesc))
1106                            )
1107                           )
1108                          )
1109                         :
1110                         (
1111                          (Datum) NULL
1112                          )
1113                 );
1114 }
1115 #endif                                                  /* defined(DISABLE_COMPLEX_MACRO) */
1116
1117
1118 /* ----------------------------------------------------------------
1119  *                                       heap access method interface
1120  * ----------------------------------------------------------------
1121  */
1122
1123
1124 TableScanDesc
1125 heap_beginscan(Relation relation, Snapshot snapshot,
1126                            int nkeys, ScanKey key,
1127                            ParallelTableScanDesc parallel_scan,
1128                            bool allow_strat,
1129                            bool allow_sync,
1130                            bool allow_pagemode,
1131                            bool is_bitmapscan,
1132                            bool is_samplescan,
1133                            bool temp_snap)
1134 {
1135         HeapScanDesc scan;
1136
1137         /*
1138          * increment relation ref count while scanning relation
1139          *
1140          * This is just to make really sure the relcache entry won't go away while
1141          * the scan has a pointer to it.  Caller should be holding the rel open
1142          * anyway, so this is redundant in all normal scenarios...
1143          */
1144         RelationIncrementReferenceCount(relation);
1145
1146         /*
1147          * allocate and initialize scan descriptor
1148          */
1149         scan = (HeapScanDesc) palloc(sizeof(HeapScanDescData));
1150
1151         scan->rs_base.rs_rd = relation;
1152         scan->rs_base.rs_snapshot = snapshot;
1153         scan->rs_base.rs_nkeys = nkeys;
1154         scan->rs_base.rs_bitmapscan = is_bitmapscan;
1155         scan->rs_base.rs_samplescan = is_samplescan;
1156         scan->rs_strategy = NULL;       /* set in initscan */
1157         scan->rs_base.rs_allow_strat = allow_strat;
1158         scan->rs_base.rs_allow_sync = allow_sync;
1159         scan->rs_base.rs_temp_snap = temp_snap;
1160         scan->rs_base.rs_parallel = parallel_scan;
1161
1162         /*
1163          * we can use page-at-a-time mode if it's an MVCC-safe snapshot
1164          */
1165         scan->rs_base.rs_pageatatime =
1166                 allow_pagemode && snapshot && IsMVCCSnapshot(snapshot);
1167
1168         /*
1169          * For a seqscan in a serializable transaction, acquire a predicate lock
1170          * on the entire relation. This is required not only to lock all the
1171          * matching tuples, but also to conflict with new insertions into the
1172          * table. In an indexscan, we take page locks on the index pages covering
1173          * the range specified in the scan qual, but in a heap scan there is
1174          * nothing more fine-grained to lock. A bitmap scan is a different story,
1175          * there we have already scanned the index and locked the index pages
1176          * covering the predicate. But in that case we still have to lock any
1177          * matching heap tuples.
1178          */
1179         if (!is_bitmapscan)
1180                 PredicateLockRelation(relation, snapshot);
1181
1182         /* we only need to set this up once */
1183         scan->rs_ctup.t_tableOid = RelationGetRelid(relation);
1184
1185         /*
1186          * we do this here instead of in initscan() because heap_rescan also calls
1187          * initscan() and we don't want to allocate memory again
1188          */
1189         if (nkeys > 0)
1190                 scan->rs_base.rs_key = (ScanKey) palloc(sizeof(ScanKeyData) * nkeys);
1191         else
1192                 scan->rs_base.rs_key = NULL;
1193
1194         initscan(scan, key, false);
1195
1196         return (TableScanDesc) scan;
1197 }
1198
1199 void
1200 heap_rescan(TableScanDesc sscan, ScanKey key, bool set_params,
1201                         bool allow_strat, bool allow_sync, bool allow_pagemode)
1202 {
1203         HeapScanDesc scan = (HeapScanDesc) sscan;
1204
1205         if (set_params)
1206         {
1207                 scan->rs_base.rs_allow_strat = allow_strat;
1208                 scan->rs_base.rs_allow_sync = allow_sync;
1209                 scan->rs_base.rs_pageatatime =
1210                         allow_pagemode && IsMVCCSnapshot(scan->rs_base.rs_snapshot);
1211         }
1212
1213         /*
1214          * unpin scan buffers
1215          */
1216         if (BufferIsValid(scan->rs_cbuf))
1217                 ReleaseBuffer(scan->rs_cbuf);
1218
1219         /*
1220          * reinitialize scan descriptor
1221          */
1222         initscan(scan, key, true);
1223 }
1224
1225 void
1226 heap_endscan(TableScanDesc sscan)
1227 {
1228         HeapScanDesc scan = (HeapScanDesc) sscan;
1229
1230         /* Note: no locking manipulations needed */
1231
1232         /*
1233          * unpin scan buffers
1234          */
1235         if (BufferIsValid(scan->rs_cbuf))
1236                 ReleaseBuffer(scan->rs_cbuf);
1237
1238         /*
1239          * decrement relation reference count and free scan descriptor storage
1240          */
1241         RelationDecrementReferenceCount(scan->rs_base.rs_rd);
1242
1243         if (scan->rs_base.rs_key)
1244                 pfree(scan->rs_base.rs_key);
1245
1246         if (scan->rs_strategy != NULL)
1247                 FreeAccessStrategy(scan->rs_strategy);
1248
1249         if (scan->rs_base.rs_temp_snap)
1250                 UnregisterSnapshot(scan->rs_base.rs_snapshot);
1251
1252         pfree(scan);
1253 }
1254
1255 #ifdef HEAPDEBUGALL
1256 #define HEAPDEBUG_1 \
1257         elog(DEBUG2, "heap_getnext([%s,nkeys=%d],dir=%d) called", \
1258                  RelationGetRelationName(scan->rs_rd), scan->rs_nkeys, (int) direction)
1259 #define HEAPDEBUG_2 \
1260         elog(DEBUG2, "heap_getnext returning EOS")
1261 #define HEAPDEBUG_3 \
1262         elog(DEBUG2, "heap_getnext returning tuple")
1263 #else
1264 #define HEAPDEBUG_1
1265 #define HEAPDEBUG_2
1266 #define HEAPDEBUG_3
1267 #endif                                                  /* !defined(HEAPDEBUGALL) */
1268
1269
1270 HeapTuple
1271 heap_getnext(TableScanDesc sscan, ScanDirection direction)
1272 {
1273         HeapScanDesc scan = (HeapScanDesc) sscan;
1274
1275         /*
1276          * This is still widely used directly, without going through table AM, so
1277          * add a safety check.  It's possible we should, at a later point,
1278          * downgrade this to an assert. The reason for checking the AM routine,
1279          * rather than the AM oid, is that this allows to write regression tests
1280          * that create another AM reusing the heap handler.
1281          */
1282         if (unlikely(sscan->rs_rd->rd_tableam != GetHeapamTableAmRoutine()))
1283                 ereport(ERROR,
1284                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1285                                  errmsg("only heap AM is supported")));
1286
1287         /* Note: no locking manipulations needed */
1288
1289         HEAPDEBUG_1;                            /* heap_getnext( info ) */
1290
1291         if (scan->rs_base.rs_pageatatime)
1292                 heapgettup_pagemode(scan, direction,
1293                                                         scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1294         else
1295                 heapgettup(scan, direction,
1296                                    scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1297
1298         if (scan->rs_ctup.t_data == NULL)
1299         {
1300                 HEAPDEBUG_2;                    /* heap_getnext returning EOS */
1301                 return NULL;
1302         }
1303
1304         /*
1305          * if we get here it means we have a new current scan tuple, so point to
1306          * the proper return buffer and return the tuple.
1307          */
1308         HEAPDEBUG_3;                            /* heap_getnext returning tuple */
1309
1310         pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1311
1312         return &scan->rs_ctup;
1313 }
1314
1315 #ifdef HEAPAMSLOTDEBUGALL
1316 #define HEAPAMSLOTDEBUG_1 \
1317         elog(DEBUG2, "heapam_getnextslot([%s,nkeys=%d],dir=%d) called", \
1318                  RelationGetRelationName(scan->rs_base.rs_rd), scan->rs_base.rs_nkeys, (int) direction)
1319 #define HEAPAMSLOTDEBUG_2 \
1320         elog(DEBUG2, "heapam_getnextslot returning EOS")
1321 #define HEAPAMSLOTDEBUG_3 \
1322         elog(DEBUG2, "heapam_getnextslot returning tuple")
1323 #else
1324 #define HEAPAMSLOTDEBUG_1
1325 #define HEAPAMSLOTDEBUG_2
1326 #define HEAPAMSLOTDEBUG_3
1327 #endif
1328
1329 bool
1330 heap_getnextslot(TableScanDesc sscan, ScanDirection direction, TupleTableSlot *slot)
1331 {
1332         HeapScanDesc scan = (HeapScanDesc) sscan;
1333
1334         /* Note: no locking manipulations needed */
1335
1336         HEAPAMSLOTDEBUG_1;                      /* heap_getnextslot( info ) */
1337
1338         if (scan->rs_base.rs_pageatatime)
1339                 heapgettup_pagemode(scan, direction,
1340                                                         scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1341         else
1342                 heapgettup(scan, direction, scan->rs_base.rs_nkeys, scan->rs_base.rs_key);
1343
1344         if (scan->rs_ctup.t_data == NULL)
1345         {
1346                 HEAPAMSLOTDEBUG_2;              /* heap_getnextslot returning EOS */
1347                 ExecClearTuple(slot);
1348                 return false;
1349         }
1350
1351         /*
1352          * if we get here it means we have a new current scan tuple, so point to
1353          * the proper return buffer and return the tuple.
1354          */
1355         HEAPAMSLOTDEBUG_3;                      /* heap_getnextslot returning tuple */
1356
1357         pgstat_count_heap_getnext(scan->rs_base.rs_rd);
1358
1359         ExecStoreBufferHeapTuple(&scan->rs_ctup, slot,
1360                                                          scan->rs_cbuf);
1361         return true;
1362 }
1363
1364 /*
1365  *      heap_fetch              - retrieve tuple with given tid
1366  *
1367  * On entry, tuple->t_self is the TID to fetch.  We pin the buffer holding
1368  * the tuple, fill in the remaining fields of *tuple, and check the tuple
1369  * against the specified snapshot.
1370  *
1371  * If successful (tuple found and passes snapshot time qual), then *userbuf
1372  * is set to the buffer holding the tuple and true is returned.  The caller
1373  * must unpin the buffer when done with the tuple.
1374  *
1375  * If the tuple is not found (ie, item number references a deleted slot),
1376  * then tuple->t_data is set to NULL and false is returned.
1377  *
1378  * If the tuple is found but fails the time qual check, then false is returned
1379  * but tuple->t_data is left pointing to the tuple.
1380  *
1381  * keep_buf determines what is done with the buffer in the false-result cases.
1382  * When the caller specifies keep_buf = true, we retain the pin on the buffer
1383  * and return it in *userbuf (so the caller must eventually unpin it); when
1384  * keep_buf = false, the pin is released and *userbuf is set to InvalidBuffer.
1385  *
1386  * stats_relation is the relation to charge the heap_fetch operation against
1387  * for statistical purposes.  (This could be the heap rel itself, an
1388  * associated index, or NULL to not count the fetch at all.)
1389  *
1390  * heap_fetch does not follow HOT chains: only the exact TID requested will
1391  * be fetched.
1392  *
1393  * It is somewhat inconsistent that we ereport() on invalid block number but
1394  * return false on invalid item number.  There are a couple of reasons though.
1395  * One is that the caller can relatively easily check the block number for
1396  * validity, but cannot check the item number without reading the page
1397  * himself.  Another is that when we are following a t_ctid link, we can be
1398  * reasonably confident that the page number is valid (since VACUUM shouldn't
1399  * truncate off the destination page without having killed the referencing
1400  * tuple first), but the item number might well not be good.
1401  */
1402 bool
1403 heap_fetch(Relation relation,
1404                    Snapshot snapshot,
1405                    HeapTuple tuple,
1406                    Buffer *userbuf)
1407 {
1408         ItemPointer tid = &(tuple->t_self);
1409         ItemId          lp;
1410         Buffer          buffer;
1411         Page            page;
1412         OffsetNumber offnum;
1413         bool            valid;
1414
1415         /*
1416          * Fetch and pin the appropriate page of the relation.
1417          */
1418         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
1419
1420         /*
1421          * Need share lock on buffer to examine tuple commit status.
1422          */
1423         LockBuffer(buffer, BUFFER_LOCK_SHARE);
1424         page = BufferGetPage(buffer);
1425         TestForOldSnapshot(snapshot, relation, page);
1426
1427         /*
1428          * We'd better check for out-of-range offnum in case of VACUUM since the
1429          * TID was obtained.
1430          */
1431         offnum = ItemPointerGetOffsetNumber(tid);
1432         if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1433         {
1434                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1435                 ReleaseBuffer(buffer);
1436                 *userbuf = InvalidBuffer;
1437                 tuple->t_data = NULL;
1438                 return false;
1439         }
1440
1441         /*
1442          * get the item line pointer corresponding to the requested tid
1443          */
1444         lp = PageGetItemId(page, offnum);
1445
1446         /*
1447          * Must check for deleted tuple.
1448          */
1449         if (!ItemIdIsNormal(lp))
1450         {
1451                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1452                 ReleaseBuffer(buffer);
1453                 *userbuf = InvalidBuffer;
1454                 tuple->t_data = NULL;
1455                 return false;
1456         }
1457
1458         /*
1459          * fill in *tuple fields
1460          */
1461         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
1462         tuple->t_len = ItemIdGetLength(lp);
1463         tuple->t_tableOid = RelationGetRelid(relation);
1464
1465         /*
1466          * check tuple visibility, then release lock
1467          */
1468         valid = HeapTupleSatisfiesVisibility(tuple, snapshot, buffer);
1469
1470         if (valid)
1471                 PredicateLockTuple(relation, tuple, snapshot);
1472
1473         CheckForSerializableConflictOut(valid, relation, tuple, buffer, snapshot);
1474
1475         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
1476
1477         if (valid)
1478         {
1479                 /*
1480                  * All checks passed, so return the tuple as valid. Caller is now
1481                  * responsible for releasing the buffer.
1482                  */
1483                 *userbuf = buffer;
1484
1485                 return true;
1486         }
1487
1488         /* Tuple failed time qual */
1489         ReleaseBuffer(buffer);
1490         *userbuf = InvalidBuffer;
1491
1492         return false;
1493 }
1494
1495 /*
1496  *      heap_hot_search_buffer  - search HOT chain for tuple satisfying snapshot
1497  *
1498  * On entry, *tid is the TID of a tuple (either a simple tuple, or the root
1499  * of a HOT chain), and buffer is the buffer holding this tuple.  We search
1500  * for the first chain member satisfying the given snapshot.  If one is
1501  * found, we update *tid to reference that tuple's offset number, and
1502  * return true.  If no match, return false without modifying *tid.
1503  *
1504  * heapTuple is a caller-supplied buffer.  When a match is found, we return
1505  * the tuple here, in addition to updating *tid.  If no match is found, the
1506  * contents of this buffer on return are undefined.
1507  *
1508  * If all_dead is not NULL, we check non-visible tuples to see if they are
1509  * globally dead; *all_dead is set true if all members of the HOT chain
1510  * are vacuumable, false if not.
1511  *
1512  * Unlike heap_fetch, the caller must already have pin and (at least) share
1513  * lock on the buffer; it is still pinned/locked at exit.  Also unlike
1514  * heap_fetch, we do not report any pgstats count; caller may do so if wanted.
1515  */
1516 bool
1517 heap_hot_search_buffer(ItemPointer tid, Relation relation, Buffer buffer,
1518                                            Snapshot snapshot, HeapTuple heapTuple,
1519                                            bool *all_dead, bool first_call)
1520 {
1521         Page            dp = (Page) BufferGetPage(buffer);
1522         TransactionId prev_xmax = InvalidTransactionId;
1523         OffsetNumber offnum;
1524         bool            at_chain_start;
1525         bool            valid;
1526         bool            skip;
1527
1528         /* If this is not the first call, previous call returned a (live!) tuple */
1529         if (all_dead)
1530                 *all_dead = first_call;
1531
1532         Assert(TransactionIdIsValid(RecentGlobalXmin));
1533
1534         Assert(ItemPointerGetBlockNumber(tid) == BufferGetBlockNumber(buffer));
1535         offnum = ItemPointerGetOffsetNumber(tid);
1536         at_chain_start = first_call;
1537         skip = !first_call;
1538
1539         heapTuple->t_self = *tid;
1540
1541         /* Scan through possible multiple members of HOT-chain */
1542         for (;;)
1543         {
1544                 ItemId          lp;
1545
1546                 /* check for bogus TID */
1547                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(dp))
1548                         break;
1549
1550                 lp = PageGetItemId(dp, offnum);
1551
1552                 /* check for unused, dead, or redirected items */
1553                 if (!ItemIdIsNormal(lp))
1554                 {
1555                         /* We should only see a redirect at start of chain */
1556                         if (ItemIdIsRedirected(lp) && at_chain_start)
1557                         {
1558                                 /* Follow the redirect */
1559                                 offnum = ItemIdGetRedirect(lp);
1560                                 at_chain_start = false;
1561                                 continue;
1562                         }
1563                         /* else must be end of chain */
1564                         break;
1565                 }
1566
1567                 heapTuple->t_data = (HeapTupleHeader) PageGetItem(dp, lp);
1568                 heapTuple->t_len = ItemIdGetLength(lp);
1569                 heapTuple->t_tableOid = RelationGetRelid(relation);
1570                 ItemPointerSetOffsetNumber(&heapTuple->t_self, offnum);
1571
1572                 /*
1573                  * Shouldn't see a HEAP_ONLY tuple at chain start.
1574                  */
1575                 if (at_chain_start && HeapTupleIsHeapOnly(heapTuple))
1576                         break;
1577
1578                 /*
1579                  * The xmin should match the previous xmax value, else chain is
1580                  * broken.
1581                  */
1582                 if (TransactionIdIsValid(prev_xmax) &&
1583                         !TransactionIdEquals(prev_xmax,
1584                                                                  HeapTupleHeaderGetXmin(heapTuple->t_data)))
1585                         break;
1586
1587                 /*
1588                  * When first_call is true (and thus, skip is initially false) we'll
1589                  * return the first tuple we find.  But on later passes, heapTuple
1590                  * will initially be pointing to the tuple we returned last time.
1591                  * Returning it again would be incorrect (and would loop forever), so
1592                  * we skip it and return the next match we find.
1593                  */
1594                 if (!skip)
1595                 {
1596                         /*
1597                          * For the benefit of logical decoding, have t_self point at the
1598                          * element of the HOT chain we're currently investigating instead
1599                          * of the root tuple of the HOT chain. This is important because
1600                          * the *Satisfies routine for historical mvcc snapshots needs the
1601                          * correct tid to decide about the visibility in some cases.
1602                          */
1603                         ItemPointerSet(&(heapTuple->t_self), BufferGetBlockNumber(buffer), offnum);
1604
1605                         /* If it's visible per the snapshot, we must return it */
1606                         valid = HeapTupleSatisfiesVisibility(heapTuple, snapshot, buffer);
1607                         CheckForSerializableConflictOut(valid, relation, heapTuple,
1608                                                                                         buffer, snapshot);
1609                         /* reset to original, non-redirected, tid */
1610                         heapTuple->t_self = *tid;
1611
1612                         if (valid)
1613                         {
1614                                 ItemPointerSetOffsetNumber(tid, offnum);
1615                                 PredicateLockTuple(relation, heapTuple, snapshot);
1616                                 if (all_dead)
1617                                         *all_dead = false;
1618                                 return true;
1619                         }
1620                 }
1621                 skip = false;
1622
1623                 /*
1624                  * If we can't see it, maybe no one else can either.  At caller
1625                  * request, check whether all chain members are dead to all
1626                  * transactions.
1627                  *
1628                  * Note: if you change the criterion here for what is "dead", fix the
1629                  * planner's get_actual_variable_range() function to match.
1630                  */
1631                 if (all_dead && *all_dead &&
1632                         !HeapTupleIsSurelyDead(heapTuple, RecentGlobalXmin))
1633                         *all_dead = false;
1634
1635                 /*
1636                  * Check to see if HOT chain continues past this tuple; if so fetch
1637                  * the next offnum and loop around.
1638                  */
1639                 if (HeapTupleIsHotUpdated(heapTuple))
1640                 {
1641                         Assert(ItemPointerGetBlockNumber(&heapTuple->t_data->t_ctid) ==
1642                                    ItemPointerGetBlockNumber(tid));
1643                         offnum = ItemPointerGetOffsetNumber(&heapTuple->t_data->t_ctid);
1644                         at_chain_start = false;
1645                         prev_xmax = HeapTupleHeaderGetUpdateXid(heapTuple->t_data);
1646                 }
1647                 else
1648                         break;                          /* end of chain */
1649         }
1650
1651         return false;
1652 }
1653
1654 /*
1655  *      heap_get_latest_tid -  get the latest tid of a specified tuple
1656  *
1657  * Actually, this gets the latest version that is visible according to
1658  * the passed snapshot.  You can pass SnapshotDirty to get the very latest,
1659  * possibly uncommitted version.
1660  *
1661  * *tid is both an input and an output parameter: it is updated to
1662  * show the latest version of the row.  Note that it will not be changed
1663  * if no version of the row passes the snapshot test.
1664  */
1665 void
1666 heap_get_latest_tid(Relation relation,
1667                                         Snapshot snapshot,
1668                                         ItemPointer tid)
1669 {
1670         BlockNumber blk;
1671         ItemPointerData ctid;
1672         TransactionId priorXmax;
1673
1674         /* this is to avoid Assert failures on bad input */
1675         if (!ItemPointerIsValid(tid))
1676                 return;
1677
1678         /*
1679          * Since this can be called with user-supplied TID, don't trust the input
1680          * too much.  (RelationGetNumberOfBlocks is an expensive check, so we
1681          * don't check t_ctid links again this way.  Note that it would not do to
1682          * call it just once and save the result, either.)
1683          */
1684         blk = ItemPointerGetBlockNumber(tid);
1685         if (blk >= RelationGetNumberOfBlocks(relation))
1686                 elog(ERROR, "block number %u is out of range for relation \"%s\"",
1687                          blk, RelationGetRelationName(relation));
1688
1689         /*
1690          * Loop to chase down t_ctid links.  At top of loop, ctid is the tuple we
1691          * need to examine, and *tid is the TID we will return if ctid turns out
1692          * to be bogus.
1693          *
1694          * Note that we will loop until we reach the end of the t_ctid chain.
1695          * Depending on the snapshot passed, there might be at most one visible
1696          * version of the row, but we don't try to optimize for that.
1697          */
1698         ctid = *tid;
1699         priorXmax = InvalidTransactionId;       /* cannot check first XMIN */
1700         for (;;)
1701         {
1702                 Buffer          buffer;
1703                 Page            page;
1704                 OffsetNumber offnum;
1705                 ItemId          lp;
1706                 HeapTupleData tp;
1707                 bool            valid;
1708
1709                 /*
1710                  * Read, pin, and lock the page.
1711                  */
1712                 buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&ctid));
1713                 LockBuffer(buffer, BUFFER_LOCK_SHARE);
1714                 page = BufferGetPage(buffer);
1715                 TestForOldSnapshot(snapshot, relation, page);
1716
1717                 /*
1718                  * Check for bogus item number.  This is not treated as an error
1719                  * condition because it can happen while following a t_ctid link. We
1720                  * just assume that the prior tid is OK and return it unchanged.
1721                  */
1722                 offnum = ItemPointerGetOffsetNumber(&ctid);
1723                 if (offnum < FirstOffsetNumber || offnum > PageGetMaxOffsetNumber(page))
1724                 {
1725                         UnlockReleaseBuffer(buffer);
1726                         break;
1727                 }
1728                 lp = PageGetItemId(page, offnum);
1729                 if (!ItemIdIsNormal(lp))
1730                 {
1731                         UnlockReleaseBuffer(buffer);
1732                         break;
1733                 }
1734
1735                 /* OK to access the tuple */
1736                 tp.t_self = ctid;
1737                 tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
1738                 tp.t_len = ItemIdGetLength(lp);
1739                 tp.t_tableOid = RelationGetRelid(relation);
1740
1741                 /*
1742                  * After following a t_ctid link, we might arrive at an unrelated
1743                  * tuple.  Check for XMIN match.
1744                  */
1745                 if (TransactionIdIsValid(priorXmax) &&
1746                         !TransactionIdEquals(priorXmax, HeapTupleHeaderGetXmin(tp.t_data)))
1747                 {
1748                         UnlockReleaseBuffer(buffer);
1749                         break;
1750                 }
1751
1752                 /*
1753                  * Check tuple visibility; if visible, set it as the new result
1754                  * candidate.
1755                  */
1756                 valid = HeapTupleSatisfiesVisibility(&tp, snapshot, buffer);
1757                 CheckForSerializableConflictOut(valid, relation, &tp, buffer, snapshot);
1758                 if (valid)
1759                         *tid = ctid;
1760
1761                 /*
1762                  * If there's a valid t_ctid link, follow it, else we're done.
1763                  */
1764                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
1765                         HeapTupleHeaderIsOnlyLocked(tp.t_data) ||
1766                         HeapTupleHeaderIndicatesMovedPartitions(tp.t_data) ||
1767                         ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid))
1768                 {
1769                         UnlockReleaseBuffer(buffer);
1770                         break;
1771                 }
1772
1773                 ctid = tp.t_data->t_ctid;
1774                 priorXmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
1775                 UnlockReleaseBuffer(buffer);
1776         }                                                       /* end of loop */
1777 }
1778
1779
1780 /*
1781  * UpdateXmaxHintBits - update tuple hint bits after xmax transaction ends
1782  *
1783  * This is called after we have waited for the XMAX transaction to terminate.
1784  * If the transaction aborted, we guarantee the XMAX_INVALID hint bit will
1785  * be set on exit.  If the transaction committed, we set the XMAX_COMMITTED
1786  * hint bit if possible --- but beware that that may not yet be possible,
1787  * if the transaction committed asynchronously.
1788  *
1789  * Note that if the transaction was a locker only, we set HEAP_XMAX_INVALID
1790  * even if it commits.
1791  *
1792  * Hence callers should look only at XMAX_INVALID.
1793  *
1794  * Note this is not allowed for tuples whose xmax is a multixact.
1795  */
1796 static void
1797 UpdateXmaxHintBits(HeapTupleHeader tuple, Buffer buffer, TransactionId xid)
1798 {
1799         Assert(TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple), xid));
1800         Assert(!(tuple->t_infomask & HEAP_XMAX_IS_MULTI));
1801
1802         if (!(tuple->t_infomask & (HEAP_XMAX_COMMITTED | HEAP_XMAX_INVALID)))
1803         {
1804                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
1805                         TransactionIdDidCommit(xid))
1806                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_COMMITTED,
1807                                                                  xid);
1808                 else
1809                         HeapTupleSetHintBits(tuple, buffer, HEAP_XMAX_INVALID,
1810                                                                  InvalidTransactionId);
1811         }
1812 }
1813
1814
1815 /*
1816  * GetBulkInsertState - prepare status object for a bulk insert
1817  */
1818 BulkInsertState
1819 GetBulkInsertState(void)
1820 {
1821         BulkInsertState bistate;
1822
1823         bistate = (BulkInsertState) palloc(sizeof(BulkInsertStateData));
1824         bistate->strategy = GetAccessStrategy(BAS_BULKWRITE);
1825         bistate->current_buf = InvalidBuffer;
1826         return bistate;
1827 }
1828
1829 /*
1830  * FreeBulkInsertState - clean up after finishing a bulk insert
1831  */
1832 void
1833 FreeBulkInsertState(BulkInsertState bistate)
1834 {
1835         if (bistate->current_buf != InvalidBuffer)
1836                 ReleaseBuffer(bistate->current_buf);
1837         FreeAccessStrategy(bistate->strategy);
1838         pfree(bistate);
1839 }
1840
1841 /*
1842  * ReleaseBulkInsertStatePin - release a buffer currently held in bistate
1843  */
1844 void
1845 ReleaseBulkInsertStatePin(BulkInsertState bistate)
1846 {
1847         if (bistate->current_buf != InvalidBuffer)
1848                 ReleaseBuffer(bistate->current_buf);
1849         bistate->current_buf = InvalidBuffer;
1850 }
1851
1852
1853 /*
1854  *      heap_insert             - insert tuple into a heap
1855  *
1856  * The new tuple is stamped with current transaction ID and the specified
1857  * command ID.
1858  *
1859  * See table_insert for comments about most of the input flags, except that
1860  * this routine directly takes a tuple rather than a slot.
1861  *
1862  * There's corresponding HEAP_INSERT_ options to all the TABLE_INSERT_
1863  * options, and there additionally is HEAP_INSERT_SPECULATIVE which is used to
1864  * implement table_insert_speculative().
1865  *
1866  * On return the header fields of *tup are updated to match the stored tuple;
1867  * in particular tup->t_self receives the actual TID where the tuple was
1868  * stored.  But note that any toasting of fields within the tuple data is NOT
1869  * reflected into *tup.
1870  */
1871 void
1872 heap_insert(Relation relation, HeapTuple tup, CommandId cid,
1873                         int options, BulkInsertState bistate)
1874 {
1875         TransactionId xid = GetCurrentTransactionId();
1876         HeapTuple       heaptup;
1877         Buffer          buffer;
1878         Buffer          vmbuffer = InvalidBuffer;
1879         bool            all_visible_cleared = false;
1880
1881         /*
1882          * Fill in tuple header fields and toast the tuple if necessary.
1883          *
1884          * Note: below this point, heaptup is the data we actually intend to store
1885          * into the relation; tup is the caller's original untoasted data.
1886          */
1887         heaptup = heap_prepare_insert(relation, tup, xid, cid, options);
1888
1889         /*
1890          * Find buffer to insert this tuple into.  If the page is all visible,
1891          * this will also pin the requisite visibility map page.
1892          */
1893         buffer = RelationGetBufferForTuple(relation, heaptup->t_len,
1894                                                                            InvalidBuffer, options, bistate,
1895                                                                            &vmbuffer, NULL);
1896
1897         /*
1898          * We're about to do the actual insert -- but check for conflict first, to
1899          * avoid possibly having to roll back work we've just done.
1900          *
1901          * This is safe without a recheck as long as there is no possibility of
1902          * another process scanning the page between this check and the insert
1903          * being visible to the scan (i.e., an exclusive buffer content lock is
1904          * continuously held from this point until the tuple insert is visible).
1905          *
1906          * For a heap insert, we only need to check for table-level SSI locks. Our
1907          * new tuple can't possibly conflict with existing tuple locks, and heap
1908          * page locks are only consolidated versions of tuple locks; they do not
1909          * lock "gaps" as index page locks do.  So we don't need to specify a
1910          * buffer when making the call, which makes for a faster check.
1911          */
1912         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
1913
1914         /* NO EREPORT(ERROR) from here till changes are logged */
1915         START_CRIT_SECTION();
1916
1917         RelationPutHeapTuple(relation, buffer, heaptup,
1918                                                  (options & HEAP_INSERT_SPECULATIVE) != 0);
1919
1920         if (PageIsAllVisible(BufferGetPage(buffer)))
1921         {
1922                 all_visible_cleared = true;
1923                 PageClearAllVisible(BufferGetPage(buffer));
1924                 visibilitymap_clear(relation,
1925                                                         ItemPointerGetBlockNumber(&(heaptup->t_self)),
1926                                                         vmbuffer, VISIBILITYMAP_VALID_BITS);
1927         }
1928
1929         /*
1930          * XXX Should we set PageSetPrunable on this page ?
1931          *
1932          * The inserting transaction may eventually abort thus making this tuple
1933          * DEAD and hence available for pruning. Though we don't want to optimize
1934          * for aborts, if no other tuple in this page is UPDATEd/DELETEd, the
1935          * aborted tuple will never be pruned until next vacuum is triggered.
1936          *
1937          * If you do add PageSetPrunable here, add it in heap_xlog_insert too.
1938          */
1939
1940         MarkBufferDirty(buffer);
1941
1942         /* XLOG stuff */
1943         if (!(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation))
1944         {
1945                 xl_heap_insert xlrec;
1946                 xl_heap_header xlhdr;
1947                 XLogRecPtr      recptr;
1948                 Page            page = BufferGetPage(buffer);
1949                 uint8           info = XLOG_HEAP_INSERT;
1950                 int                     bufflags = 0;
1951
1952                 /*
1953                  * If this is a catalog, we need to transmit combocids to properly
1954                  * decode, so log that as well.
1955                  */
1956                 if (RelationIsAccessibleInLogicalDecoding(relation))
1957                         log_heap_new_cid(relation, heaptup);
1958
1959                 /*
1960                  * If this is the single and first tuple on page, we can reinit the
1961                  * page instead of restoring the whole thing.  Set flag, and hide
1962                  * buffer references from XLogInsert.
1963                  */
1964                 if (ItemPointerGetOffsetNumber(&(heaptup->t_self)) == FirstOffsetNumber &&
1965                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
1966                 {
1967                         info |= XLOG_HEAP_INIT_PAGE;
1968                         bufflags |= REGBUF_WILL_INIT;
1969                 }
1970
1971                 xlrec.offnum = ItemPointerGetOffsetNumber(&heaptup->t_self);
1972                 xlrec.flags = 0;
1973                 if (all_visible_cleared)
1974                         xlrec.flags |= XLH_INSERT_ALL_VISIBLE_CLEARED;
1975                 if (options & HEAP_INSERT_SPECULATIVE)
1976                         xlrec.flags |= XLH_INSERT_IS_SPECULATIVE;
1977                 Assert(ItemPointerGetBlockNumber(&heaptup->t_self) == BufferGetBlockNumber(buffer));
1978
1979                 /*
1980                  * For logical decoding, we need the tuple even if we're doing a full
1981                  * page write, so make sure it's included even if we take a full-page
1982                  * image. (XXX We could alternatively store a pointer into the FPW).
1983                  */
1984                 if (RelationIsLogicallyLogged(relation) &&
1985                         !(options & HEAP_INSERT_NO_LOGICAL))
1986                 {
1987                         xlrec.flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
1988                         bufflags |= REGBUF_KEEP_DATA;
1989                 }
1990
1991                 XLogBeginInsert();
1992                 XLogRegisterData((char *) &xlrec, SizeOfHeapInsert);
1993
1994                 xlhdr.t_infomask2 = heaptup->t_data->t_infomask2;
1995                 xlhdr.t_infomask = heaptup->t_data->t_infomask;
1996                 xlhdr.t_hoff = heaptup->t_data->t_hoff;
1997
1998                 /*
1999                  * note we mark xlhdr as belonging to buffer; if XLogInsert decides to
2000                  * write the whole page to the xlog, we don't need to store
2001                  * xl_heap_header in the xlog.
2002                  */
2003                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2004                 XLogRegisterBufData(0, (char *) &xlhdr, SizeOfHeapHeader);
2005                 /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
2006                 XLogRegisterBufData(0,
2007                                                         (char *) heaptup->t_data + SizeofHeapTupleHeader,
2008                                                         heaptup->t_len - SizeofHeapTupleHeader);
2009
2010                 /* filtering by origin on a row level is much more efficient */
2011                 XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2012
2013                 recptr = XLogInsert(RM_HEAP_ID, info);
2014
2015                 PageSetLSN(page, recptr);
2016         }
2017
2018         END_CRIT_SECTION();
2019
2020         UnlockReleaseBuffer(buffer);
2021         if (vmbuffer != InvalidBuffer)
2022                 ReleaseBuffer(vmbuffer);
2023
2024         /*
2025          * If tuple is cachable, mark it for invalidation from the caches in case
2026          * we abort.  Note it is OK to do this after releasing the buffer, because
2027          * the heaptup data structure is all in local memory, not in the shared
2028          * buffer.
2029          */
2030         CacheInvalidateHeapTuple(relation, heaptup, NULL);
2031
2032         /* Note: speculative insertions are counted too, even if aborted later */
2033         pgstat_count_heap_insert(relation, 1);
2034
2035         /*
2036          * If heaptup is a private copy, release it.  Don't forget to copy t_self
2037          * back to the caller's image, too.
2038          */
2039         if (heaptup != tup)
2040         {
2041                 tup->t_self = heaptup->t_self;
2042                 heap_freetuple(heaptup);
2043         }
2044 }
2045
2046 /*
2047  * Subroutine for heap_insert(). Prepares a tuple for insertion. This sets the
2048  * tuple header fields and toasts the tuple if necessary.  Returns a toasted
2049  * version of the tuple if it was toasted, or the original tuple if not. Note
2050  * that in any case, the header fields are also set in the original tuple.
2051  */
2052 static HeapTuple
2053 heap_prepare_insert(Relation relation, HeapTuple tup, TransactionId xid,
2054                                         CommandId cid, int options)
2055 {
2056         /*
2057          * Parallel operations are required to be strictly read-only in a parallel
2058          * worker.  Parallel inserts are not safe even in the leader in the
2059          * general case, because group locking means that heavyweight locks for
2060          * relation extension or GIN page locks will not conflict between members
2061          * of a lock group, but we don't prohibit that case here because there are
2062          * useful special cases that we can safely allow, such as CREATE TABLE AS.
2063          */
2064         if (IsParallelWorker())
2065                 ereport(ERROR,
2066                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2067                                  errmsg("cannot insert tuples in a parallel worker")));
2068
2069         tup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
2070         tup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
2071         tup->t_data->t_infomask |= HEAP_XMAX_INVALID;
2072         HeapTupleHeaderSetXmin(tup->t_data, xid);
2073         if (options & HEAP_INSERT_FROZEN)
2074                 HeapTupleHeaderSetXminFrozen(tup->t_data);
2075
2076         HeapTupleHeaderSetCmin(tup->t_data, cid);
2077         HeapTupleHeaderSetXmax(tup->t_data, 0); /* for cleanliness */
2078         tup->t_tableOid = RelationGetRelid(relation);
2079
2080         /*
2081          * If the new tuple is too big for storage or contains already toasted
2082          * out-of-line attributes from some other relation, invoke the toaster.
2083          */
2084         if (relation->rd_rel->relkind != RELKIND_RELATION &&
2085                 relation->rd_rel->relkind != RELKIND_MATVIEW)
2086         {
2087                 /* toast table entries should never be recursively toasted */
2088                 Assert(!HeapTupleHasExternal(tup));
2089                 return tup;
2090         }
2091         else if (HeapTupleHasExternal(tup) || tup->t_len > TOAST_TUPLE_THRESHOLD)
2092                 return toast_insert_or_update(relation, tup, NULL, options);
2093         else
2094                 return tup;
2095 }
2096
2097 /*
2098  *      heap_multi_insert       - insert multiple tuple into a heap
2099  *
2100  * This is like heap_insert(), but inserts multiple tuples in one operation.
2101  * That's faster than calling heap_insert() in a loop, because when multiple
2102  * tuples can be inserted on a single page, we can write just a single WAL
2103  * record covering all of them, and only need to lock/unlock the page once.
2104  *
2105  * Note: this leaks memory into the current memory context. You can create a
2106  * temporary context before calling this, if that's a problem.
2107  */
2108 void
2109 heap_multi_insert(Relation relation, TupleTableSlot **slots, int ntuples,
2110                                   CommandId cid, int options, BulkInsertState bistate)
2111 {
2112         TransactionId xid = GetCurrentTransactionId();
2113         HeapTuple  *heaptuples;
2114         int                     i;
2115         int                     ndone;
2116         PGAlignedBlock scratch;
2117         Page            page;
2118         bool            needwal;
2119         Size            saveFreeSpace;
2120         bool            need_tuple_data = RelationIsLogicallyLogged(relation);
2121         bool            need_cids = RelationIsAccessibleInLogicalDecoding(relation);
2122
2123         /* currently not needed (thus unsupported) for heap_multi_insert() */
2124         AssertArg(!(options & HEAP_INSERT_NO_LOGICAL));
2125
2126         needwal = !(options & HEAP_INSERT_SKIP_WAL) && RelationNeedsWAL(relation);
2127         saveFreeSpace = RelationGetTargetPageFreeSpace(relation,
2128                                                                                                    HEAP_DEFAULT_FILLFACTOR);
2129
2130         /* Toast and set header data in all the slots */
2131         heaptuples = palloc(ntuples * sizeof(HeapTuple));
2132         for (i = 0; i < ntuples; i++)
2133         {
2134                 HeapTuple       tuple;
2135
2136                 tuple = ExecFetchSlotHeapTuple(slots[i], true, NULL);
2137                 slots[i]->tts_tableOid = RelationGetRelid(relation);
2138                 tuple->t_tableOid = slots[i]->tts_tableOid;
2139                 heaptuples[i] = heap_prepare_insert(relation, tuple, xid, cid,
2140                                                                                         options);
2141         }
2142
2143         /*
2144          * We're about to do the actual inserts -- but check for conflict first,
2145          * to minimize the possibility of having to roll back work we've just
2146          * done.
2147          *
2148          * A check here does not definitively prevent a serialization anomaly;
2149          * that check MUST be done at least past the point of acquiring an
2150          * exclusive buffer content lock on every buffer that will be affected,
2151          * and MAY be done after all inserts are reflected in the buffers and
2152          * those locks are released; otherwise there race condition.  Since
2153          * multiple buffers can be locked and unlocked in the loop below, and it
2154          * would not be feasible to identify and lock all of those buffers before
2155          * the loop, we must do a final check at the end.
2156          *
2157          * The check here could be omitted with no loss of correctness; it is
2158          * present strictly as an optimization.
2159          *
2160          * For heap inserts, we only need to check for table-level SSI locks. Our
2161          * new tuples can't possibly conflict with existing tuple locks, and heap
2162          * page locks are only consolidated versions of tuple locks; they do not
2163          * lock "gaps" as index page locks do.  So we don't need to specify a
2164          * buffer when making the call, which makes for a faster check.
2165          */
2166         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2167
2168         ndone = 0;
2169         while (ndone < ntuples)
2170         {
2171                 Buffer          buffer;
2172                 Buffer          vmbuffer = InvalidBuffer;
2173                 bool            all_visible_cleared = false;
2174                 int                     nthispage;
2175
2176                 CHECK_FOR_INTERRUPTS();
2177
2178                 /*
2179                  * Find buffer where at least the next tuple will fit.  If the page is
2180                  * all-visible, this will also pin the requisite visibility map page.
2181                  */
2182                 buffer = RelationGetBufferForTuple(relation, heaptuples[ndone]->t_len,
2183                                                                                    InvalidBuffer, options, bistate,
2184                                                                                    &vmbuffer, NULL);
2185                 page = BufferGetPage(buffer);
2186
2187                 /* NO EREPORT(ERROR) from here till changes are logged */
2188                 START_CRIT_SECTION();
2189
2190                 /*
2191                  * RelationGetBufferForTuple has ensured that the first tuple fits.
2192                  * Put that on the page, and then as many other tuples as fit.
2193                  */
2194                 RelationPutHeapTuple(relation, buffer, heaptuples[ndone], false);
2195                 for (nthispage = 1; ndone + nthispage < ntuples; nthispage++)
2196                 {
2197                         HeapTuple       heaptup = heaptuples[ndone + nthispage];
2198
2199                         if (PageGetHeapFreeSpace(page) < MAXALIGN(heaptup->t_len) + saveFreeSpace)
2200                                 break;
2201
2202                         RelationPutHeapTuple(relation, buffer, heaptup, false);
2203
2204                         /*
2205                          * We don't use heap_multi_insert for catalog tuples yet, but
2206                          * better be prepared...
2207                          */
2208                         if (needwal && need_cids)
2209                                 log_heap_new_cid(relation, heaptup);
2210                 }
2211
2212                 if (PageIsAllVisible(page))
2213                 {
2214                         all_visible_cleared = true;
2215                         PageClearAllVisible(page);
2216                         visibilitymap_clear(relation,
2217                                                                 BufferGetBlockNumber(buffer),
2218                                                                 vmbuffer, VISIBILITYMAP_VALID_BITS);
2219                 }
2220
2221                 /*
2222                  * XXX Should we set PageSetPrunable on this page ? See heap_insert()
2223                  */
2224
2225                 MarkBufferDirty(buffer);
2226
2227                 /* XLOG stuff */
2228                 if (needwal)
2229                 {
2230                         XLogRecPtr      recptr;
2231                         xl_heap_multi_insert *xlrec;
2232                         uint8           info = XLOG_HEAP2_MULTI_INSERT;
2233                         char       *tupledata;
2234                         int                     totaldatalen;
2235                         char       *scratchptr = scratch.data;
2236                         bool            init;
2237                         int                     bufflags = 0;
2238
2239                         /*
2240                          * If the page was previously empty, we can reinit the page
2241                          * instead of restoring the whole thing.
2242                          */
2243                         init = (ItemPointerGetOffsetNumber(&(heaptuples[ndone]->t_self)) == FirstOffsetNumber &&
2244                                         PageGetMaxOffsetNumber(page) == FirstOffsetNumber + nthispage - 1);
2245
2246                         /* allocate xl_heap_multi_insert struct from the scratch area */
2247                         xlrec = (xl_heap_multi_insert *) scratchptr;
2248                         scratchptr += SizeOfHeapMultiInsert;
2249
2250                         /*
2251                          * Allocate offsets array. Unless we're reinitializing the page,
2252                          * in that case the tuples are stored in order starting at
2253                          * FirstOffsetNumber and we don't need to store the offsets
2254                          * explicitly.
2255                          */
2256                         if (!init)
2257                                 scratchptr += nthispage * sizeof(OffsetNumber);
2258
2259                         /* the rest of the scratch space is used for tuple data */
2260                         tupledata = scratchptr;
2261
2262                         xlrec->flags = all_visible_cleared ? XLH_INSERT_ALL_VISIBLE_CLEARED : 0;
2263                         xlrec->ntuples = nthispage;
2264
2265                         /*
2266                          * Write out an xl_multi_insert_tuple and the tuple data itself
2267                          * for each tuple.
2268                          */
2269                         for (i = 0; i < nthispage; i++)
2270                         {
2271                                 HeapTuple       heaptup = heaptuples[ndone + i];
2272                                 xl_multi_insert_tuple *tuphdr;
2273                                 int                     datalen;
2274
2275                                 if (!init)
2276                                         xlrec->offsets[i] = ItemPointerGetOffsetNumber(&heaptup->t_self);
2277                                 /* xl_multi_insert_tuple needs two-byte alignment. */
2278                                 tuphdr = (xl_multi_insert_tuple *) SHORTALIGN(scratchptr);
2279                                 scratchptr = ((char *) tuphdr) + SizeOfMultiInsertTuple;
2280
2281                                 tuphdr->t_infomask2 = heaptup->t_data->t_infomask2;
2282                                 tuphdr->t_infomask = heaptup->t_data->t_infomask;
2283                                 tuphdr->t_hoff = heaptup->t_data->t_hoff;
2284
2285                                 /* write bitmap [+ padding] [+ oid] + data */
2286                                 datalen = heaptup->t_len - SizeofHeapTupleHeader;
2287                                 memcpy(scratchptr,
2288                                            (char *) heaptup->t_data + SizeofHeapTupleHeader,
2289                                            datalen);
2290                                 tuphdr->datalen = datalen;
2291                                 scratchptr += datalen;
2292                         }
2293                         totaldatalen = scratchptr - tupledata;
2294                         Assert((scratchptr - scratch.data) < BLCKSZ);
2295
2296                         if (need_tuple_data)
2297                                 xlrec->flags |= XLH_INSERT_CONTAINS_NEW_TUPLE;
2298
2299                         /*
2300                          * Signal that this is the last xl_heap_multi_insert record
2301                          * emitted by this call to heap_multi_insert(). Needed for logical
2302                          * decoding so it knows when to cleanup temporary data.
2303                          */
2304                         if (ndone + nthispage == ntuples)
2305                                 xlrec->flags |= XLH_INSERT_LAST_IN_MULTI;
2306
2307                         if (init)
2308                         {
2309                                 info |= XLOG_HEAP_INIT_PAGE;
2310                                 bufflags |= REGBUF_WILL_INIT;
2311                         }
2312
2313                         /*
2314                          * If we're doing logical decoding, include the new tuple data
2315                          * even if we take a full-page image of the page.
2316                          */
2317                         if (need_tuple_data)
2318                                 bufflags |= REGBUF_KEEP_DATA;
2319
2320                         XLogBeginInsert();
2321                         XLogRegisterData((char *) xlrec, tupledata - scratch.data);
2322                         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD | bufflags);
2323
2324                         XLogRegisterBufData(0, tupledata, totaldatalen);
2325
2326                         /* filtering by origin on a row level is much more efficient */
2327                         XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2328
2329                         recptr = XLogInsert(RM_HEAP2_ID, info);
2330
2331                         PageSetLSN(page, recptr);
2332                 }
2333
2334                 END_CRIT_SECTION();
2335
2336                 UnlockReleaseBuffer(buffer);
2337                 if (vmbuffer != InvalidBuffer)
2338                         ReleaseBuffer(vmbuffer);
2339
2340                 ndone += nthispage;
2341         }
2342
2343         /*
2344          * We're done with the actual inserts.  Check for conflicts again, to
2345          * ensure that all rw-conflicts in to these inserts are detected.  Without
2346          * this final check, a sequential scan of the heap may have locked the
2347          * table after the "before" check, missing one opportunity to detect the
2348          * conflict, and then scanned the table before the new tuples were there,
2349          * missing the other chance to detect the conflict.
2350          *
2351          * For heap inserts, we only need to check for table-level SSI locks. Our
2352          * new tuples can't possibly conflict with existing tuple locks, and heap
2353          * page locks are only consolidated versions of tuple locks; they do not
2354          * lock "gaps" as index page locks do.  So we don't need to specify a
2355          * buffer when making the call.
2356          */
2357         CheckForSerializableConflictIn(relation, NULL, InvalidBuffer);
2358
2359         /*
2360          * If tuples are cachable, mark them for invalidation from the caches in
2361          * case we abort.  Note it is OK to do this after releasing the buffer,
2362          * because the heaptuples data structure is all in local memory, not in
2363          * the shared buffer.
2364          */
2365         if (IsCatalogRelation(relation))
2366         {
2367                 for (i = 0; i < ntuples; i++)
2368                         CacheInvalidateHeapTuple(relation, heaptuples[i], NULL);
2369         }
2370
2371         /* copy t_self fields back to the caller's slots */
2372         for (i = 0; i < ntuples; i++)
2373                 slots[i]->tts_tid = heaptuples[i]->t_self;
2374
2375         pgstat_count_heap_insert(relation, ntuples);
2376 }
2377
2378 /*
2379  *      simple_heap_insert - insert a tuple
2380  *
2381  * Currently, this routine differs from heap_insert only in supplying
2382  * a default command ID and not allowing access to the speedup options.
2383  *
2384  * This should be used rather than using heap_insert directly in most places
2385  * where we are modifying system catalogs.
2386  */
2387 void
2388 simple_heap_insert(Relation relation, HeapTuple tup)
2389 {
2390         heap_insert(relation, tup, GetCurrentCommandId(true), 0, NULL);
2391 }
2392
2393 /*
2394  * Given infomask/infomask2, compute the bits that must be saved in the
2395  * "infobits" field of xl_heap_delete, xl_heap_update, xl_heap_lock,
2396  * xl_heap_lock_updated WAL records.
2397  *
2398  * See fix_infomask_from_infobits.
2399  */
2400 static uint8
2401 compute_infobits(uint16 infomask, uint16 infomask2)
2402 {
2403         return
2404                 ((infomask & HEAP_XMAX_IS_MULTI) != 0 ? XLHL_XMAX_IS_MULTI : 0) |
2405                 ((infomask & HEAP_XMAX_LOCK_ONLY) != 0 ? XLHL_XMAX_LOCK_ONLY : 0) |
2406                 ((infomask & HEAP_XMAX_EXCL_LOCK) != 0 ? XLHL_XMAX_EXCL_LOCK : 0) |
2407         /* note we ignore HEAP_XMAX_SHR_LOCK here */
2408                 ((infomask & HEAP_XMAX_KEYSHR_LOCK) != 0 ? XLHL_XMAX_KEYSHR_LOCK : 0) |
2409                 ((infomask2 & HEAP_KEYS_UPDATED) != 0 ?
2410                  XLHL_KEYS_UPDATED : 0);
2411 }
2412
2413 /*
2414  * Given two versions of the same t_infomask for a tuple, compare them and
2415  * return whether the relevant status for a tuple Xmax has changed.  This is
2416  * used after a buffer lock has been released and reacquired: we want to ensure
2417  * that the tuple state continues to be the same it was when we previously
2418  * examined it.
2419  *
2420  * Note the Xmax field itself must be compared separately.
2421  */
2422 static inline bool
2423 xmax_infomask_changed(uint16 new_infomask, uint16 old_infomask)
2424 {
2425         const uint16 interesting =
2426         HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY | HEAP_LOCK_MASK;
2427
2428         if ((new_infomask & interesting) != (old_infomask & interesting))
2429                 return true;
2430
2431         return false;
2432 }
2433
2434 /*
2435  *      heap_delete - delete a tuple
2436  *
2437  * See table_delete() for an explanation of the parameters, except that this
2438  * routine directly takes a tuple rather than a slot.
2439  *
2440  * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
2441  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
2442  * only for TM_SelfModified, since we cannot obtain cmax from a combocid
2443  * generated by another transaction).
2444  */
2445 TM_Result
2446 heap_delete(Relation relation, ItemPointer tid,
2447                         CommandId cid, Snapshot crosscheck, bool wait,
2448                         TM_FailureData *tmfd, bool changingPart)
2449 {
2450         TM_Result       result;
2451         TransactionId xid = GetCurrentTransactionId();
2452         ItemId          lp;
2453         HeapTupleData tp;
2454         Page            page;
2455         BlockNumber block;
2456         Buffer          buffer;
2457         Buffer          vmbuffer = InvalidBuffer;
2458         TransactionId new_xmax;
2459         uint16          new_infomask,
2460                                 new_infomask2;
2461         bool            have_tuple_lock = false;
2462         bool            iscombo;
2463         bool            all_visible_cleared = false;
2464         HeapTuple       old_key_tuple = NULL;   /* replica identity of the tuple */
2465         bool            old_key_copied = false;
2466
2467         Assert(ItemPointerIsValid(tid));
2468
2469         /*
2470          * Forbid this during a parallel operation, lest it allocate a combocid.
2471          * Other workers might need that combocid for visibility checks, and we
2472          * have no provision for broadcasting it to them.
2473          */
2474         if (IsInParallelMode())
2475                 ereport(ERROR,
2476                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2477                                  errmsg("cannot delete tuples during a parallel operation")));
2478
2479         block = ItemPointerGetBlockNumber(tid);
2480         buffer = ReadBuffer(relation, block);
2481         page = BufferGetPage(buffer);
2482
2483         /*
2484          * Before locking the buffer, pin the visibility map page if it appears to
2485          * be necessary.  Since we haven't got the lock yet, someone else might be
2486          * in the middle of changing this, so we'll need to recheck after we have
2487          * the lock.
2488          */
2489         if (PageIsAllVisible(page))
2490                 visibilitymap_pin(relation, block, &vmbuffer);
2491
2492         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2493
2494         /*
2495          * If we didn't pin the visibility map page and the page has become all
2496          * visible while we were busy locking the buffer, we'll have to unlock and
2497          * re-lock, to avoid holding the buffer lock across an I/O.  That's a bit
2498          * unfortunate, but hopefully shouldn't happen often.
2499          */
2500         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
2501         {
2502                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2503                 visibilitymap_pin(relation, block, &vmbuffer);
2504                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2505         }
2506
2507         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
2508         Assert(ItemIdIsNormal(lp));
2509
2510         tp.t_tableOid = RelationGetRelid(relation);
2511         tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
2512         tp.t_len = ItemIdGetLength(lp);
2513         tp.t_self = *tid;
2514
2515 l1:
2516         result = HeapTupleSatisfiesUpdate(&tp, cid, buffer);
2517
2518         if (result == TM_Invisible)
2519         {
2520                 UnlockReleaseBuffer(buffer);
2521                 ereport(ERROR,
2522                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
2523                                  errmsg("attempted to delete invisible tuple")));
2524         }
2525         else if (result == TM_BeingModified && wait)
2526         {
2527                 TransactionId xwait;
2528                 uint16          infomask;
2529
2530                 /* must copy state data before unlocking buffer */
2531                 xwait = HeapTupleHeaderGetRawXmax(tp.t_data);
2532                 infomask = tp.t_data->t_infomask;
2533
2534                 /*
2535                  * Sleep until concurrent transaction ends -- except when there's a
2536                  * single locker and it's our own transaction.  Note we don't care
2537                  * which lock mode the locker has, because we need the strongest one.
2538                  *
2539                  * Before sleeping, we need to acquire tuple lock to establish our
2540                  * priority for the tuple (see heap_lock_tuple).  LockTuple will
2541                  * release us when we are next-in-line for the tuple.
2542                  *
2543                  * If we are forced to "start over" below, we keep the tuple lock;
2544                  * this arranges that we stay at the head of the line while rechecking
2545                  * tuple state.
2546                  */
2547                 if (infomask & HEAP_XMAX_IS_MULTI)
2548                 {
2549                         /* wait for multixact */
2550                         if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
2551                                                                                 LockTupleExclusive))
2552                         {
2553                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2554
2555                                 /* acquire tuple lock, if necessary */
2556                                 heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2557                                                                          LockWaitBlock, &have_tuple_lock);
2558
2559                                 /* wait for multixact */
2560                                 MultiXactIdWait((MultiXactId) xwait, MultiXactStatusUpdate, infomask,
2561                                                                 relation, &(tp.t_self), XLTW_Delete,
2562                                                                 NULL);
2563                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2564
2565                                 /*
2566                                  * If xwait had just locked the tuple then some other xact
2567                                  * could update this tuple before we get to this point.  Check
2568                                  * for xmax change, and start over if so.
2569                                  */
2570                                 if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2571                                         !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2572                                                                                  xwait))
2573                                         goto l1;
2574                         }
2575
2576                         /*
2577                          * You might think the multixact is necessarily done here, but not
2578                          * so: it could have surviving members, namely our own xact or
2579                          * other subxacts of this backend.  It is legal for us to delete
2580                          * the tuple in either case, however (the latter case is
2581                          * essentially a situation of upgrading our former shared lock to
2582                          * exclusive).  We don't bother changing the on-disk hint bits
2583                          * since we are about to overwrite the xmax altogether.
2584                          */
2585                 }
2586                 else if (!TransactionIdIsCurrentTransactionId(xwait))
2587                 {
2588                         /*
2589                          * Wait for regular transaction to end; but first, acquire tuple
2590                          * lock.
2591                          */
2592                         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2593                         heap_acquire_tuplock(relation, &(tp.t_self), LockTupleExclusive,
2594                                                                  LockWaitBlock, &have_tuple_lock);
2595                         XactLockTableWait(xwait, relation, &(tp.t_self), XLTW_Delete);
2596                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
2597
2598                         /*
2599                          * xwait is done, but if xwait had just locked the tuple then some
2600                          * other xact could update this tuple before we get to this point.
2601                          * Check for xmax change, and start over if so.
2602                          */
2603                         if (xmax_infomask_changed(tp.t_data->t_infomask, infomask) ||
2604                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tp.t_data),
2605                                                                          xwait))
2606                                 goto l1;
2607
2608                         /* Otherwise check if it committed or aborted */
2609                         UpdateXmaxHintBits(tp.t_data, buffer, xwait);
2610                 }
2611
2612                 /*
2613                  * We may overwrite if previous xmax aborted, or if it committed but
2614                  * only locked the tuple without updating it.
2615                  */
2616                 if ((tp.t_data->t_infomask & HEAP_XMAX_INVALID) ||
2617                         HEAP_XMAX_IS_LOCKED_ONLY(tp.t_data->t_infomask) ||
2618                         HeapTupleHeaderIsOnlyLocked(tp.t_data))
2619                         result = TM_Ok;
2620                 else if (!ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid) ||
2621                                  HeapTupleHeaderIndicatesMovedPartitions(tp.t_data))
2622                         result = TM_Updated;
2623                 else
2624                         result = TM_Deleted;
2625         }
2626
2627         if (crosscheck != InvalidSnapshot && result == TM_Ok)
2628         {
2629                 /* Perform additional check for transaction-snapshot mode RI updates */
2630                 if (!HeapTupleSatisfiesVisibility(&tp, crosscheck, buffer))
2631                         result = TM_Updated;
2632         }
2633
2634         if (result != TM_Ok)
2635         {
2636                 Assert(result == TM_SelfModified ||
2637                            result == TM_Updated ||
2638                            result == TM_Deleted ||
2639                            result == TM_BeingModified);
2640                 Assert(!(tp.t_data->t_infomask & HEAP_XMAX_INVALID));
2641                 Assert(result != TM_Updated ||
2642                            !ItemPointerEquals(&tp.t_self, &tp.t_data->t_ctid));
2643                 tmfd->ctid = tp.t_data->t_ctid;
2644                 tmfd->xmax = HeapTupleHeaderGetUpdateXid(tp.t_data);
2645                 if (result == TM_SelfModified)
2646                         tmfd->cmax = HeapTupleHeaderGetCmax(tp.t_data);
2647                 else
2648                         tmfd->cmax = InvalidCommandId;
2649                 UnlockReleaseBuffer(buffer);
2650                 if (have_tuple_lock)
2651                         UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2652                 if (vmbuffer != InvalidBuffer)
2653                         ReleaseBuffer(vmbuffer);
2654                 return result;
2655         }
2656
2657         /*
2658          * We're about to do the actual delete -- check for conflict first, to
2659          * avoid possibly having to roll back work we've just done.
2660          *
2661          * This is safe without a recheck as long as there is no possibility of
2662          * another process scanning the page between this check and the delete
2663          * being visible to the scan (i.e., an exclusive buffer content lock is
2664          * continuously held from this point until the tuple delete is visible).
2665          */
2666         CheckForSerializableConflictIn(relation, &tp, buffer);
2667
2668         /* replace cid with a combo cid if necessary */
2669         HeapTupleHeaderAdjustCmax(tp.t_data, &cid, &iscombo);
2670
2671         /*
2672          * Compute replica identity tuple before entering the critical section so
2673          * we don't PANIC upon a memory allocation failure.
2674          */
2675         old_key_tuple = ExtractReplicaIdentity(relation, &tp, true, &old_key_copied);
2676
2677         /*
2678          * If this is the first possibly-multixact-able operation in the current
2679          * transaction, set my per-backend OldestMemberMXactId setting. We can be
2680          * certain that the transaction will never become a member of any older
2681          * MultiXactIds than that.  (We have to do this even if we end up just
2682          * using our own TransactionId below, since some other backend could
2683          * incorporate our XID into a MultiXact immediately afterwards.)
2684          */
2685         MultiXactIdSetOldestMember();
2686
2687         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(tp.t_data),
2688                                                           tp.t_data->t_infomask, tp.t_data->t_infomask2,
2689                                                           xid, LockTupleExclusive, true,
2690                                                           &new_xmax, &new_infomask, &new_infomask2);
2691
2692         START_CRIT_SECTION();
2693
2694         /*
2695          * If this transaction commits, the tuple will become DEAD sooner or
2696          * later.  Set flag that this page is a candidate for pruning once our xid
2697          * falls below the OldestXmin horizon.  If the transaction finally aborts,
2698          * the subsequent page pruning will be a no-op and the hint will be
2699          * cleared.
2700          */
2701         PageSetPrunable(page, xid);
2702
2703         if (PageIsAllVisible(page))
2704         {
2705                 all_visible_cleared = true;
2706                 PageClearAllVisible(page);
2707                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
2708                                                         vmbuffer, VISIBILITYMAP_VALID_BITS);
2709         }
2710
2711         /* store transaction information of xact deleting the tuple */
2712         tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
2713         tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
2714         tp.t_data->t_infomask |= new_infomask;
2715         tp.t_data->t_infomask2 |= new_infomask2;
2716         HeapTupleHeaderClearHotUpdated(tp.t_data);
2717         HeapTupleHeaderSetXmax(tp.t_data, new_xmax);
2718         HeapTupleHeaderSetCmax(tp.t_data, cid, iscombo);
2719         /* Make sure there is no forward chain link in t_ctid */
2720         tp.t_data->t_ctid = tp.t_self;
2721
2722         /* Signal that this is actually a move into another partition */
2723         if (changingPart)
2724                 HeapTupleHeaderSetMovedPartitions(tp.t_data);
2725
2726         MarkBufferDirty(buffer);
2727
2728         /*
2729          * XLOG stuff
2730          *
2731          * NB: heap_abort_speculative() uses the same xlog record and replay
2732          * routines.
2733          */
2734         if (RelationNeedsWAL(relation))
2735         {
2736                 xl_heap_delete xlrec;
2737                 xl_heap_header xlhdr;
2738                 XLogRecPtr      recptr;
2739
2740                 /* For logical decode we need combocids to properly decode the catalog */
2741                 if (RelationIsAccessibleInLogicalDecoding(relation))
2742                         log_heap_new_cid(relation, &tp);
2743
2744                 xlrec.flags = 0;
2745                 if (all_visible_cleared)
2746                         xlrec.flags |= XLH_DELETE_ALL_VISIBLE_CLEARED;
2747                 if (changingPart)
2748                         xlrec.flags |= XLH_DELETE_IS_PARTITION_MOVE;
2749                 xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
2750                                                                                           tp.t_data->t_infomask2);
2751                 xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
2752                 xlrec.xmax = new_xmax;
2753
2754                 if (old_key_tuple != NULL)
2755                 {
2756                         if (relation->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
2757                                 xlrec.flags |= XLH_DELETE_CONTAINS_OLD_TUPLE;
2758                         else
2759                                 xlrec.flags |= XLH_DELETE_CONTAINS_OLD_KEY;
2760                 }
2761
2762                 XLogBeginInsert();
2763                 XLogRegisterData((char *) &xlrec, SizeOfHeapDelete);
2764
2765                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
2766
2767                 /*
2768                  * Log replica identity of the deleted tuple if there is one
2769                  */
2770                 if (old_key_tuple != NULL)
2771                 {
2772                         xlhdr.t_infomask2 = old_key_tuple->t_data->t_infomask2;
2773                         xlhdr.t_infomask = old_key_tuple->t_data->t_infomask;
2774                         xlhdr.t_hoff = old_key_tuple->t_data->t_hoff;
2775
2776                         XLogRegisterData((char *) &xlhdr, SizeOfHeapHeader);
2777                         XLogRegisterData((char *) old_key_tuple->t_data
2778                                                          + SizeofHeapTupleHeader,
2779                                                          old_key_tuple->t_len
2780                                                          - SizeofHeapTupleHeader);
2781                 }
2782
2783                 /* filtering by origin on a row level is much more efficient */
2784                 XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
2785
2786                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
2787
2788                 PageSetLSN(page, recptr);
2789         }
2790
2791         END_CRIT_SECTION();
2792
2793         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
2794
2795         if (vmbuffer != InvalidBuffer)
2796                 ReleaseBuffer(vmbuffer);
2797
2798         /*
2799          * If the tuple has toasted out-of-line attributes, we need to delete
2800          * those items too.  We have to do this before releasing the buffer
2801          * because we need to look at the contents of the tuple, but it's OK to
2802          * release the content lock on the buffer first.
2803          */
2804         if (relation->rd_rel->relkind != RELKIND_RELATION &&
2805                 relation->rd_rel->relkind != RELKIND_MATVIEW)
2806         {
2807                 /* toast table entries should never be recursively toasted */
2808                 Assert(!HeapTupleHasExternal(&tp));
2809         }
2810         else if (HeapTupleHasExternal(&tp))
2811                 toast_delete(relation, &tp, false);
2812
2813         /*
2814          * Mark tuple for invalidation from system caches at next command
2815          * boundary. We have to do this before releasing the buffer because we
2816          * need to look at the contents of the tuple.
2817          */
2818         CacheInvalidateHeapTuple(relation, &tp, NULL);
2819
2820         /* Now we can release the buffer */
2821         ReleaseBuffer(buffer);
2822
2823         /*
2824          * Release the lmgr tuple lock, if we had it.
2825          */
2826         if (have_tuple_lock)
2827                 UnlockTupleTuplock(relation, &(tp.t_self), LockTupleExclusive);
2828
2829         pgstat_count_heap_delete(relation);
2830
2831         if (old_key_tuple != NULL && old_key_copied)
2832                 heap_freetuple(old_key_tuple);
2833
2834         return TM_Ok;
2835 }
2836
2837 /*
2838  *      simple_heap_delete - delete a tuple
2839  *
2840  * This routine may be used to delete a tuple when concurrent updates of
2841  * the target tuple are not expected (for example, because we have a lock
2842  * on the relation associated with the tuple).  Any failure is reported
2843  * via ereport().
2844  */
2845 void
2846 simple_heap_delete(Relation relation, ItemPointer tid)
2847 {
2848         TM_Result       result;
2849         TM_FailureData tmfd;
2850
2851         result = heap_delete(relation, tid,
2852                                                  GetCurrentCommandId(true), InvalidSnapshot,
2853                                                  true /* wait for commit */ ,
2854                                                  &tmfd, false /* changingPart */ );
2855         switch (result)
2856         {
2857                 case TM_SelfModified:
2858                         /* Tuple was already updated in current command? */
2859                         elog(ERROR, "tuple already updated by self");
2860                         break;
2861
2862                 case TM_Ok:
2863                         /* done successfully */
2864                         break;
2865
2866                 case TM_Updated:
2867                         elog(ERROR, "tuple concurrently updated");
2868                         break;
2869
2870                 case TM_Deleted:
2871                         elog(ERROR, "tuple concurrently deleted");
2872                         break;
2873
2874                 default:
2875                         elog(ERROR, "unrecognized heap_delete status: %u", result);
2876                         break;
2877         }
2878 }
2879
2880 /*
2881  *      heap_update - replace a tuple
2882  *
2883  * See table_update() for an explanation of the parameters, except that this
2884  * routine directly takes a tuple rather than a slot.
2885  *
2886  * In the failure cases, the routine fills *tmfd with the tuple's t_ctid,
2887  * t_xmax (resolving a possible MultiXact, if necessary), and t_cmax (the last
2888  * only for TM_SelfModified, since we cannot obtain cmax from a combocid
2889  * generated by another transaction).
2890  */
2891 TM_Result
2892 heap_update(Relation relation, ItemPointer otid, HeapTuple newtup,
2893                         CommandId cid, Snapshot crosscheck, bool wait,
2894                         TM_FailureData *tmfd, LockTupleMode *lockmode)
2895 {
2896         TM_Result       result;
2897         TransactionId xid = GetCurrentTransactionId();
2898         Bitmapset  *hot_attrs;
2899         Bitmapset  *key_attrs;
2900         Bitmapset  *id_attrs;
2901         Bitmapset  *interesting_attrs;
2902         Bitmapset  *modified_attrs;
2903         ItemId          lp;
2904         HeapTupleData oldtup;
2905         HeapTuple       heaptup;
2906         HeapTuple       old_key_tuple = NULL;
2907         bool            old_key_copied = false;
2908         Page            page;
2909         BlockNumber block;
2910         MultiXactStatus mxact_status;
2911         Buffer          buffer,
2912                                 newbuf,
2913                                 vmbuffer = InvalidBuffer,
2914                                 vmbuffer_new = InvalidBuffer;
2915         bool            need_toast;
2916         Size            newtupsize,
2917                                 pagefree;
2918         bool            have_tuple_lock = false;
2919         bool            iscombo;
2920         bool            use_hot_update = false;
2921         bool            hot_attrs_checked = false;
2922         bool            key_intact;
2923         bool            all_visible_cleared = false;
2924         bool            all_visible_cleared_new = false;
2925         bool            checked_lockers;
2926         bool            locker_remains;
2927         TransactionId xmax_new_tuple,
2928                                 xmax_old_tuple;
2929         uint16          infomask_old_tuple,
2930                                 infomask2_old_tuple,
2931                                 infomask_new_tuple,
2932                                 infomask2_new_tuple;
2933
2934         Assert(ItemPointerIsValid(otid));
2935
2936         /*
2937          * Forbid this during a parallel operation, lest it allocate a combocid.
2938          * Other workers might need that combocid for visibility checks, and we
2939          * have no provision for broadcasting it to them.
2940          */
2941         if (IsInParallelMode())
2942                 ereport(ERROR,
2943                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
2944                                  errmsg("cannot update tuples during a parallel operation")));
2945
2946         /*
2947          * Fetch the list of attributes to be checked for various operations.
2948          *
2949          * For HOT considerations, this is wasted effort if we fail to update or
2950          * have to put the new tuple on a different page.  But we must compute the
2951          * list before obtaining buffer lock --- in the worst case, if we are
2952          * doing an update on one of the relevant system catalogs, we could
2953          * deadlock if we try to fetch the list later.  In any case, the relcache
2954          * caches the data so this is usually pretty cheap.
2955          *
2956          * We also need columns used by the replica identity and columns that are
2957          * considered the "key" of rows in the table.
2958          *
2959          * Note that we get copies of each bitmap, so we need not worry about
2960          * relcache flush happening midway through.
2961          */
2962         hot_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_ALL);
2963         key_attrs = RelationGetIndexAttrBitmap(relation, INDEX_ATTR_BITMAP_KEY);
2964         id_attrs = RelationGetIndexAttrBitmap(relation,
2965                                                                                   INDEX_ATTR_BITMAP_IDENTITY_KEY);
2966
2967
2968         block = ItemPointerGetBlockNumber(otid);
2969         buffer = ReadBuffer(relation, block);
2970         page = BufferGetPage(buffer);
2971
2972         interesting_attrs = NULL;
2973
2974         /*
2975          * If the page is already full, there is hardly any chance of doing a HOT
2976          * update on this page. It might be wasteful effort to look for index
2977          * column updates only to later reject HOT updates for lack of space in
2978          * the same page. So we be conservative and only fetch hot_attrs if the
2979          * page is not already full. Since we are already holding a pin on the
2980          * buffer, there is no chance that the buffer can get cleaned up
2981          * concurrently and even if that was possible, in the worst case we lose a
2982          * chance to do a HOT update.
2983          */
2984         if (!PageIsFull(page))
2985         {
2986                 interesting_attrs = bms_add_members(interesting_attrs, hot_attrs);
2987                 hot_attrs_checked = true;
2988         }
2989         interesting_attrs = bms_add_members(interesting_attrs, key_attrs);
2990         interesting_attrs = bms_add_members(interesting_attrs, id_attrs);
2991
2992         /*
2993          * Before locking the buffer, pin the visibility map page if it appears to
2994          * be necessary.  Since we haven't got the lock yet, someone else might be
2995          * in the middle of changing this, so we'll need to recheck after we have
2996          * the lock.
2997          */
2998         if (PageIsAllVisible(page))
2999                 visibilitymap_pin(relation, block, &vmbuffer);
3000
3001         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3002
3003         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(otid));
3004         Assert(ItemIdIsNormal(lp));
3005
3006         /*
3007          * Fill in enough data in oldtup for HeapDetermineModifiedColumns to work
3008          * properly.
3009          */
3010         oldtup.t_tableOid = RelationGetRelid(relation);
3011         oldtup.t_data = (HeapTupleHeader) PageGetItem(page, lp);
3012         oldtup.t_len = ItemIdGetLength(lp);
3013         oldtup.t_self = *otid;
3014
3015         /* the new tuple is ready, except for this: */
3016         newtup->t_tableOid = RelationGetRelid(relation);
3017
3018         /* Determine columns modified by the update. */
3019         modified_attrs = HeapDetermineModifiedColumns(relation, interesting_attrs,
3020                                                                                                   &oldtup, newtup);
3021
3022         /*
3023          * If we're not updating any "key" column, we can grab a weaker lock type.
3024          * This allows for more concurrency when we are running simultaneously
3025          * with foreign key checks.
3026          *
3027          * Note that if a column gets detoasted while executing the update, but
3028          * the value ends up being the same, this test will fail and we will use
3029          * the stronger lock.  This is acceptable; the important case to optimize
3030          * is updates that don't manipulate key columns, not those that
3031          * serendipitiously arrive at the same key values.
3032          */
3033         if (!bms_overlap(modified_attrs, key_attrs))
3034         {
3035                 *lockmode = LockTupleNoKeyExclusive;
3036                 mxact_status = MultiXactStatusNoKeyUpdate;
3037                 key_intact = true;
3038
3039                 /*
3040                  * If this is the first possibly-multixact-able operation in the
3041                  * current transaction, set my per-backend OldestMemberMXactId
3042                  * setting. We can be certain that the transaction will never become a
3043                  * member of any older MultiXactIds than that.  (We have to do this
3044                  * even if we end up just using our own TransactionId below, since
3045                  * some other backend could incorporate our XID into a MultiXact
3046                  * immediately afterwards.)
3047                  */
3048                 MultiXactIdSetOldestMember();
3049         }
3050         else
3051         {
3052                 *lockmode = LockTupleExclusive;
3053                 mxact_status = MultiXactStatusUpdate;
3054                 key_intact = false;
3055         }
3056
3057         /*
3058          * Note: beyond this point, use oldtup not otid to refer to old tuple.
3059          * otid may very well point at newtup->t_self, which we will overwrite
3060          * with the new tuple's location, so there's great risk of confusion if we
3061          * use otid anymore.
3062          */
3063
3064 l2:
3065         checked_lockers = false;
3066         locker_remains = false;
3067         result = HeapTupleSatisfiesUpdate(&oldtup, cid, buffer);
3068
3069         /* see below about the "no wait" case */
3070         Assert(result != TM_BeingModified || wait);
3071
3072         if (result == TM_Invisible)
3073         {
3074                 UnlockReleaseBuffer(buffer);
3075                 ereport(ERROR,
3076                                 (errcode(ERRCODE_OBJECT_NOT_IN_PREREQUISITE_STATE),
3077                                  errmsg("attempted to update invisible tuple")));
3078         }
3079         else if (result == TM_BeingModified && wait)
3080         {
3081                 TransactionId xwait;
3082                 uint16          infomask;
3083                 bool            can_continue = false;
3084
3085                 /*
3086                  * XXX note that we don't consider the "no wait" case here.  This
3087                  * isn't a problem currently because no caller uses that case, but it
3088                  * should be fixed if such a caller is introduced.  It wasn't a
3089                  * problem previously because this code would always wait, but now
3090                  * that some tuple locks do not conflict with one of the lock modes we
3091                  * use, it is possible that this case is interesting to handle
3092                  * specially.
3093                  *
3094                  * This may cause failures with third-party code that calls
3095                  * heap_update directly.
3096                  */
3097
3098                 /* must copy state data before unlocking buffer */
3099                 xwait = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3100                 infomask = oldtup.t_data->t_infomask;
3101
3102                 /*
3103                  * Now we have to do something about the existing locker.  If it's a
3104                  * multi, sleep on it; we might be awakened before it is completely
3105                  * gone (or even not sleep at all in some cases); we need to preserve
3106                  * it as locker, unless it is gone completely.
3107                  *
3108                  * If it's not a multi, we need to check for sleeping conditions
3109                  * before actually going to sleep.  If the update doesn't conflict
3110                  * with the locks, we just continue without sleeping (but making sure
3111                  * it is preserved).
3112                  *
3113                  * Before sleeping, we need to acquire tuple lock to establish our
3114                  * priority for the tuple (see heap_lock_tuple).  LockTuple will
3115                  * release us when we are next-in-line for the tuple.  Note we must
3116                  * not acquire the tuple lock until we're sure we're going to sleep;
3117                  * otherwise we're open for race conditions with other transactions
3118                  * holding the tuple lock which sleep on us.
3119                  *
3120                  * If we are forced to "start over" below, we keep the tuple lock;
3121                  * this arranges that we stay at the head of the line while rechecking
3122                  * tuple state.
3123                  */
3124                 if (infomask & HEAP_XMAX_IS_MULTI)
3125                 {
3126                         TransactionId update_xact;
3127                         int                     remain;
3128
3129                         if (DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
3130                                                                                 *lockmode))
3131                         {
3132                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3133
3134                                 /* acquire tuple lock, if necessary */
3135                                 heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3136                                                                          LockWaitBlock, &have_tuple_lock);
3137
3138                                 /* wait for multixact */
3139                                 MultiXactIdWait((MultiXactId) xwait, mxact_status, infomask,
3140                                                                 relation, &oldtup.t_self, XLTW_Update,
3141                                                                 &remain);
3142                                 checked_lockers = true;
3143                                 locker_remains = remain != 0;
3144                                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3145
3146                                 /*
3147                                  * If xwait had just locked the tuple then some other xact
3148                                  * could update this tuple before we get to this point.  Check
3149                                  * for xmax change, and start over if so.
3150                                  */
3151                                 if (xmax_infomask_changed(oldtup.t_data->t_infomask,
3152                                                                                   infomask) ||
3153                                         !TransactionIdEquals(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3154                                                                                  xwait))
3155                                         goto l2;
3156                         }
3157
3158                         /*
3159                          * Note that the multixact may not be done by now.  It could have
3160                          * surviving members; our own xact or other subxacts of this
3161                          * backend, and also any other concurrent transaction that locked
3162                          * the tuple with KeyShare if we only got TupleLockUpdate.  If
3163                          * this is the case, we have to be careful to mark the updated
3164                          * tuple with the surviving members in Xmax.
3165                          *
3166                          * Note that there could have been another update in the
3167                          * MultiXact. In that case, we need to check whether it committed
3168                          * or aborted. If it aborted we are safe to update it again;
3169                          * otherwise there is an update conflict, and we have to return
3170                          * TableTuple{Deleted, Updated} below.
3171                          *
3172                          * In the LockTupleExclusive case, we still need to preserve the
3173                          * surviving members: those would include the tuple locks we had
3174                          * before this one, which are important to keep in case this
3175                          * subxact aborts.
3176                          */
3177                         if (!HEAP_XMAX_IS_LOCKED_ONLY(oldtup.t_data->t_infomask))
3178                                 update_xact = HeapTupleGetUpdateXid(oldtup.t_data);
3179                         else
3180                                 update_xact = InvalidTransactionId;
3181
3182                         /*
3183                          * There was no UPDATE in the MultiXact; or it aborted. No
3184                          * TransactionIdIsInProgress() call needed here, since we called
3185                          * MultiXactIdWait() above.
3186                          */
3187                         if (!TransactionIdIsValid(update_xact) ||
3188                                 TransactionIdDidAbort(update_xact))
3189                                 can_continue = true;
3190                 }
3191                 else if (TransactionIdIsCurrentTransactionId(xwait))
3192                 {
3193                         /*
3194                          * The only locker is ourselves; we can avoid grabbing the tuple
3195                          * lock here, but must preserve our locking information.
3196                          */
3197                         checked_lockers = true;
3198                         locker_remains = true;
3199                         can_continue = true;
3200                 }
3201                 else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) && key_intact)
3202                 {
3203                         /*
3204                          * If it's just a key-share locker, and we're not changing the key
3205                          * columns, we don't need to wait for it to end; but we need to
3206                          * preserve it as locker.
3207                          */
3208                         checked_lockers = true;
3209                         locker_remains = true;
3210                         can_continue = true;
3211                 }
3212                 else
3213                 {
3214                         /*
3215                          * Wait for regular transaction to end; but first, acquire tuple
3216                          * lock.
3217                          */
3218                         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3219                         heap_acquire_tuplock(relation, &(oldtup.t_self), *lockmode,
3220                                                                  LockWaitBlock, &have_tuple_lock);
3221                         XactLockTableWait(xwait, relation, &oldtup.t_self,
3222                                                           XLTW_Update);
3223                         checked_lockers = true;
3224                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3225
3226                         /*
3227                          * xwait is done, but if xwait had just locked the tuple then some
3228                          * other xact could update this tuple before we get to this point.
3229                          * Check for xmax change, and start over if so.
3230                          */
3231                         if (xmax_infomask_changed(oldtup.t_data->t_infomask, infomask) ||
3232                                 !TransactionIdEquals(xwait,
3233                                                                          HeapTupleHeaderGetRawXmax(oldtup.t_data)))
3234                                 goto l2;
3235
3236                         /* Otherwise check if it committed or aborted */
3237                         UpdateXmaxHintBits(oldtup.t_data, buffer, xwait);
3238                         if (oldtup.t_data->t_infomask & HEAP_XMAX_INVALID)
3239                                 can_continue = true;
3240                 }
3241
3242                 if (can_continue)
3243                         result = TM_Ok;
3244                 else if (!ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid) ||
3245                                  HeapTupleHeaderIndicatesMovedPartitions(oldtup.t_data))
3246                         result = TM_Updated;
3247                 else
3248                         result = TM_Deleted;
3249         }
3250
3251         if (crosscheck != InvalidSnapshot && result == TM_Ok)
3252         {
3253                 /* Perform additional check for transaction-snapshot mode RI updates */
3254                 if (!HeapTupleSatisfiesVisibility(&oldtup, crosscheck, buffer))
3255                 {
3256                         result = TM_Updated;
3257                         Assert(!ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid));
3258                 }
3259         }
3260
3261         if (result != TM_Ok)
3262         {
3263                 Assert(result == TM_SelfModified ||
3264                            result == TM_Updated ||
3265                            result == TM_Deleted ||
3266                            result == TM_BeingModified);
3267                 Assert(!(oldtup.t_data->t_infomask & HEAP_XMAX_INVALID));
3268                 Assert(result != TM_Updated ||
3269                            !ItemPointerEquals(&oldtup.t_self, &oldtup.t_data->t_ctid));
3270                 tmfd->ctid = oldtup.t_data->t_ctid;
3271                 tmfd->xmax = HeapTupleHeaderGetUpdateXid(oldtup.t_data);
3272                 if (result == TM_SelfModified)
3273                         tmfd->cmax = HeapTupleHeaderGetCmax(oldtup.t_data);
3274                 else
3275                         tmfd->cmax = InvalidCommandId;
3276                 UnlockReleaseBuffer(buffer);
3277                 if (have_tuple_lock)
3278                         UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3279                 if (vmbuffer != InvalidBuffer)
3280                         ReleaseBuffer(vmbuffer);
3281                 bms_free(hot_attrs);
3282                 bms_free(key_attrs);
3283                 bms_free(id_attrs);
3284                 bms_free(modified_attrs);
3285                 bms_free(interesting_attrs);
3286                 return result;
3287         }
3288
3289         /*
3290          * If we didn't pin the visibility map page and the page has become all
3291          * visible while we were busy locking the buffer, or during some
3292          * subsequent window during which we had it unlocked, we'll have to unlock
3293          * and re-lock, to avoid holding the buffer lock across an I/O.  That's a
3294          * bit unfortunate, especially since we'll now have to recheck whether the
3295          * tuple has been locked or updated under us, but hopefully it won't
3296          * happen very often.
3297          */
3298         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
3299         {
3300                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3301                 visibilitymap_pin(relation, block, &vmbuffer);
3302                 LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3303                 goto l2;
3304         }
3305
3306         /* Fill in transaction status data */
3307
3308         /*
3309          * If the tuple we're updating is locked, we need to preserve the locking
3310          * info in the old tuple's Xmax.  Prepare a new Xmax value for this.
3311          */
3312         compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3313                                                           oldtup.t_data->t_infomask,
3314                                                           oldtup.t_data->t_infomask2,
3315                                                           xid, *lockmode, true,
3316                                                           &xmax_old_tuple, &infomask_old_tuple,
3317                                                           &infomask2_old_tuple);
3318
3319         /*
3320          * And also prepare an Xmax value for the new copy of the tuple.  If there
3321          * was no xmax previously, or there was one but all lockers are now gone,
3322          * then use InvalidXid; otherwise, get the xmax from the old tuple.  (In
3323          * rare cases that might also be InvalidXid and yet not have the
3324          * HEAP_XMAX_INVALID bit set; that's fine.)
3325          */
3326         if ((oldtup.t_data->t_infomask & HEAP_XMAX_INVALID) ||
3327                 HEAP_LOCKED_UPGRADED(oldtup.t_data->t_infomask) ||
3328                 (checked_lockers && !locker_remains))
3329                 xmax_new_tuple = InvalidTransactionId;
3330         else
3331                 xmax_new_tuple = HeapTupleHeaderGetRawXmax(oldtup.t_data);
3332
3333         if (!TransactionIdIsValid(xmax_new_tuple))
3334         {
3335                 infomask_new_tuple = HEAP_XMAX_INVALID;
3336                 infomask2_new_tuple = 0;
3337         }
3338         else
3339         {
3340                 /*
3341                  * If we found a valid Xmax for the new tuple, then the infomask bits
3342                  * to use on the new tuple depend on what was there on the old one.
3343                  * Note that since we're doing an update, the only possibility is that
3344                  * the lockers had FOR KEY SHARE lock.
3345                  */
3346                 if (oldtup.t_data->t_infomask & HEAP_XMAX_IS_MULTI)
3347                 {
3348                         GetMultiXactIdHintBits(xmax_new_tuple, &infomask_new_tuple,
3349                                                                    &infomask2_new_tuple);
3350                 }
3351                 else
3352                 {
3353                         infomask_new_tuple = HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_LOCK_ONLY;
3354                         infomask2_new_tuple = 0;
3355                 }
3356         }
3357
3358         /*
3359          * Prepare the new tuple with the appropriate initial values of Xmin and
3360          * Xmax, as well as initial infomask bits as computed above.
3361          */
3362         newtup->t_data->t_infomask &= ~(HEAP_XACT_MASK);
3363         newtup->t_data->t_infomask2 &= ~(HEAP2_XACT_MASK);
3364         HeapTupleHeaderSetXmin(newtup->t_data, xid);
3365         HeapTupleHeaderSetCmin(newtup->t_data, cid);
3366         newtup->t_data->t_infomask |= HEAP_UPDATED | infomask_new_tuple;
3367         newtup->t_data->t_infomask2 |= infomask2_new_tuple;
3368         HeapTupleHeaderSetXmax(newtup->t_data, xmax_new_tuple);
3369
3370         /*
3371          * Replace cid with a combo cid if necessary.  Note that we already put
3372          * the plain cid into the new tuple.
3373          */
3374         HeapTupleHeaderAdjustCmax(oldtup.t_data, &cid, &iscombo);
3375
3376         /*
3377          * If the toaster needs to be activated, OR if the new tuple will not fit
3378          * on the same page as the old, then we need to release the content lock
3379          * (but not the pin!) on the old tuple's buffer while we are off doing
3380          * TOAST and/or table-file-extension work.  We must mark the old tuple to
3381          * show that it's locked, else other processes may try to update it
3382          * themselves.
3383          *
3384          * We need to invoke the toaster if there are already any out-of-line
3385          * toasted values present, or if the new tuple is over-threshold.
3386          */
3387         if (relation->rd_rel->relkind != RELKIND_RELATION &&
3388                 relation->rd_rel->relkind != RELKIND_MATVIEW)
3389         {
3390                 /* toast table entries should never be recursively toasted */
3391                 Assert(!HeapTupleHasExternal(&oldtup));
3392                 Assert(!HeapTupleHasExternal(newtup));
3393                 need_toast = false;
3394         }
3395         else
3396                 need_toast = (HeapTupleHasExternal(&oldtup) ||
3397                                           HeapTupleHasExternal(newtup) ||
3398                                           newtup->t_len > TOAST_TUPLE_THRESHOLD);
3399
3400         pagefree = PageGetHeapFreeSpace(page);
3401
3402         newtupsize = MAXALIGN(newtup->t_len);
3403
3404         if (need_toast || newtupsize > pagefree)
3405         {
3406                 TransactionId xmax_lock_old_tuple;
3407                 uint16          infomask_lock_old_tuple,
3408                                         infomask2_lock_old_tuple;
3409                 bool            cleared_all_frozen = false;
3410
3411                 /*
3412                  * To prevent concurrent sessions from updating the tuple, we have to
3413                  * temporarily mark it locked, while we release the page-level lock.
3414                  *
3415                  * To satisfy the rule that any xid potentially appearing in a buffer
3416                  * written out to disk, we unfortunately have to WAL log this
3417                  * temporary modification.  We can reuse xl_heap_lock for this
3418                  * purpose.  If we crash/error before following through with the
3419                  * actual update, xmax will be of an aborted transaction, allowing
3420                  * other sessions to proceed.
3421                  */
3422
3423                 /*
3424                  * Compute xmax / infomask appropriate for locking the tuple. This has
3425                  * to be done separately from the combo that's going to be used for
3426                  * updating, because the potentially created multixact would otherwise
3427                  * be wrong.
3428                  */
3429                 compute_new_xmax_infomask(HeapTupleHeaderGetRawXmax(oldtup.t_data),
3430                                                                   oldtup.t_data->t_infomask,
3431                                                                   oldtup.t_data->t_infomask2,
3432                                                                   xid, *lockmode, false,
3433                                                                   &xmax_lock_old_tuple, &infomask_lock_old_tuple,
3434                                                                   &infomask2_lock_old_tuple);
3435
3436                 Assert(HEAP_XMAX_IS_LOCKED_ONLY(infomask_lock_old_tuple));
3437
3438                 START_CRIT_SECTION();
3439
3440                 /* Clear obsolete visibility flags ... */
3441                 oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3442                 oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3443                 HeapTupleClearHotUpdated(&oldtup);
3444                 /* ... and store info about transaction updating this tuple */
3445                 Assert(TransactionIdIsValid(xmax_lock_old_tuple));
3446                 HeapTupleHeaderSetXmax(oldtup.t_data, xmax_lock_old_tuple);
3447                 oldtup.t_data->t_infomask |= infomask_lock_old_tuple;
3448                 oldtup.t_data->t_infomask2 |= infomask2_lock_old_tuple;
3449                 HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3450
3451                 /* temporarily make it look not-updated, but locked */
3452                 oldtup.t_data->t_ctid = oldtup.t_self;
3453
3454                 /*
3455                  * Clear all-frozen bit on visibility map if needed. We could
3456                  * immediately reset ALL_VISIBLE, but given that the WAL logging
3457                  * overhead would be unchanged, that doesn't seem necessarily
3458                  * worthwhile.
3459                  */
3460                 if (PageIsAllVisible(BufferGetPage(buffer)) &&
3461                         visibilitymap_clear(relation, block, vmbuffer,
3462                                                                 VISIBILITYMAP_ALL_FROZEN))
3463                         cleared_all_frozen = true;
3464
3465                 MarkBufferDirty(buffer);
3466
3467                 if (RelationNeedsWAL(relation))
3468                 {
3469                         xl_heap_lock xlrec;
3470                         XLogRecPtr      recptr;
3471
3472                         XLogBeginInsert();
3473                         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
3474
3475                         xlrec.offnum = ItemPointerGetOffsetNumber(&oldtup.t_self);
3476                         xlrec.locking_xid = xmax_lock_old_tuple;
3477                         xlrec.infobits_set = compute_infobits(oldtup.t_data->t_infomask,
3478                                                                                                   oldtup.t_data->t_infomask2);
3479                         xlrec.flags =
3480                                 cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
3481                         XLogRegisterData((char *) &xlrec, SizeOfHeapLock);
3482                         recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
3483                         PageSetLSN(page, recptr);
3484                 }
3485
3486                 END_CRIT_SECTION();
3487
3488                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3489
3490                 /*
3491                  * Let the toaster do its thing, if needed.
3492                  *
3493                  * Note: below this point, heaptup is the data we actually intend to
3494                  * store into the relation; newtup is the caller's original untoasted
3495                  * data.
3496                  */
3497                 if (need_toast)
3498                 {
3499                         /* Note we always use WAL and FSM during updates */
3500                         heaptup = toast_insert_or_update(relation, newtup, &oldtup, 0);
3501                         newtupsize = MAXALIGN(heaptup->t_len);
3502                 }
3503                 else
3504                         heaptup = newtup;
3505
3506                 /*
3507                  * Now, do we need a new page for the tuple, or not?  This is a bit
3508                  * tricky since someone else could have added tuples to the page while
3509                  * we weren't looking.  We have to recheck the available space after
3510                  * reacquiring the buffer lock.  But don't bother to do that if the
3511                  * former amount of free space is still not enough; it's unlikely
3512                  * there's more free now than before.
3513                  *
3514                  * What's more, if we need to get a new page, we will need to acquire
3515                  * buffer locks on both old and new pages.  To avoid deadlock against
3516                  * some other backend trying to get the same two locks in the other
3517                  * order, we must be consistent about the order we get the locks in.
3518                  * We use the rule "lock the lower-numbered page of the relation
3519                  * first".  To implement this, we must do RelationGetBufferForTuple
3520                  * while not holding the lock on the old page, and we must rely on it
3521                  * to get the locks on both pages in the correct order.
3522                  */
3523                 if (newtupsize > pagefree)
3524                 {
3525                         /* Assume there's no chance to put heaptup on same page. */
3526                         newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3527                                                                                            buffer, 0, NULL,
3528                                                                                            &vmbuffer_new, &vmbuffer);
3529                 }
3530                 else
3531                 {
3532                         /* Re-acquire the lock on the old tuple's page. */
3533                         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
3534                         /* Re-check using the up-to-date free space */
3535                         pagefree = PageGetHeapFreeSpace(page);
3536                         if (newtupsize > pagefree)
3537                         {
3538                                 /*
3539                                  * Rats, it doesn't fit anymore.  We must now unlock and
3540                                  * relock to avoid deadlock.  Fortunately, this path should
3541                                  * seldom be taken.
3542                                  */
3543                                 LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3544                                 newbuf = RelationGetBufferForTuple(relation, heaptup->t_len,
3545                                                                                                    buffer, 0, NULL,
3546                                                                                                    &vmbuffer_new, &vmbuffer);
3547                         }
3548                         else
3549                         {
3550                                 /* OK, it fits here, so we're done. */
3551                                 newbuf = buffer;
3552                         }
3553                 }
3554         }
3555         else
3556         {
3557                 /* No TOAST work needed, and it'll fit on same page */
3558                 newbuf = buffer;
3559                 heaptup = newtup;
3560         }
3561
3562         /*
3563          * We're about to do the actual update -- check for conflict first, to
3564          * avoid possibly having to roll back work we've just done.
3565          *
3566          * This is safe without a recheck as long as there is no possibility of
3567          * another process scanning the pages between this check and the update
3568          * being visible to the scan (i.e., exclusive buffer content lock(s) are
3569          * continuously held from this point until the tuple update is visible).
3570          *
3571          * For the new tuple the only check needed is at the relation level, but
3572          * since both tuples are in the same relation and the check for oldtup
3573          * will include checking the relation level, there is no benefit to a
3574          * separate check for the new tuple.
3575          */
3576         CheckForSerializableConflictIn(relation, &oldtup, buffer);
3577
3578         /*
3579          * At this point newbuf and buffer are both pinned and locked, and newbuf
3580          * has enough space for the new tuple.  If they are the same buffer, only
3581          * one pin is held.
3582          */
3583
3584         if (newbuf == buffer)
3585         {
3586                 /*
3587                  * Since the new tuple is going into the same page, we might be able
3588                  * to do a HOT update.  Check if any of the index columns have been
3589                  * changed. If the page was already full, we may have skipped checking
3590                  * for index columns, and also can't do a HOT update.
3591                  */
3592                 if (hot_attrs_checked && !bms_overlap(modified_attrs, hot_attrs))
3593                         use_hot_update = true;
3594         }
3595         else
3596         {
3597                 /* Set a hint that the old page could use prune/defrag */
3598                 PageSetFull(page);
3599         }
3600
3601         /*
3602          * Compute replica identity tuple before entering the critical section so
3603          * we don't PANIC upon a memory allocation failure.
3604          * ExtractReplicaIdentity() will return NULL if nothing needs to be
3605          * logged.
3606          */
3607         old_key_tuple = ExtractReplicaIdentity(relation, &oldtup,
3608                                                                                    bms_overlap(modified_attrs, id_attrs),
3609                                                                                    &old_key_copied);
3610
3611         /* NO EREPORT(ERROR) from here till changes are logged */
3612         START_CRIT_SECTION();
3613
3614         /*
3615          * If this transaction commits, the old tuple will become DEAD sooner or
3616          * later.  Set flag that this page is a candidate for pruning once our xid
3617          * falls below the OldestXmin horizon.  If the transaction finally aborts,
3618          * the subsequent page pruning will be a no-op and the hint will be
3619          * cleared.
3620          *
3621          * XXX Should we set hint on newbuf as well?  If the transaction aborts,
3622          * there would be a prunable tuple in the newbuf; but for now we choose
3623          * not to optimize for aborts.  Note that heap_xlog_update must be kept in
3624          * sync if this decision changes.
3625          */
3626         PageSetPrunable(page, xid);
3627
3628         if (use_hot_update)
3629         {
3630                 /* Mark the old tuple as HOT-updated */
3631                 HeapTupleSetHotUpdated(&oldtup);
3632                 /* And mark the new tuple as heap-only */
3633                 HeapTupleSetHeapOnly(heaptup);
3634                 /* Mark the caller's copy too, in case different from heaptup */
3635                 HeapTupleSetHeapOnly(newtup);
3636         }
3637         else
3638         {
3639                 /* Make sure tuples are correctly marked as not-HOT */
3640                 HeapTupleClearHotUpdated(&oldtup);
3641                 HeapTupleClearHeapOnly(heaptup);
3642                 HeapTupleClearHeapOnly(newtup);
3643         }
3644
3645         RelationPutHeapTuple(relation, newbuf, heaptup, false); /* insert new tuple */
3646
3647
3648         /* Clear obsolete visibility flags, possibly set by ourselves above... */
3649         oldtup.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
3650         oldtup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
3651         /* ... and store info about transaction updating this tuple */
3652         Assert(TransactionIdIsValid(xmax_old_tuple));
3653         HeapTupleHeaderSetXmax(oldtup.t_data, xmax_old_tuple);
3654         oldtup.t_data->t_infomask |= infomask_old_tuple;
3655         oldtup.t_data->t_infomask2 |= infomask2_old_tuple;
3656         HeapTupleHeaderSetCmax(oldtup.t_data, cid, iscombo);
3657
3658         /* record address of new tuple in t_ctid of old one */
3659         oldtup.t_data->t_ctid = heaptup->t_self;
3660
3661         /* clear PD_ALL_VISIBLE flags, reset all visibilitymap bits */
3662         if (PageIsAllVisible(BufferGetPage(buffer)))
3663         {
3664                 all_visible_cleared = true;
3665                 PageClearAllVisible(BufferGetPage(buffer));
3666                 visibilitymap_clear(relation, BufferGetBlockNumber(buffer),
3667                                                         vmbuffer, VISIBILITYMAP_VALID_BITS);
3668         }
3669         if (newbuf != buffer && PageIsAllVisible(BufferGetPage(newbuf)))
3670         {
3671                 all_visible_cleared_new = true;
3672                 PageClearAllVisible(BufferGetPage(newbuf));
3673                 visibilitymap_clear(relation, BufferGetBlockNumber(newbuf),
3674                                                         vmbuffer_new, VISIBILITYMAP_VALID_BITS);
3675         }
3676
3677         if (newbuf != buffer)
3678                 MarkBufferDirty(newbuf);
3679         MarkBufferDirty(buffer);
3680
3681         /* XLOG stuff */
3682         if (RelationNeedsWAL(relation))
3683         {
3684                 XLogRecPtr      recptr;
3685
3686                 /*
3687                  * For logical decoding we need combocids to properly decode the
3688                  * catalog.
3689                  */
3690                 if (RelationIsAccessibleInLogicalDecoding(relation))
3691                 {
3692                         log_heap_new_cid(relation, &oldtup);
3693                         log_heap_new_cid(relation, heaptup);
3694                 }
3695
3696                 recptr = log_heap_update(relation, buffer,
3697                                                                  newbuf, &oldtup, heaptup,
3698                                                                  old_key_tuple,
3699                                                                  all_visible_cleared,
3700                                                                  all_visible_cleared_new);
3701                 if (newbuf != buffer)
3702                 {
3703                         PageSetLSN(BufferGetPage(newbuf), recptr);
3704                 }
3705                 PageSetLSN(BufferGetPage(buffer), recptr);
3706         }
3707
3708         END_CRIT_SECTION();
3709
3710         if (newbuf != buffer)
3711                 LockBuffer(newbuf, BUFFER_LOCK_UNLOCK);
3712         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
3713
3714         /*
3715          * Mark old tuple for invalidation from system caches at next command
3716          * boundary, and mark the new tuple for invalidation in case we abort. We
3717          * have to do this before releasing the buffer because oldtup is in the
3718          * buffer.  (heaptup is all in local memory, but it's necessary to process
3719          * both tuple versions in one call to inval.c so we can avoid redundant
3720          * sinval messages.)
3721          */
3722         CacheInvalidateHeapTuple(relation, &oldtup, heaptup);
3723
3724         /* Now we can release the buffer(s) */
3725         if (newbuf != buffer)
3726                 ReleaseBuffer(newbuf);
3727         ReleaseBuffer(buffer);
3728         if (BufferIsValid(vmbuffer_new))
3729                 ReleaseBuffer(vmbuffer_new);
3730         if (BufferIsValid(vmbuffer))
3731                 ReleaseBuffer(vmbuffer);
3732
3733         /*
3734          * Release the lmgr tuple lock, if we had it.
3735          */
3736         if (have_tuple_lock)
3737                 UnlockTupleTuplock(relation, &(oldtup.t_self), *lockmode);
3738
3739         pgstat_count_heap_update(relation, use_hot_update);
3740
3741         /*
3742          * If heaptup is a private copy, release it.  Don't forget to copy t_self
3743          * back to the caller's image, too.
3744          */
3745         if (heaptup != newtup)
3746         {
3747                 newtup->t_self = heaptup->t_self;
3748                 heap_freetuple(heaptup);
3749         }
3750
3751         if (old_key_tuple != NULL && old_key_copied)
3752                 heap_freetuple(old_key_tuple);
3753
3754         bms_free(hot_attrs);
3755         bms_free(key_attrs);
3756         bms_free(id_attrs);
3757         bms_free(modified_attrs);
3758         bms_free(interesting_attrs);
3759
3760         return TM_Ok;
3761 }
3762
3763 /*
3764  * Check if the specified attribute's value is same in both given tuples.
3765  * Subroutine for HeapDetermineModifiedColumns.
3766  */
3767 static bool
3768 heap_tuple_attr_equals(TupleDesc tupdesc, int attrnum,
3769                                            HeapTuple tup1, HeapTuple tup2)
3770 {
3771         Datum           value1,
3772                                 value2;
3773         bool            isnull1,
3774                                 isnull2;
3775         Form_pg_attribute att;
3776
3777         /*
3778          * If it's a whole-tuple reference, say "not equal".  It's not really
3779          * worth supporting this case, since it could only succeed after a no-op
3780          * update, which is hardly a case worth optimizing for.
3781          */
3782         if (attrnum == 0)
3783                 return false;
3784
3785         /*
3786          * Likewise, automatically say "not equal" for any system attribute other
3787          * than tableOID; we cannot expect these to be consistent in a HOT chain,
3788          * or even to be set correctly yet in the new tuple.
3789          */
3790         if (attrnum < 0)
3791         {
3792                 if (attrnum != TableOidAttributeNumber)
3793                         return false;
3794         }
3795
3796         /*
3797          * Extract the corresponding values.  XXX this is pretty inefficient if
3798          * there are many indexed columns.  Should HeapDetermineModifiedColumns do
3799          * a single heap_deform_tuple call on each tuple, instead?      But that
3800          * doesn't work for system columns ...
3801          */
3802         value1 = heap_getattr(tup1, attrnum, tupdesc, &isnull1);
3803         value2 = heap_getattr(tup2, attrnum, tupdesc, &isnull2);
3804
3805         /*
3806          * If one value is NULL and other is not, then they are certainly not
3807          * equal
3808          */
3809         if (isnull1 != isnull2)
3810                 return false;
3811
3812         /*
3813          * If both are NULL, they can be considered equal.
3814          */
3815         if (isnull1)
3816                 return true;
3817
3818         /*
3819          * We do simple binary comparison of the two datums.  This may be overly
3820          * strict because there can be multiple binary representations for the
3821          * same logical value.  But we should be OK as long as there are no false
3822          * positives.  Using a type-specific equality operator is messy because
3823          * there could be multiple notions of equality in different operator
3824          * classes; furthermore, we cannot safely invoke user-defined functions
3825          * while holding exclusive buffer lock.
3826          */
3827         if (attrnum <= 0)
3828         {
3829                 /* The only allowed system columns are OIDs, so do this */
3830                 return (DatumGetObjectId(value1) == DatumGetObjectId(value2));
3831         }
3832         else
3833         {
3834                 Assert(attrnum <= tupdesc->natts);
3835                 att = TupleDescAttr(tupdesc, attrnum - 1);
3836                 return datumIsEqual(value1, value2, att->attbyval, att->attlen);
3837         }
3838 }
3839
3840 /*
3841  * Check which columns are being updated.
3842  *
3843  * Given an updated tuple, determine (and return into the output bitmapset),
3844  * from those listed as interesting, the set of columns that changed.
3845  *
3846  * The input bitmapset is destructively modified; that is OK since this is
3847  * invoked at most once in heap_update.
3848  */
3849 static Bitmapset *
3850 HeapDetermineModifiedColumns(Relation relation, Bitmapset *interesting_cols,
3851                                                          HeapTuple oldtup, HeapTuple newtup)
3852 {
3853         int                     attnum;
3854         Bitmapset  *modified = NULL;
3855
3856         while ((attnum = bms_first_member(interesting_cols)) >= 0)
3857         {
3858                 attnum += FirstLowInvalidHeapAttributeNumber;
3859
3860                 if (!heap_tuple_attr_equals(RelationGetDescr(relation),
3861                                                                         attnum, oldtup, newtup))
3862                         modified = bms_add_member(modified,
3863                                                                           attnum - FirstLowInvalidHeapAttributeNumber);
3864         }
3865
3866         return modified;
3867 }
3868
3869 /*
3870  *      simple_heap_update - replace a tuple
3871  *
3872  * This routine may be used to update a tuple when concurrent updates of
3873  * the target tuple are not expected (for example, because we have a lock
3874  * on the relation associated with the tuple).  Any failure is reported
3875  * via ereport().
3876  */
3877 void
3878 simple_heap_update(Relation relation, ItemPointer otid, HeapTuple tup)
3879 {
3880         TM_Result       result;
3881         TM_FailureData tmfd;
3882         LockTupleMode lockmode;
3883
3884         result = heap_update(relation, otid, tup,
3885                                                  GetCurrentCommandId(true), InvalidSnapshot,
3886                                                  true /* wait for commit */ ,
3887                                                  &tmfd, &lockmode);
3888         switch (result)
3889         {
3890                 case TM_SelfModified:
3891                         /* Tuple was already updated in current command? */
3892                         elog(ERROR, "tuple already updated by self");
3893                         break;
3894
3895                 case TM_Ok:
3896                         /* done successfully */
3897                         break;
3898
3899                 case TM_Updated:
3900                         elog(ERROR, "tuple concurrently updated");
3901                         break;
3902
3903                 case TM_Deleted:
3904                         elog(ERROR, "tuple concurrently deleted");
3905                         break;
3906
3907                 default:
3908                         elog(ERROR, "unrecognized heap_update status: %u", result);
3909                         break;
3910         }
3911 }
3912
3913
3914 /*
3915  * Return the MultiXactStatus corresponding to the given tuple lock mode.
3916  */
3917 static MultiXactStatus
3918 get_mxact_status_for_lock(LockTupleMode mode, bool is_update)
3919 {
3920         int                     retval;
3921
3922         if (is_update)
3923                 retval = tupleLockExtraInfo[mode].updstatus;
3924         else
3925                 retval = tupleLockExtraInfo[mode].lockstatus;
3926
3927         if (retval == -1)
3928                 elog(ERROR, "invalid lock tuple mode %d/%s", mode,
3929                          is_update ? "true" : "false");
3930
3931         return (MultiXactStatus) retval;
3932 }
3933
3934 /*
3935  *      heap_lock_tuple - lock a tuple in shared or exclusive mode
3936  *
3937  * Note that this acquires a buffer pin, which the caller must release.
3938  *
3939  * Input parameters:
3940  *      relation: relation containing tuple (caller must hold suitable lock)
3941  *      tid: TID of tuple to lock
3942  *      cid: current command ID (used for visibility test, and stored into
3943  *              tuple's cmax if lock is successful)
3944  *      mode: indicates if shared or exclusive tuple lock is desired
3945  *      wait_policy: what to do if tuple lock is not available
3946  *      follow_updates: if true, follow the update chain to also lock descendant
3947  *              tuples.
3948  *
3949  * Output parameters:
3950  *      *tuple: all fields filled in
3951  *      *buffer: set to buffer holding tuple (pinned but not locked at exit)
3952  *      *tmfd: filled in failure cases (see below)
3953  *
3954  * Function results are the same as the ones for table_lock_tuple().
3955  *
3956  * In the failure cases other than TM_Invisible, the routine fills
3957  * *tmfd with the tuple's t_ctid, t_xmax (resolving a possible MultiXact,
3958  * if necessary), and t_cmax (the last only for TM_SelfModified,
3959  * since we cannot obtain cmax from a combocid generated by another
3960  * transaction).
3961  * See comments for struct TM_FailureData for additional info.
3962  *
3963  * See README.tuplock for a thorough explanation of this mechanism.
3964  */
3965 TM_Result
3966 heap_lock_tuple(Relation relation, HeapTuple tuple,
3967                                 CommandId cid, LockTupleMode mode, LockWaitPolicy wait_policy,
3968                                 bool follow_updates,
3969                                 Buffer *buffer, TM_FailureData *tmfd)
3970 {
3971         TM_Result       result;
3972         ItemPointer tid = &(tuple->t_self);
3973         ItemId          lp;
3974         Page            page;
3975         Buffer          vmbuffer = InvalidBuffer;
3976         BlockNumber block;
3977         TransactionId xid,
3978                                 xmax;
3979         uint16          old_infomask,
3980                                 new_infomask,
3981                                 new_infomask2;
3982         bool            first_time = true;
3983         bool            have_tuple_lock = false;
3984         bool            cleared_all_frozen = false;
3985
3986         *buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
3987         block = ItemPointerGetBlockNumber(tid);
3988
3989         /*
3990          * Before locking the buffer, pin the visibility map page if it appears to
3991          * be necessary.  Since we haven't got the lock yet, someone else might be
3992          * in the middle of changing this, so we'll need to recheck after we have
3993          * the lock.
3994          */
3995         if (PageIsAllVisible(BufferGetPage(*buffer)))
3996                 visibilitymap_pin(relation, block, &vmbuffer);
3997
3998         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
3999
4000         page = BufferGetPage(*buffer);
4001         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
4002         Assert(ItemIdIsNormal(lp));
4003
4004         tuple->t_data = (HeapTupleHeader) PageGetItem(page, lp);
4005         tuple->t_len = ItemIdGetLength(lp);
4006         tuple->t_tableOid = RelationGetRelid(relation);
4007
4008 l3:
4009         result = HeapTupleSatisfiesUpdate(tuple, cid, *buffer);
4010
4011         if (result == TM_Invisible)
4012         {
4013                 /*
4014                  * This is possible, but only when locking a tuple for ON CONFLICT
4015                  * UPDATE.  We return this value here rather than throwing an error in
4016                  * order to give that case the opportunity to throw a more specific
4017                  * error.
4018                  */
4019                 result = TM_Invisible;
4020                 goto out_locked;
4021         }
4022         else if (result == TM_BeingModified ||
4023                          result == TM_Updated ||
4024                          result == TM_Deleted)
4025         {
4026                 TransactionId xwait;
4027                 uint16          infomask;
4028                 uint16          infomask2;
4029                 bool            require_sleep;
4030                 ItemPointerData t_ctid;
4031
4032                 /* must copy state data before unlocking buffer */
4033                 xwait = HeapTupleHeaderGetRawXmax(tuple->t_data);
4034                 infomask = tuple->t_data->t_infomask;
4035                 infomask2 = tuple->t_data->t_infomask2;
4036                 ItemPointerCopy(&tuple->t_data->t_ctid, &t_ctid);
4037
4038                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4039
4040                 /*
4041                  * If any subtransaction of the current top transaction already holds
4042                  * a lock as strong as or stronger than what we're requesting, we
4043                  * effectively hold the desired lock already.  We *must* succeed
4044                  * without trying to take the tuple lock, else we will deadlock
4045                  * against anyone wanting to acquire a stronger lock.
4046                  *
4047                  * Note we only do this the first time we loop on the HTSU result;
4048                  * there is no point in testing in subsequent passes, because
4049                  * evidently our own transaction cannot have acquired a new lock after
4050                  * the first time we checked.
4051                  */
4052                 if (first_time)
4053                 {
4054                         first_time = false;
4055
4056                         if (infomask & HEAP_XMAX_IS_MULTI)
4057                         {
4058                                 int                     i;
4059                                 int                     nmembers;
4060                                 MultiXactMember *members;
4061
4062                                 /*
4063                                  * We don't need to allow old multixacts here; if that had
4064                                  * been the case, HeapTupleSatisfiesUpdate would have returned
4065                                  * MayBeUpdated and we wouldn't be here.
4066                                  */
4067                                 nmembers =
4068                                         GetMultiXactIdMembers(xwait, &members, false,
4069                                                                                   HEAP_XMAX_IS_LOCKED_ONLY(infomask));
4070
4071                                 for (i = 0; i < nmembers; i++)
4072                                 {
4073                                         /* only consider members of our own transaction */
4074                                         if (!TransactionIdIsCurrentTransactionId(members[i].xid))
4075                                                 continue;
4076
4077                                         if (TUPLOCK_from_mxstatus(members[i].status) >= mode)
4078                                         {
4079                                                 pfree(members);
4080                                                 result = TM_Ok;
4081                                                 goto out_unlocked;
4082                                         }
4083                                 }
4084
4085                                 if (members)
4086                                         pfree(members);
4087                         }
4088                         else if (TransactionIdIsCurrentTransactionId(xwait))
4089                         {
4090                                 switch (mode)
4091                                 {
4092                                         case LockTupleKeyShare:
4093                                                 Assert(HEAP_XMAX_IS_KEYSHR_LOCKED(infomask) ||
4094                                                            HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4095                                                            HEAP_XMAX_IS_EXCL_LOCKED(infomask));
4096                                                 result = TM_Ok;
4097                                                 goto out_unlocked;
4098                                         case LockTupleShare:
4099                                                 if (HEAP_XMAX_IS_SHR_LOCKED(infomask) ||
4100                                                         HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4101                                                 {
4102                                                         result = TM_Ok;
4103                                                         goto out_unlocked;
4104                                                 }
4105                                                 break;
4106                                         case LockTupleNoKeyExclusive:
4107                                                 if (HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4108                                                 {
4109                                                         result = TM_Ok;
4110                                                         goto out_unlocked;
4111                                                 }
4112                                                 break;
4113                                         case LockTupleExclusive:
4114                                                 if (HEAP_XMAX_IS_EXCL_LOCKED(infomask) &&
4115                                                         infomask2 & HEAP_KEYS_UPDATED)
4116                                                 {
4117                                                         result = TM_Ok;
4118                                                         goto out_unlocked;
4119                                                 }
4120                                                 break;
4121                                 }
4122                         }
4123                 }
4124
4125                 /*
4126                  * Initially assume that we will have to wait for the locking
4127                  * transaction(s) to finish.  We check various cases below in which
4128                  * this can be turned off.
4129                  */
4130                 require_sleep = true;
4131                 if (mode == LockTupleKeyShare)
4132                 {
4133                         /*
4134                          * If we're requesting KeyShare, and there's no update present, we
4135                          * don't need to wait.  Even if there is an update, we can still
4136                          * continue if the key hasn't been modified.
4137                          *
4138                          * However, if there are updates, we need to walk the update chain
4139                          * to mark future versions of the row as locked, too.  That way,
4140                          * if somebody deletes that future version, we're protected
4141                          * against the key going away.  This locking of future versions
4142                          * could block momentarily, if a concurrent transaction is
4143                          * deleting a key; or it could return a value to the effect that
4144                          * the transaction deleting the key has already committed.  So we
4145                          * do this before re-locking the buffer; otherwise this would be
4146                          * prone to deadlocks.
4147                          *
4148                          * Note that the TID we're locking was grabbed before we unlocked
4149                          * the buffer.  For it to change while we're not looking, the
4150                          * other properties we're testing for below after re-locking the
4151                          * buffer would also change, in which case we would restart this
4152                          * loop above.
4153                          */
4154                         if (!(infomask2 & HEAP_KEYS_UPDATED))
4155                         {
4156                                 bool            updated;
4157
4158                                 updated = !HEAP_XMAX_IS_LOCKED_ONLY(infomask);
4159
4160                                 /*
4161                                  * If there are updates, follow the update chain; bail out if
4162                                  * that cannot be done.
4163                                  */
4164                                 if (follow_updates && updated)
4165                                 {
4166                                         TM_Result       res;
4167
4168                                         res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4169                                                                                                   GetCurrentTransactionId(),
4170                                                                                                   mode);
4171                                         if (res != TM_Ok)
4172                                         {
4173                                                 result = res;
4174                                                 /* recovery code expects to have buffer lock held */
4175                                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4176                                                 goto failed;
4177                                         }
4178                                 }
4179
4180                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4181
4182                                 /*
4183                                  * Make sure it's still an appropriate lock, else start over.
4184                                  * Also, if it wasn't updated before we released the lock, but
4185                                  * is updated now, we start over too; the reason is that we
4186                                  * now need to follow the update chain to lock the new
4187                                  * versions.
4188                                  */
4189                                 if (!HeapTupleHeaderIsOnlyLocked(tuple->t_data) &&
4190                                         ((tuple->t_data->t_infomask2 & HEAP_KEYS_UPDATED) ||
4191                                          !updated))
4192                                         goto l3;
4193
4194                                 /* Things look okay, so we can skip sleeping */
4195                                 require_sleep = false;
4196
4197                                 /*
4198                                  * Note we allow Xmax to change here; other updaters/lockers
4199                                  * could have modified it before we grabbed the buffer lock.
4200                                  * However, this is not a problem, because with the recheck we
4201                                  * just did we ensure that they still don't conflict with the
4202                                  * lock we want.
4203                                  */
4204                         }
4205                 }
4206                 else if (mode == LockTupleShare)
4207                 {
4208                         /*
4209                          * If we're requesting Share, we can similarly avoid sleeping if
4210                          * there's no update and no exclusive lock present.
4211                          */
4212                         if (HEAP_XMAX_IS_LOCKED_ONLY(infomask) &&
4213                                 !HEAP_XMAX_IS_EXCL_LOCKED(infomask))
4214                         {
4215                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4216
4217                                 /*
4218                                  * Make sure it's still an appropriate lock, else start over.
4219                                  * See above about allowing xmax to change.
4220                                  */
4221                                 if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4222                                         HEAP_XMAX_IS_EXCL_LOCKED(tuple->t_data->t_infomask))
4223                                         goto l3;
4224                                 require_sleep = false;
4225                         }
4226                 }
4227                 else if (mode == LockTupleNoKeyExclusive)
4228                 {
4229                         /*
4230                          * If we're requesting NoKeyExclusive, we might also be able to
4231                          * avoid sleeping; just ensure that there no conflicting lock
4232                          * already acquired.
4233                          */
4234                         if (infomask & HEAP_XMAX_IS_MULTI)
4235                         {
4236                                 if (!DoesMultiXactIdConflict((MultiXactId) xwait, infomask,
4237                                                                                          mode))
4238                                 {
4239                                         /*
4240                                          * No conflict, but if the xmax changed under us in the
4241                                          * meantime, start over.
4242                                          */
4243                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4244                                         if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4245                                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4246                                                                                          xwait))
4247                                                 goto l3;
4248
4249                                         /* otherwise, we're good */
4250                                         require_sleep = false;
4251                                 }
4252                         }
4253                         else if (HEAP_XMAX_IS_KEYSHR_LOCKED(infomask))
4254                         {
4255                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4256
4257                                 /* if the xmax changed in the meantime, start over */
4258                                 if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4259                                         !TransactionIdEquals(
4260                                                                                  HeapTupleHeaderGetRawXmax(tuple->t_data),
4261                                                                                  xwait))
4262                                         goto l3;
4263                                 /* otherwise, we're good */
4264                                 require_sleep = false;
4265                         }
4266                 }
4267
4268                 /*
4269                  * As a check independent from those above, we can also avoid sleeping
4270                  * if the current transaction is the sole locker of the tuple.  Note
4271                  * that the strength of the lock already held is irrelevant; this is
4272                  * not about recording the lock in Xmax (which will be done regardless
4273                  * of this optimization, below).  Also, note that the cases where we
4274                  * hold a lock stronger than we are requesting are already handled
4275                  * above by not doing anything.
4276                  *
4277                  * Note we only deal with the non-multixact case here; MultiXactIdWait
4278                  * is well equipped to deal with this situation on its own.
4279                  */
4280                 if (require_sleep && !(infomask & HEAP_XMAX_IS_MULTI) &&
4281                         TransactionIdIsCurrentTransactionId(xwait))
4282                 {
4283                         /* ... but if the xmax changed in the meantime, start over */
4284                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4285                         if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4286                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4287                                                                          xwait))
4288                                 goto l3;
4289                         Assert(HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask));
4290                         require_sleep = false;
4291                 }
4292
4293                 /*
4294                  * Time to sleep on the other transaction/multixact, if necessary.
4295                  *
4296                  * If the other transaction is an update/delete that's already
4297                  * committed, then sleeping cannot possibly do any good: if we're
4298                  * required to sleep, get out to raise an error instead.
4299                  *
4300                  * By here, we either have already acquired the buffer exclusive lock,
4301                  * or we must wait for the locking transaction or multixact; so below
4302                  * we ensure that we grab buffer lock after the sleep.
4303                  */
4304                 if (require_sleep && (result == TM_Updated || result == TM_Deleted))
4305                 {
4306                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4307                         goto failed;
4308                 }
4309                 else if (require_sleep)
4310                 {
4311                         /*
4312                          * Acquire tuple lock to establish our priority for the tuple, or
4313                          * die trying.  LockTuple will release us when we are next-in-line
4314                          * for the tuple.  We must do this even if we are share-locking.
4315                          *
4316                          * If we are forced to "start over" below, we keep the tuple lock;
4317                          * this arranges that we stay at the head of the line while
4318                          * rechecking tuple state.
4319                          */
4320                         if (!heap_acquire_tuplock(relation, tid, mode, wait_policy,
4321                                                                           &have_tuple_lock))
4322                         {
4323                                 /*
4324                                  * This can only happen if wait_policy is Skip and the lock
4325                                  * couldn't be obtained.
4326                                  */
4327                                 result = TM_WouldBlock;
4328                                 /* recovery code expects to have buffer lock held */
4329                                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4330                                 goto failed;
4331                         }
4332
4333                         if (infomask & HEAP_XMAX_IS_MULTI)
4334                         {
4335                                 MultiXactStatus status = get_mxact_status_for_lock(mode, false);
4336
4337                                 /* We only ever lock tuples, never update them */
4338                                 if (status >= MultiXactStatusNoKeyUpdate)
4339                                         elog(ERROR, "invalid lock mode in heap_lock_tuple");
4340
4341                                 /* wait for multixact to end, or die trying  */
4342                                 switch (wait_policy)
4343                                 {
4344                                         case LockWaitBlock:
4345                                                 MultiXactIdWait((MultiXactId) xwait, status, infomask,
4346                                                                                 relation, &tuple->t_self, XLTW_Lock, NULL);
4347                                                 break;
4348                                         case LockWaitSkip:
4349                                                 if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4350                                                                                                                 status, infomask, relation,
4351                                                                                                                 NULL))
4352                                                 {
4353                                                         result = TM_WouldBlock;
4354                                                         /* recovery code expects to have buffer lock held */
4355                                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4356                                                         goto failed;
4357                                                 }
4358                                                 break;
4359                                         case LockWaitError:
4360                                                 if (!ConditionalMultiXactIdWait((MultiXactId) xwait,
4361                                                                                                                 status, infomask, relation,
4362                                                                                                                 NULL))
4363                                                         ereport(ERROR,
4364                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4365                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
4366                                                                                         RelationGetRelationName(relation))));
4367
4368                                                 break;
4369                                 }
4370
4371                                 /*
4372                                  * Of course, the multixact might not be done here: if we're
4373                                  * requesting a light lock mode, other transactions with light
4374                                  * locks could still be alive, as well as locks owned by our
4375                                  * own xact or other subxacts of this backend.  We need to
4376                                  * preserve the surviving MultiXact members.  Note that it
4377                                  * isn't absolutely necessary in the latter case, but doing so
4378                                  * is simpler.
4379                                  */
4380                         }
4381                         else
4382                         {
4383                                 /* wait for regular transaction to end, or die trying */
4384                                 switch (wait_policy)
4385                                 {
4386                                         case LockWaitBlock:
4387                                                 XactLockTableWait(xwait, relation, &tuple->t_self,
4388                                                                                   XLTW_Lock);
4389                                                 break;
4390                                         case LockWaitSkip:
4391                                                 if (!ConditionalXactLockTableWait(xwait))
4392                                                 {
4393                                                         result = TM_WouldBlock;
4394                                                         /* recovery code expects to have buffer lock held */
4395                                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4396                                                         goto failed;
4397                                                 }
4398                                                 break;
4399                                         case LockWaitError:
4400                                                 if (!ConditionalXactLockTableWait(xwait))
4401                                                         ereport(ERROR,
4402                                                                         (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4403                                                                          errmsg("could not obtain lock on row in relation \"%s\"",
4404                                                                                         RelationGetRelationName(relation))));
4405                                                 break;
4406                                 }
4407                         }
4408
4409                         /* if there are updates, follow the update chain */
4410                         if (follow_updates && !HEAP_XMAX_IS_LOCKED_ONLY(infomask))
4411                         {
4412                                 TM_Result       res;
4413
4414                                 res = heap_lock_updated_tuple(relation, tuple, &t_ctid,
4415                                                                                           GetCurrentTransactionId(),
4416                                                                                           mode);
4417                                 if (res != TM_Ok)
4418                                 {
4419                                         result = res;
4420                                         /* recovery code expects to have buffer lock held */
4421                                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4422                                         goto failed;
4423                                 }
4424                         }
4425
4426                         LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4427
4428                         /*
4429                          * xwait is done, but if xwait had just locked the tuple then some
4430                          * other xact could update this tuple before we get to this point.
4431                          * Check for xmax change, and start over if so.
4432                          */
4433                         if (xmax_infomask_changed(tuple->t_data->t_infomask, infomask) ||
4434                                 !TransactionIdEquals(HeapTupleHeaderGetRawXmax(tuple->t_data),
4435                                                                          xwait))
4436                                 goto l3;
4437
4438                         if (!(infomask & HEAP_XMAX_IS_MULTI))
4439                         {
4440                                 /*
4441                                  * Otherwise check if it committed or aborted.  Note we cannot
4442                                  * be here if the tuple was only locked by somebody who didn't
4443                                  * conflict with us; that would have been handled above.  So
4444                                  * that transaction must necessarily be gone by now.  But
4445                                  * don't check for this in the multixact case, because some
4446                                  * locker transactions might still be running.
4447                                  */
4448                                 UpdateXmaxHintBits(tuple->t_data, *buffer, xwait);
4449                         }
4450                 }
4451
4452                 /* By here, we're certain that we hold buffer exclusive lock again */
4453
4454                 /*
4455                  * We may lock if previous xmax aborted, or if it committed but only
4456                  * locked the tuple without updating it; or if we didn't have to wait
4457                  * at all for whatever reason.
4458                  */
4459                 if (!require_sleep ||
4460                         (tuple->t_data->t_infomask & HEAP_XMAX_INVALID) ||
4461                         HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_data->t_infomask) ||
4462                         HeapTupleHeaderIsOnlyLocked(tuple->t_data))
4463                         result = TM_Ok;
4464                 else if (!ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid) ||
4465                                  HeapTupleHeaderIndicatesMovedPartitions(tuple->t_data))
4466                         result = TM_Updated;
4467                 else
4468                         result = TM_Deleted;
4469         }
4470
4471 failed:
4472         if (result != TM_Ok)
4473         {
4474                 Assert(result == TM_SelfModified || result == TM_Updated ||
4475                            result == TM_Deleted || result == TM_WouldBlock);
4476                 Assert(!(tuple->t_data->t_infomask & HEAP_XMAX_INVALID));
4477                 Assert(result != TM_Updated ||
4478                            !ItemPointerEquals(&tuple->t_self, &tuple->t_data->t_ctid));
4479                 tmfd->ctid = tuple->t_data->t_ctid;
4480                 tmfd->xmax = HeapTupleHeaderGetUpdateXid(tuple->t_data);
4481                 if (result == TM_SelfModified)
4482                         tmfd->cmax = HeapTupleHeaderGetCmax(tuple->t_data);
4483                 else
4484                         tmfd->cmax = InvalidCommandId;
4485                 goto out_locked;
4486         }
4487
4488         /*
4489          * If we didn't pin the visibility map page and the page has become all
4490          * visible while we were busy locking the buffer, or during some
4491          * subsequent window during which we had it unlocked, we'll have to unlock
4492          * and re-lock, to avoid holding the buffer lock across I/O.  That's a bit
4493          * unfortunate, especially since we'll now have to recheck whether the
4494          * tuple has been locked or updated under us, but hopefully it won't
4495          * happen very often.
4496          */
4497         if (vmbuffer == InvalidBuffer && PageIsAllVisible(page))
4498         {
4499                 LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4500                 visibilitymap_pin(relation, block, &vmbuffer);
4501                 LockBuffer(*buffer, BUFFER_LOCK_EXCLUSIVE);
4502                 goto l3;
4503         }
4504
4505         xmax = HeapTupleHeaderGetRawXmax(tuple->t_data);
4506         old_infomask = tuple->t_data->t_infomask;
4507
4508         /*
4509          * If this is the first possibly-multixact-able operation in the current
4510          * transaction, set my per-backend OldestMemberMXactId setting. We can be
4511          * certain that the transaction will never become a member of any older
4512          * MultiXactIds than that.  (We have to do this even if we end up just
4513          * using our own TransactionId below, since some other backend could
4514          * incorporate our XID into a MultiXact immediately afterwards.)
4515          */
4516         MultiXactIdSetOldestMember();
4517
4518         /*
4519          * Compute the new xmax and infomask to store into the tuple.  Note we do
4520          * not modify the tuple just yet, because that would leave it in the wrong
4521          * state if multixact.c elogs.
4522          */
4523         compute_new_xmax_infomask(xmax, old_infomask, tuple->t_data->t_infomask2,
4524                                                           GetCurrentTransactionId(), mode, false,
4525                                                           &xid, &new_infomask, &new_infomask2);
4526
4527         START_CRIT_SECTION();
4528
4529         /*
4530          * Store transaction information of xact locking the tuple.
4531          *
4532          * Note: Cmax is meaningless in this context, so don't set it; this avoids
4533          * possibly generating a useless combo CID.  Moreover, if we're locking a
4534          * previously updated tuple, it's important to preserve the Cmax.
4535          *
4536          * Also reset the HOT UPDATE bit, but only if there's no update; otherwise
4537          * we would break the HOT chain.
4538          */
4539         tuple->t_data->t_infomask &= ~HEAP_XMAX_BITS;
4540         tuple->t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
4541         tuple->t_data->t_infomask |= new_infomask;
4542         tuple->t_data->t_infomask2 |= new_infomask2;
4543         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
4544                 HeapTupleHeaderClearHotUpdated(tuple->t_data);
4545         HeapTupleHeaderSetXmax(tuple->t_data, xid);
4546
4547         /*
4548          * Make sure there is no forward chain link in t_ctid.  Note that in the
4549          * cases where the tuple has been updated, we must not overwrite t_ctid,
4550          * because it was set by the updater.  Moreover, if the tuple has been
4551          * updated, we need to follow the update chain to lock the new versions of
4552          * the tuple as well.
4553          */
4554         if (HEAP_XMAX_IS_LOCKED_ONLY(new_infomask))
4555                 tuple->t_data->t_ctid = *tid;
4556
4557         /* Clear only the all-frozen bit on visibility map if needed */
4558         if (PageIsAllVisible(page) &&
4559                 visibilitymap_clear(relation, block, vmbuffer,
4560                                                         VISIBILITYMAP_ALL_FROZEN))
4561                 cleared_all_frozen = true;
4562
4563
4564         MarkBufferDirty(*buffer);
4565
4566         /*
4567          * XLOG stuff.  You might think that we don't need an XLOG record because
4568          * there is no state change worth restoring after a crash.  You would be
4569          * wrong however: we have just written either a TransactionId or a
4570          * MultiXactId that may never have been seen on disk before, and we need
4571          * to make sure that there are XLOG entries covering those ID numbers.
4572          * Else the same IDs might be re-used after a crash, which would be
4573          * disastrous if this page made it to disk before the crash.  Essentially
4574          * we have to enforce the WAL log-before-data rule even in this case.
4575          * (Also, in a PITR log-shipping or 2PC environment, we have to have XLOG
4576          * entries for everything anyway.)
4577          */
4578         if (RelationNeedsWAL(relation))
4579         {
4580                 xl_heap_lock xlrec;
4581                 XLogRecPtr      recptr;
4582
4583                 XLogBeginInsert();
4584                 XLogRegisterBuffer(0, *buffer, REGBUF_STANDARD);
4585
4586                 xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
4587                 xlrec.locking_xid = xid;
4588                 xlrec.infobits_set = compute_infobits(new_infomask,
4589                                                                                           tuple->t_data->t_infomask2);
4590                 xlrec.flags = cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
4591                 XLogRegisterData((char *) &xlrec, SizeOfHeapLock);
4592
4593                 /* we don't decode row locks atm, so no need to log the origin */
4594
4595                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_LOCK);
4596
4597                 PageSetLSN(page, recptr);
4598         }
4599
4600         END_CRIT_SECTION();
4601
4602         result = TM_Ok;
4603
4604 out_locked:
4605         LockBuffer(*buffer, BUFFER_LOCK_UNLOCK);
4606
4607 out_unlocked:
4608         if (BufferIsValid(vmbuffer))
4609                 ReleaseBuffer(vmbuffer);
4610
4611         /*
4612          * Don't update the visibility map here. Locking a tuple doesn't change
4613          * visibility info.
4614          */
4615
4616         /*
4617          * Now that we have successfully marked the tuple as locked, we can
4618          * release the lmgr tuple lock, if we had it.
4619          */
4620         if (have_tuple_lock)
4621                 UnlockTupleTuplock(relation, tid, mode);
4622
4623         return result;
4624 }
4625
4626 /*
4627  * Acquire heavyweight lock on the given tuple, in preparation for acquiring
4628  * its normal, Xmax-based tuple lock.
4629  *
4630  * have_tuple_lock is an input and output parameter: on input, it indicates
4631  * whether the lock has previously been acquired (and this function does
4632  * nothing in that case).  If this function returns success, have_tuple_lock
4633  * has been flipped to true.
4634  *
4635  * Returns false if it was unable to obtain the lock; this can only happen if
4636  * wait_policy is Skip.
4637  */
4638 static bool
4639 heap_acquire_tuplock(Relation relation, ItemPointer tid, LockTupleMode mode,
4640                                          LockWaitPolicy wait_policy, bool *have_tuple_lock)
4641 {
4642         if (*have_tuple_lock)
4643                 return true;
4644
4645         switch (wait_policy)
4646         {
4647                 case LockWaitBlock:
4648                         LockTupleTuplock(relation, tid, mode);
4649                         break;
4650
4651                 case LockWaitSkip:
4652                         if (!ConditionalLockTupleTuplock(relation, tid, mode))
4653                                 return false;
4654                         break;
4655
4656                 case LockWaitError:
4657                         if (!ConditionalLockTupleTuplock(relation, tid, mode))
4658                                 ereport(ERROR,
4659                                                 (errcode(ERRCODE_LOCK_NOT_AVAILABLE),
4660                                                  errmsg("could not obtain lock on row in relation \"%s\"",
4661                                                                 RelationGetRelationName(relation))));
4662                         break;
4663         }
4664         *have_tuple_lock = true;
4665
4666         return true;
4667 }
4668
4669 /*
4670  * Given an original set of Xmax and infomask, and a transaction (identified by
4671  * add_to_xmax) acquiring a new lock of some mode, compute the new Xmax and
4672  * corresponding infomasks to use on the tuple.
4673  *
4674  * Note that this might have side effects such as creating a new MultiXactId.
4675  *
4676  * Most callers will have called HeapTupleSatisfiesUpdate before this function;
4677  * that will have set the HEAP_XMAX_INVALID bit if the xmax was a MultiXactId
4678  * but it was not running anymore. There is a race condition, which is that the
4679  * MultiXactId may have finished since then, but that uncommon case is handled
4680  * either here, or within MultiXactIdExpand.
4681  *
4682  * There is a similar race condition possible when the old xmax was a regular
4683  * TransactionId.  We test TransactionIdIsInProgress again just to narrow the
4684  * window, but it's still possible to end up creating an unnecessary
4685  * MultiXactId.  Fortunately this is harmless.
4686  */
4687 static void
4688 compute_new_xmax_infomask(TransactionId xmax, uint16 old_infomask,
4689                                                   uint16 old_infomask2, TransactionId add_to_xmax,
4690                                                   LockTupleMode mode, bool is_update,
4691                                                   TransactionId *result_xmax, uint16 *result_infomask,
4692                                                   uint16 *result_infomask2)
4693 {
4694         TransactionId new_xmax;
4695         uint16          new_infomask,
4696                                 new_infomask2;
4697
4698         Assert(TransactionIdIsCurrentTransactionId(add_to_xmax));
4699
4700 l5:
4701         new_infomask = 0;
4702         new_infomask2 = 0;
4703         if (old_infomask & HEAP_XMAX_INVALID)
4704         {
4705                 /*
4706                  * No previous locker; we just insert our own TransactionId.
4707                  *
4708                  * Note that it's critical that this case be the first one checked,
4709                  * because there are several blocks below that come back to this one
4710                  * to implement certain optimizations; old_infomask might contain
4711                  * other dirty bits in those cases, but we don't really care.
4712                  */
4713                 if (is_update)
4714                 {
4715                         new_xmax = add_to_xmax;
4716                         if (mode == LockTupleExclusive)
4717                                 new_infomask2 |= HEAP_KEYS_UPDATED;
4718                 }
4719                 else
4720                 {
4721                         new_infomask |= HEAP_XMAX_LOCK_ONLY;
4722                         switch (mode)
4723                         {
4724                                 case LockTupleKeyShare:
4725                                         new_xmax = add_to_xmax;
4726                                         new_infomask |= HEAP_XMAX_KEYSHR_LOCK;
4727                                         break;
4728                                 case LockTupleShare:
4729                                         new_xmax = add_to_xmax;
4730                                         new_infomask |= HEAP_XMAX_SHR_LOCK;
4731                                         break;
4732                                 case LockTupleNoKeyExclusive:
4733                                         new_xmax = add_to_xmax;
4734                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
4735                                         break;
4736                                 case LockTupleExclusive:
4737                                         new_xmax = add_to_xmax;
4738                                         new_infomask |= HEAP_XMAX_EXCL_LOCK;
4739                                         new_infomask2 |= HEAP_KEYS_UPDATED;
4740                                         break;
4741                                 default:
4742                                         new_xmax = InvalidTransactionId;        /* silence compiler */
4743                                         elog(ERROR, "invalid lock mode");
4744                         }
4745                 }
4746         }
4747         else if (old_infomask & HEAP_XMAX_IS_MULTI)
4748         {
4749                 MultiXactStatus new_status;
4750
4751                 /*
4752                  * Currently we don't allow XMAX_COMMITTED to be set for multis, so
4753                  * cross-check.
4754                  */
4755                 Assert(!(old_infomask & HEAP_XMAX_COMMITTED));
4756
4757                 /*
4758                  * A multixact together with LOCK_ONLY set but neither lock bit set
4759                  * (i.e. a pg_upgraded share locked tuple) cannot possibly be running
4760                  * anymore.  This check is critical for databases upgraded by
4761                  * pg_upgrade; both MultiXactIdIsRunning and MultiXactIdExpand assume
4762                  * that such multis are never passed.
4763                  */
4764                 if (HEAP_LOCKED_UPGRADED(old_infomask))
4765                 {
4766                         old_infomask &= ~HEAP_XMAX_IS_MULTI;
4767                         old_infomask |= HEAP_XMAX_INVALID;
4768                         goto l5;
4769                 }
4770
4771                 /*
4772                  * If the XMAX is already a MultiXactId, then we need to expand it to
4773                  * include add_to_xmax; but if all the members were lockers and are
4774                  * all gone, we can do away with the IS_MULTI bit and just set
4775                  * add_to_xmax as the only locker/updater.  If all lockers are gone
4776                  * and we have an updater that aborted, we can also do without a
4777                  * multi.
4778                  *
4779                  * The cost of doing GetMultiXactIdMembers would be paid by
4780                  * MultiXactIdExpand if we weren't to do this, so this check is not
4781                  * incurring extra work anyhow.
4782                  */
4783                 if (!MultiXactIdIsRunning(xmax, HEAP_XMAX_IS_LOCKED_ONLY(old_infomask)))
4784                 {
4785                         if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) ||
4786                                 !TransactionIdDidCommit(MultiXactIdGetUpdateXid(xmax,
4787                                                                                                                                 old_infomask)))
4788                         {
4789                                 /*
4790                                  * Reset these bits and restart; otherwise fall through to
4791                                  * create a new multi below.
4792                                  */
4793                                 old_infomask &= ~HEAP_XMAX_IS_MULTI;
4794                                 old_infomask |= HEAP_XMAX_INVALID;
4795                                 goto l5;
4796                         }
4797                 }
4798
4799                 new_status = get_mxact_status_for_lock(mode, is_update);
4800
4801                 new_xmax = MultiXactIdExpand((MultiXactId) xmax, add_to_xmax,
4802                                                                          new_status);
4803                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4804         }
4805         else if (old_infomask & HEAP_XMAX_COMMITTED)
4806         {
4807                 /*
4808                  * It's a committed update, so we need to preserve him as updater of
4809                  * the tuple.
4810                  */
4811                 MultiXactStatus status;
4812                 MultiXactStatus new_status;
4813
4814                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4815                         status = MultiXactStatusUpdate;
4816                 else
4817                         status = MultiXactStatusNoKeyUpdate;
4818
4819                 new_status = get_mxact_status_for_lock(mode, is_update);
4820
4821                 /*
4822                  * since it's not running, it's obviously impossible for the old
4823                  * updater to be identical to the current one, so we need not check
4824                  * for that case as we do in the block above.
4825                  */
4826                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
4827                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4828         }
4829         else if (TransactionIdIsInProgress(xmax))
4830         {
4831                 /*
4832                  * If the XMAX is a valid, in-progress TransactionId, then we need to
4833                  * create a new MultiXactId that includes both the old locker or
4834                  * updater and our own TransactionId.
4835                  */
4836                 MultiXactStatus new_status;
4837                 MultiXactStatus old_status;
4838                 LockTupleMode old_mode;
4839
4840                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
4841                 {
4842                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
4843                                 old_status = MultiXactStatusForKeyShare;
4844                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
4845                                 old_status = MultiXactStatusForShare;
4846                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
4847                         {
4848                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4849                                         old_status = MultiXactStatusForUpdate;
4850                                 else
4851                                         old_status = MultiXactStatusForNoKeyUpdate;
4852                         }
4853                         else
4854                         {
4855                                 /*
4856                                  * LOCK_ONLY can be present alone only when a page has been
4857                                  * upgraded by pg_upgrade.  But in that case,
4858                                  * TransactionIdIsInProgress() should have returned false.  We
4859                                  * assume it's no longer locked in this case.
4860                                  */
4861                                 elog(WARNING, "LOCK_ONLY found for Xid in progress %u", xmax);
4862                                 old_infomask |= HEAP_XMAX_INVALID;
4863                                 old_infomask &= ~HEAP_XMAX_LOCK_ONLY;
4864                                 goto l5;
4865                         }
4866                 }
4867                 else
4868                 {
4869                         /* it's an update, but which kind? */
4870                         if (old_infomask2 & HEAP_KEYS_UPDATED)
4871                                 old_status = MultiXactStatusUpdate;
4872                         else
4873                                 old_status = MultiXactStatusNoKeyUpdate;
4874                 }
4875
4876                 old_mode = TUPLOCK_from_mxstatus(old_status);
4877
4878                 /*
4879                  * If the lock to be acquired is for the same TransactionId as the
4880                  * existing lock, there's an optimization possible: consider only the
4881                  * strongest of both locks as the only one present, and restart.
4882                  */
4883                 if (xmax == add_to_xmax)
4884                 {
4885                         /*
4886                          * Note that it's not possible for the original tuple to be
4887                          * updated: we wouldn't be here because the tuple would have been
4888                          * invisible and we wouldn't try to update it.  As a subtlety,
4889                          * this code can also run when traversing an update chain to lock
4890                          * future versions of a tuple.  But we wouldn't be here either,
4891                          * because the add_to_xmax would be different from the original
4892                          * updater.
4893                          */
4894                         Assert(HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
4895
4896                         /* acquire the strongest of both */
4897                         if (mode < old_mode)
4898                                 mode = old_mode;
4899                         /* mustn't touch is_update */
4900
4901                         old_infomask |= HEAP_XMAX_INVALID;
4902                         goto l5;
4903                 }
4904
4905                 /* otherwise, just fall back to creating a new multixact */
4906                 new_status = get_mxact_status_for_lock(mode, is_update);
4907                 new_xmax = MultiXactIdCreate(xmax, old_status,
4908                                                                          add_to_xmax, new_status);
4909                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4910         }
4911         else if (!HEAP_XMAX_IS_LOCKED_ONLY(old_infomask) &&
4912                          TransactionIdDidCommit(xmax))
4913         {
4914                 /*
4915                  * It's a committed update, so we gotta preserve him as updater of the
4916                  * tuple.
4917                  */
4918                 MultiXactStatus status;
4919                 MultiXactStatus new_status;
4920
4921                 if (old_infomask2 & HEAP_KEYS_UPDATED)
4922                         status = MultiXactStatusUpdate;
4923                 else
4924                         status = MultiXactStatusNoKeyUpdate;
4925
4926                 new_status = get_mxact_status_for_lock(mode, is_update);
4927
4928                 /*
4929                  * since it's not running, it's obviously impossible for the old
4930                  * updater to be identical to the current one, so we need not check
4931                  * for that case as we do in the block above.
4932                  */
4933                 new_xmax = MultiXactIdCreate(xmax, status, add_to_xmax, new_status);
4934                 GetMultiXactIdHintBits(new_xmax, &new_infomask, &new_infomask2);
4935         }
4936         else
4937         {
4938                 /*
4939                  * Can get here iff the locking/updating transaction was running when
4940                  * the infomask was extracted from the tuple, but finished before
4941                  * TransactionIdIsInProgress got to run.  Deal with it as if there was
4942                  * no locker at all in the first place.
4943                  */
4944                 old_infomask |= HEAP_XMAX_INVALID;
4945                 goto l5;
4946         }
4947
4948         *result_infomask = new_infomask;
4949         *result_infomask2 = new_infomask2;
4950         *result_xmax = new_xmax;
4951 }
4952
4953 /*
4954  * Subroutine for heap_lock_updated_tuple_rec.
4955  *
4956  * Given a hypothetical multixact status held by the transaction identified
4957  * with the given xid, does the current transaction need to wait, fail, or can
4958  * it continue if it wanted to acquire a lock of the given mode?  "needwait"
4959  * is set to true if waiting is necessary; if it can continue, then TM_Ok is
4960  * returned.  If the lock is already held by the current transaction, return
4961  * TM_SelfModified.  In case of a conflict with another transaction, a
4962  * different HeapTupleSatisfiesUpdate return code is returned.
4963  *
4964  * The held status is said to be hypothetical because it might correspond to a
4965  * lock held by a single Xid, i.e. not a real MultiXactId; we express it this
4966  * way for simplicity of API.
4967  */
4968 static TM_Result
4969 test_lockmode_for_conflict(MultiXactStatus status, TransactionId xid,
4970                                                    LockTupleMode mode, HeapTuple tup,
4971                                                    bool *needwait)
4972 {
4973         MultiXactStatus wantedstatus;
4974
4975         *needwait = false;
4976         wantedstatus = get_mxact_status_for_lock(mode, false);
4977
4978         /*
4979          * Note: we *must* check TransactionIdIsInProgress before
4980          * TransactionIdDidAbort/Commit; see comment at top of heapam_visibility.c
4981          * for an explanation.
4982          */
4983         if (TransactionIdIsCurrentTransactionId(xid))
4984         {
4985                 /*
4986                  * The tuple has already been locked by our own transaction.  This is
4987                  * very rare but can happen if multiple transactions are trying to
4988                  * lock an ancient version of the same tuple.
4989                  */
4990                 return TM_SelfModified;
4991         }
4992         else if (TransactionIdIsInProgress(xid))
4993         {
4994                 /*
4995                  * If the locking transaction is running, what we do depends on
4996                  * whether the lock modes conflict: if they do, then we must wait for
4997                  * it to finish; otherwise we can fall through to lock this tuple
4998                  * version without waiting.
4999                  */
5000                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5001                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5002                 {
5003                         *needwait = true;
5004                 }
5005
5006                 /*
5007                  * If we set needwait above, then this value doesn't matter;
5008                  * otherwise, this value signals to caller that it's okay to proceed.
5009                  */
5010                 return TM_Ok;
5011         }
5012         else if (TransactionIdDidAbort(xid))
5013                 return TM_Ok;
5014         else if (TransactionIdDidCommit(xid))
5015         {
5016                 /*
5017                  * The other transaction committed.  If it was only a locker, then the
5018                  * lock is completely gone now and we can return success; but if it
5019                  * was an update, then what we do depends on whether the two lock
5020                  * modes conflict.  If they conflict, then we must report error to
5021                  * caller. But if they don't, we can fall through to allow the current
5022                  * transaction to lock the tuple.
5023                  *
5024                  * Note: the reason we worry about ISUPDATE here is because as soon as
5025                  * a transaction ends, all its locks are gone and meaningless, and
5026                  * thus we can ignore them; whereas its updates persist.  In the
5027                  * TransactionIdIsInProgress case, above, we don't need to check
5028                  * because we know the lock is still "alive" and thus a conflict needs
5029                  * always be checked.
5030                  */
5031                 if (!ISUPDATE_from_mxstatus(status))
5032                         return TM_Ok;
5033
5034                 if (DoLockModesConflict(LOCKMODE_from_mxstatus(status),
5035                                                                 LOCKMODE_from_mxstatus(wantedstatus)))
5036                 {
5037                         /* bummer */
5038                         if (!ItemPointerEquals(&tup->t_self, &tup->t_data->t_ctid) ||
5039                                 HeapTupleHeaderIndicatesMovedPartitions(tup->t_data))
5040                                 return TM_Updated;
5041                         else
5042                                 return TM_Deleted;
5043                 }
5044
5045                 return TM_Ok;
5046         }
5047
5048         /* Not in progress, not aborted, not committed -- must have crashed */
5049         return TM_Ok;
5050 }
5051
5052
5053 /*
5054  * Recursive part of heap_lock_updated_tuple
5055  *
5056  * Fetch the tuple pointed to by tid in rel, and mark it as locked by the given
5057  * xid with the given mode; if this tuple is updated, recurse to lock the new
5058  * version as well.
5059  */
5060 static TM_Result
5061 heap_lock_updated_tuple_rec(Relation rel, ItemPointer tid, TransactionId xid,
5062                                                         LockTupleMode mode)
5063 {
5064         TM_Result       result;
5065         ItemPointerData tupid;
5066         HeapTupleData mytup;
5067         Buffer          buf;
5068         uint16          new_infomask,
5069                                 new_infomask2,
5070                                 old_infomask,
5071                                 old_infomask2;
5072         TransactionId xmax,
5073                                 new_xmax;
5074         TransactionId priorXmax = InvalidTransactionId;
5075         bool            cleared_all_frozen = false;
5076         bool            pinned_desired_page;
5077         Buffer          vmbuffer = InvalidBuffer;
5078         BlockNumber block;
5079
5080         ItemPointerCopy(tid, &tupid);
5081
5082         for (;;)
5083         {
5084                 new_infomask = 0;
5085                 new_xmax = InvalidTransactionId;
5086                 block = ItemPointerGetBlockNumber(&tupid);
5087                 ItemPointerCopy(&tupid, &(mytup.t_self));
5088
5089                 if (!heap_fetch(rel, SnapshotAny, &mytup, &buf))
5090                 {
5091                         /*
5092                          * if we fail to find the updated version of the tuple, it's
5093                          * because it was vacuumed/pruned away after its creator
5094                          * transaction aborted.  So behave as if we got to the end of the
5095                          * chain, and there's no further tuple to lock: return success to
5096                          * caller.
5097                          */
5098                         result = TM_Ok;
5099                         goto out_unlocked;
5100                 }
5101
5102 l4:
5103                 CHECK_FOR_INTERRUPTS();
5104
5105                 /*
5106                  * Before locking the buffer, pin the visibility map page if it
5107                  * appears to be necessary.  Since we haven't got the lock yet,
5108                  * someone else might be in the middle of changing this, so we'll need
5109                  * to recheck after we have the lock.
5110                  */
5111                 if (PageIsAllVisible(BufferGetPage(buf)))
5112                 {
5113                         visibilitymap_pin(rel, block, &vmbuffer);
5114                         pinned_desired_page = true;
5115                 }
5116                 else
5117                         pinned_desired_page = false;
5118
5119                 LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5120
5121                 /*
5122                  * If we didn't pin the visibility map page and the page has become
5123                  * all visible while we were busy locking the buffer, we'll have to
5124                  * unlock and re-lock, to avoid holding the buffer lock across I/O.
5125                  * That's a bit unfortunate, but hopefully shouldn't happen often.
5126                  *
5127                  * Note: in some paths through this function, we will reach here
5128                  * holding a pin on a vm page that may or may not be the one matching
5129                  * this page.  If this page isn't all-visible, we won't use the vm
5130                  * page, but we hold onto such a pin till the end of the function.
5131                  */
5132                 if (!pinned_desired_page && PageIsAllVisible(BufferGetPage(buf)))
5133                 {
5134                         LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5135                         visibilitymap_pin(rel, block, &vmbuffer);
5136                         LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
5137                 }
5138
5139                 /*
5140                  * Check the tuple XMIN against prior XMAX, if any.  If we reached the
5141                  * end of the chain, we're done, so return success.
5142                  */
5143                 if (TransactionIdIsValid(priorXmax) &&
5144                         !TransactionIdEquals(HeapTupleHeaderGetXmin(mytup.t_data),
5145                                                                  priorXmax))
5146                 {
5147                         result = TM_Ok;
5148                         goto out_locked;
5149                 }
5150
5151                 /*
5152                  * Also check Xmin: if this tuple was created by an aborted
5153                  * (sub)transaction, then we already locked the last live one in the
5154                  * chain, thus we're done, so return success.
5155                  */
5156                 if (TransactionIdDidAbort(HeapTupleHeaderGetXmin(mytup.t_data)))
5157                 {
5158                         result = TM_Ok;
5159                         goto out_locked;
5160                 }
5161
5162                 old_infomask = mytup.t_data->t_infomask;
5163                 old_infomask2 = mytup.t_data->t_infomask2;
5164                 xmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5165
5166                 /*
5167                  * If this tuple version has been updated or locked by some concurrent
5168                  * transaction(s), what we do depends on whether our lock mode
5169                  * conflicts with what those other transactions hold, and also on the
5170                  * status of them.
5171                  */
5172                 if (!(old_infomask & HEAP_XMAX_INVALID))
5173                 {
5174                         TransactionId rawxmax;
5175                         bool            needwait;
5176
5177                         rawxmax = HeapTupleHeaderGetRawXmax(mytup.t_data);
5178                         if (old_infomask & HEAP_XMAX_IS_MULTI)
5179                         {
5180                                 int                     nmembers;
5181                                 int                     i;
5182                                 MultiXactMember *members;
5183
5184                                 /*
5185                                  * We don't need a test for pg_upgrade'd tuples: this is only
5186                                  * applied to tuples after the first in an update chain.  Said
5187                                  * first tuple in the chain may well be locked-in-9.2-and-
5188                                  * pg_upgraded, but that one was already locked by our caller,
5189                                  * not us; and any subsequent ones cannot be because our
5190                                  * caller must necessarily have obtained a snapshot later than
5191                                  * the pg_upgrade itself.
5192                                  */
5193                                 Assert(!HEAP_LOCKED_UPGRADED(mytup.t_data->t_infomask));
5194
5195                                 nmembers = GetMultiXactIdMembers(rawxmax, &members, false,
5196                                                                                                  HEAP_XMAX_IS_LOCKED_ONLY(old_infomask));
5197                                 for (i = 0; i < nmembers; i++)
5198                                 {
5199                                         result = test_lockmode_for_conflict(members[i].status,
5200                                                                                                                 members[i].xid,
5201                                                                                                                 mode,
5202                                                                                                                 &mytup,
5203                                                                                                                 &needwait);
5204
5205                                         /*
5206                                          * If the tuple was already locked by ourselves in a
5207                                          * previous iteration of this (say heap_lock_tuple was
5208                                          * forced to restart the locking loop because of a change
5209                                          * in xmax), then we hold the lock already on this tuple
5210                                          * version and we don't need to do anything; and this is
5211                                          * not an error condition either.  We just need to skip
5212                                          * this tuple and continue locking the next version in the
5213                                          * update chain.
5214                                          */
5215                                         if (result == TM_SelfModified)
5216                                         {
5217                                                 pfree(members);
5218                                                 goto next;
5219                                         }
5220
5221                                         if (needwait)
5222                                         {
5223                                                 LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5224                                                 XactLockTableWait(members[i].xid, rel,
5225                                                                                   &mytup.t_self,
5226                                                                                   XLTW_LockUpdated);
5227                                                 pfree(members);
5228                                                 goto l4;
5229                                         }
5230                                         if (result != TM_Ok)
5231                                         {
5232                                                 pfree(members);
5233                                                 goto out_locked;
5234                                         }
5235                                 }
5236                                 if (members)
5237                                         pfree(members);
5238                         }
5239                         else
5240                         {
5241                                 MultiXactStatus status;
5242
5243                                 /*
5244                                  * For a non-multi Xmax, we first need to compute the
5245                                  * corresponding MultiXactStatus by using the infomask bits.
5246                                  */
5247                                 if (HEAP_XMAX_IS_LOCKED_ONLY(old_infomask))
5248                                 {
5249                                         if (HEAP_XMAX_IS_KEYSHR_LOCKED(old_infomask))
5250                                                 status = MultiXactStatusForKeyShare;
5251                                         else if (HEAP_XMAX_IS_SHR_LOCKED(old_infomask))
5252                                                 status = MultiXactStatusForShare;
5253                                         else if (HEAP_XMAX_IS_EXCL_LOCKED(old_infomask))
5254                                         {
5255                                                 if (old_infomask2 & HEAP_KEYS_UPDATED)
5256                                                         status = MultiXactStatusForUpdate;
5257                                                 else
5258                                                         status = MultiXactStatusForNoKeyUpdate;
5259                                         }
5260                                         else
5261                                         {
5262                                                 /*
5263                                                  * LOCK_ONLY present alone (a pg_upgraded tuple marked
5264                                                  * as share-locked in the old cluster) shouldn't be
5265                                                  * seen in the middle of an update chain.
5266                                                  */
5267                                                 elog(ERROR, "invalid lock status in tuple");
5268                                         }
5269                                 }
5270                                 else
5271                                 {
5272                                         /* it's an update, but which kind? */
5273                                         if (old_infomask2 & HEAP_KEYS_UPDATED)
5274                                                 status = MultiXactStatusUpdate;
5275                                         else
5276                                                 status = MultiXactStatusNoKeyUpdate;
5277                                 }
5278
5279                                 result = test_lockmode_for_conflict(status, rawxmax, mode,
5280                                                                                                         &mytup, &needwait);
5281
5282                                 /*
5283                                  * If the tuple was already locked by ourselves in a previous
5284                                  * iteration of this (say heap_lock_tuple was forced to
5285                                  * restart the locking loop because of a change in xmax), then
5286                                  * we hold the lock already on this tuple version and we don't
5287                                  * need to do anything; and this is not an error condition
5288                                  * either.  We just need to skip this tuple and continue
5289                                  * locking the next version in the update chain.
5290                                  */
5291                                 if (result == TM_SelfModified)
5292                                         goto next;
5293
5294                                 if (needwait)
5295                                 {
5296                                         LockBuffer(buf, BUFFER_LOCK_UNLOCK);
5297                                         XactLockTableWait(rawxmax, rel, &mytup.t_self,
5298                                                                           XLTW_LockUpdated);
5299                                         goto l4;
5300                                 }
5301                                 if (result != TM_Ok)
5302                                 {
5303                                         goto out_locked;
5304                                 }
5305                         }
5306                 }
5307
5308                 /* compute the new Xmax and infomask values for the tuple ... */
5309                 compute_new_xmax_infomask(xmax, old_infomask, mytup.t_data->t_infomask2,
5310                                                                   xid, mode, false,
5311                                                                   &new_xmax, &new_infomask, &new_infomask2);
5312
5313                 if (PageIsAllVisible(BufferGetPage(buf)) &&
5314                         visibilitymap_clear(rel, block, vmbuffer,
5315                                                                 VISIBILITYMAP_ALL_FROZEN))
5316                         cleared_all_frozen = true;
5317
5318                 START_CRIT_SECTION();
5319
5320                 /* ... and set them */
5321                 HeapTupleHeaderSetXmax(mytup.t_data, new_xmax);
5322                 mytup.t_data->t_infomask &= ~HEAP_XMAX_BITS;
5323                 mytup.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5324                 mytup.t_data->t_infomask |= new_infomask;
5325                 mytup.t_data->t_infomask2 |= new_infomask2;
5326
5327                 MarkBufferDirty(buf);
5328
5329                 /* XLOG stuff */
5330                 if (RelationNeedsWAL(rel))
5331                 {
5332                         xl_heap_lock_updated xlrec;
5333                         XLogRecPtr      recptr;
5334                         Page            page = BufferGetPage(buf);
5335
5336                         XLogBeginInsert();
5337                         XLogRegisterBuffer(0, buf, REGBUF_STANDARD);
5338
5339                         xlrec.offnum = ItemPointerGetOffsetNumber(&mytup.t_self);
5340                         xlrec.xmax = new_xmax;
5341                         xlrec.infobits_set = compute_infobits(new_infomask, new_infomask2);
5342                         xlrec.flags =
5343                                 cleared_all_frozen ? XLH_LOCK_ALL_FROZEN_CLEARED : 0;
5344
5345                         XLogRegisterData((char *) &xlrec, SizeOfHeapLockUpdated);
5346
5347                         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_LOCK_UPDATED);
5348
5349                         PageSetLSN(page, recptr);
5350                 }
5351
5352                 END_CRIT_SECTION();
5353
5354 next:
5355                 /* if we find the end of update chain, we're done. */
5356                 if (mytup.t_data->t_infomask & HEAP_XMAX_INVALID ||
5357                         HeapTupleHeaderIndicatesMovedPartitions(mytup.t_data) ||
5358                         ItemPointerEquals(&mytup.t_self, &mytup.t_data->t_ctid) ||
5359                         HeapTupleHeaderIsOnlyLocked(mytup.t_data))
5360                 {
5361                         result = TM_Ok;
5362                         goto out_locked;
5363                 }
5364
5365                 /* tail recursion */
5366                 priorXmax = HeapTupleHeaderGetUpdateXid(mytup.t_data);
5367                 ItemPointerCopy(&(mytup.t_data->t_ctid), &tupid);
5368                 UnlockReleaseBuffer(buf);
5369         }
5370
5371         result = TM_Ok;
5372
5373 out_locked:
5374         UnlockReleaseBuffer(buf);
5375
5376 out_unlocked:
5377         if (vmbuffer != InvalidBuffer)
5378                 ReleaseBuffer(vmbuffer);
5379
5380         return result;
5381 }
5382
5383 /*
5384  * heap_lock_updated_tuple
5385  *              Follow update chain when locking an updated tuple, acquiring locks (row
5386  *              marks) on the updated versions.
5387  *
5388  * The initial tuple is assumed to be already locked.
5389  *
5390  * This function doesn't check visibility, it just unconditionally marks the
5391  * tuple(s) as locked.  If any tuple in the updated chain is being deleted
5392  * concurrently (or updated with the key being modified), sleep until the
5393  * transaction doing it is finished.
5394  *
5395  * Note that we don't acquire heavyweight tuple locks on the tuples we walk
5396  * when we have to wait for other transactions to release them, as opposed to
5397  * what heap_lock_tuple does.  The reason is that having more than one
5398  * transaction walking the chain is probably uncommon enough that risk of
5399  * starvation is not likely: one of the preconditions for being here is that
5400  * the snapshot in use predates the update that created this tuple (because we
5401  * started at an earlier version of the tuple), but at the same time such a
5402  * transaction cannot be using repeatable read or serializable isolation
5403  * levels, because that would lead to a serializability failure.
5404  */
5405 static TM_Result
5406 heap_lock_updated_tuple(Relation rel, HeapTuple tuple, ItemPointer ctid,
5407                                                 TransactionId xid, LockTupleMode mode)
5408 {
5409         /*
5410          * If the tuple has not been updated, or has moved into another partition
5411          * (effectively a delete) stop here.
5412          */
5413         if (!HeapTupleHeaderIndicatesMovedPartitions(tuple->t_data) &&
5414                 !ItemPointerEquals(&tuple->t_self, ctid))
5415         {
5416                 /*
5417                  * If this is the first possibly-multixact-able operation in the
5418                  * current transaction, set my per-backend OldestMemberMXactId
5419                  * setting. We can be certain that the transaction will never become a
5420                  * member of any older MultiXactIds than that.  (We have to do this
5421                  * even if we end up just using our own TransactionId below, since
5422                  * some other backend could incorporate our XID into a MultiXact
5423                  * immediately afterwards.)
5424                  */
5425                 MultiXactIdSetOldestMember();
5426
5427                 return heap_lock_updated_tuple_rec(rel, ctid, xid, mode);
5428         }
5429
5430         /* nothing to lock */
5431         return TM_Ok;
5432 }
5433
5434 /*
5435  *      heap_finish_speculative - mark speculative insertion as successful
5436  *
5437  * To successfully finish a speculative insertion we have to clear speculative
5438  * token from tuple.  To do so the t_ctid field, which will contain a
5439  * speculative token value, is modified in place to point to the tuple itself,
5440  * which is characteristic of a newly inserted ordinary tuple.
5441  *
5442  * NB: It is not ok to commit without either finishing or aborting a
5443  * speculative insertion.  We could treat speculative tuples of committed
5444  * transactions implicitly as completed, but then we would have to be prepared
5445  * to deal with speculative tokens on committed tuples.  That wouldn't be
5446  * difficult - no-one looks at the ctid field of a tuple with invalid xmax -
5447  * but clearing the token at completion isn't very expensive either.
5448  * An explicit confirmation WAL record also makes logical decoding simpler.
5449  */
5450 void
5451 heap_finish_speculative(Relation relation, ItemPointer tid)
5452 {
5453         Buffer          buffer;
5454         Page            page;
5455         OffsetNumber offnum;
5456         ItemId          lp = NULL;
5457         HeapTupleHeader htup;
5458
5459         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(tid));
5460         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5461         page = (Page) BufferGetPage(buffer);
5462
5463         offnum = ItemPointerGetOffsetNumber(tid);
5464         if (PageGetMaxOffsetNumber(page) >= offnum)
5465                 lp = PageGetItemId(page, offnum);
5466
5467         if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
5468                 elog(ERROR, "invalid lp");
5469
5470         htup = (HeapTupleHeader) PageGetItem(page, lp);
5471
5472         /* SpecTokenOffsetNumber should be distinguishable from any real offset */
5473         StaticAssertStmt(MaxOffsetNumber < SpecTokenOffsetNumber,
5474                                          "invalid speculative token constant");
5475
5476         /* NO EREPORT(ERROR) from here till changes are logged */
5477         START_CRIT_SECTION();
5478
5479         Assert(HeapTupleHeaderIsSpeculative(htup));
5480
5481         MarkBufferDirty(buffer);
5482
5483         /*
5484          * Replace the speculative insertion token with a real t_ctid, pointing to
5485          * itself like it does on regular tuples.
5486          */
5487         htup->t_ctid = *tid;
5488
5489         /* XLOG stuff */
5490         if (RelationNeedsWAL(relation))
5491         {
5492                 xl_heap_confirm xlrec;
5493                 XLogRecPtr      recptr;
5494
5495                 xlrec.offnum = ItemPointerGetOffsetNumber(tid);
5496
5497                 XLogBeginInsert();
5498
5499                 /* We want the same filtering on this as on a plain insert */
5500                 XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
5501
5502                 XLogRegisterData((char *) &xlrec, SizeOfHeapConfirm);
5503                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
5504
5505                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_CONFIRM);
5506
5507                 PageSetLSN(page, recptr);
5508         }
5509
5510         END_CRIT_SECTION();
5511
5512         UnlockReleaseBuffer(buffer);
5513 }
5514
5515 /*
5516  *      heap_abort_speculative - kill a speculatively inserted tuple
5517  *
5518  * Marks a tuple that was speculatively inserted in the same command as dead,
5519  * by setting its xmin as invalid.  That makes it immediately appear as dead
5520  * to all transactions, including our own.  In particular, it makes
5521  * HeapTupleSatisfiesDirty() regard the tuple as dead, so that another backend
5522  * inserting a duplicate key value won't unnecessarily wait for our whole
5523  * transaction to finish (it'll just wait for our speculative insertion to
5524  * finish).
5525  *
5526  * Killing the tuple prevents "unprincipled deadlocks", which are deadlocks
5527  * that arise due to a mutual dependency that is not user visible.  By
5528  * definition, unprincipled deadlocks cannot be prevented by the user
5529  * reordering lock acquisition in client code, because the implementation level
5530  * lock acquisitions are not under the user's direct control.  If speculative
5531  * inserters did not take this precaution, then under high concurrency they
5532  * could deadlock with each other, which would not be acceptable.
5533  *
5534  * This is somewhat redundant with heap_delete, but we prefer to have a
5535  * dedicated routine with stripped down requirements.  Note that this is also
5536  * used to delete the TOAST tuples created during speculative insertion.
5537  *
5538  * This routine does not affect logical decoding as it only looks at
5539  * confirmation records.
5540  */
5541 void
5542 heap_abort_speculative(Relation relation, ItemPointer tid)
5543 {
5544         TransactionId xid = GetCurrentTransactionId();
5545         ItemId          lp;
5546         HeapTupleData tp;
5547         Page            page;
5548         BlockNumber block;
5549         Buffer          buffer;
5550
5551         Assert(ItemPointerIsValid(tid));
5552
5553         block = ItemPointerGetBlockNumber(tid);
5554         buffer = ReadBuffer(relation, block);
5555         page = BufferGetPage(buffer);
5556
5557         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5558
5559         /*
5560          * Page can't be all visible, we just inserted into it, and are still
5561          * running.
5562          */
5563         Assert(!PageIsAllVisible(page));
5564
5565         lp = PageGetItemId(page, ItemPointerGetOffsetNumber(tid));
5566         Assert(ItemIdIsNormal(lp));
5567
5568         tp.t_tableOid = RelationGetRelid(relation);
5569         tp.t_data = (HeapTupleHeader) PageGetItem(page, lp);
5570         tp.t_len = ItemIdGetLength(lp);
5571         tp.t_self = *tid;
5572
5573         /*
5574          * Sanity check that the tuple really is a speculatively inserted tuple,
5575          * inserted by us.
5576          */
5577         if (tp.t_data->t_choice.t_heap.t_xmin != xid)
5578                 elog(ERROR, "attempted to kill a tuple inserted by another transaction");
5579         if (!(IsToastRelation(relation) || HeapTupleHeaderIsSpeculative(tp.t_data)))
5580                 elog(ERROR, "attempted to kill a non-speculative tuple");
5581         Assert(!HeapTupleHeaderIsHeapOnly(tp.t_data));
5582
5583         /*
5584          * No need to check for serializable conflicts here.  There is never a
5585          * need for a combocid, either.  No need to extract replica identity, or
5586          * do anything special with infomask bits.
5587          */
5588
5589         START_CRIT_SECTION();
5590
5591         /*
5592          * The tuple will become DEAD immediately.  Flag that this page
5593          * immediately is a candidate for pruning by setting xmin to
5594          * RecentGlobalXmin.  That's not pretty, but it doesn't seem worth
5595          * inventing a nicer API for this.
5596          */
5597         Assert(TransactionIdIsValid(RecentGlobalXmin));
5598         PageSetPrunable(page, RecentGlobalXmin);
5599
5600         /* store transaction information of xact deleting the tuple */
5601         tp.t_data->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
5602         tp.t_data->t_infomask2 &= ~HEAP_KEYS_UPDATED;
5603
5604         /*
5605          * Set the tuple header xmin to InvalidTransactionId.  This makes the
5606          * tuple immediately invisible everyone.  (In particular, to any
5607          * transactions waiting on the speculative token, woken up later.)
5608          */
5609         HeapTupleHeaderSetXmin(tp.t_data, InvalidTransactionId);
5610
5611         /* Clear the speculative insertion token too */
5612         tp.t_data->t_ctid = tp.t_self;
5613
5614         MarkBufferDirty(buffer);
5615
5616         /*
5617          * XLOG stuff
5618          *
5619          * The WAL records generated here match heap_delete().  The same recovery
5620          * routines are used.
5621          */
5622         if (RelationNeedsWAL(relation))
5623         {
5624                 xl_heap_delete xlrec;
5625                 XLogRecPtr      recptr;
5626
5627                 xlrec.flags = XLH_DELETE_IS_SUPER;
5628                 xlrec.infobits_set = compute_infobits(tp.t_data->t_infomask,
5629                                                                                           tp.t_data->t_infomask2);
5630                 xlrec.offnum = ItemPointerGetOffsetNumber(&tp.t_self);
5631                 xlrec.xmax = xid;
5632
5633                 XLogBeginInsert();
5634                 XLogRegisterData((char *) &xlrec, SizeOfHeapDelete);
5635                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
5636
5637                 /* No replica identity & replication origin logged */
5638
5639                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_DELETE);
5640
5641                 PageSetLSN(page, recptr);
5642         }
5643
5644         END_CRIT_SECTION();
5645
5646         LockBuffer(buffer, BUFFER_LOCK_UNLOCK);
5647
5648         if (HeapTupleHasExternal(&tp))
5649         {
5650                 Assert(!IsToastRelation(relation));
5651                 toast_delete(relation, &tp, true);
5652         }
5653
5654         /*
5655          * Never need to mark tuple for invalidation, since catalogs don't support
5656          * speculative insertion
5657          */
5658
5659         /* Now we can release the buffer */
5660         ReleaseBuffer(buffer);
5661
5662         /* count deletion, as we counted the insertion too */
5663         pgstat_count_heap_delete(relation);
5664 }
5665
5666 /*
5667  * heap_inplace_update - update a tuple "in place" (ie, overwrite it)
5668  *
5669  * Overwriting violates both MVCC and transactional safety, so the uses
5670  * of this function in Postgres are extremely limited.  Nonetheless we
5671  * find some places to use it.
5672  *
5673  * The tuple cannot change size, and therefore it's reasonable to assume
5674  * that its null bitmap (if any) doesn't change either.  So we just
5675  * overwrite the data portion of the tuple without touching the null
5676  * bitmap or any of the header fields.
5677  *
5678  * tuple is an in-memory tuple structure containing the data to be written
5679  * over the target tuple.  Also, tuple->t_self identifies the target tuple.
5680  */
5681 void
5682 heap_inplace_update(Relation relation, HeapTuple tuple)
5683 {
5684         Buffer          buffer;
5685         Page            page;
5686         OffsetNumber offnum;
5687         ItemId          lp = NULL;
5688         HeapTupleHeader htup;
5689         uint32          oldlen;
5690         uint32          newlen;
5691
5692         /*
5693          * For now, parallel operations are required to be strictly read-only.
5694          * Unlike a regular update, this should never create a combo CID, so it
5695          * might be possible to relax this restriction, but not without more
5696          * thought and testing.  It's not clear that it would be useful, anyway.
5697          */
5698         if (IsInParallelMode())
5699                 ereport(ERROR,
5700                                 (errcode(ERRCODE_INVALID_TRANSACTION_STATE),
5701                                  errmsg("cannot update tuples during a parallel operation")));
5702
5703         buffer = ReadBuffer(relation, ItemPointerGetBlockNumber(&(tuple->t_self)));
5704         LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
5705         page = (Page) BufferGetPage(buffer);
5706
5707         offnum = ItemPointerGetOffsetNumber(&(tuple->t_self));
5708         if (PageGetMaxOffsetNumber(page) >= offnum)
5709                 lp = PageGetItemId(page, offnum);
5710
5711         if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
5712                 elog(ERROR, "invalid lp");
5713
5714         htup = (HeapTupleHeader) PageGetItem(page, lp);
5715
5716         oldlen = ItemIdGetLength(lp) - htup->t_hoff;
5717         newlen = tuple->t_len - tuple->t_data->t_hoff;
5718         if (oldlen != newlen || htup->t_hoff != tuple->t_data->t_hoff)
5719                 elog(ERROR, "wrong tuple length");
5720
5721         /* NO EREPORT(ERROR) from here till changes are logged */
5722         START_CRIT_SECTION();
5723
5724         memcpy((char *) htup + htup->t_hoff,
5725                    (char *) tuple->t_data + tuple->t_data->t_hoff,
5726                    newlen);
5727
5728         MarkBufferDirty(buffer);
5729
5730         /* XLOG stuff */
5731         if (RelationNeedsWAL(relation))
5732         {
5733                 xl_heap_inplace xlrec;
5734                 XLogRecPtr      recptr;
5735
5736                 xlrec.offnum = ItemPointerGetOffsetNumber(&tuple->t_self);
5737
5738                 XLogBeginInsert();
5739                 XLogRegisterData((char *) &xlrec, SizeOfHeapInplace);
5740
5741                 XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
5742                 XLogRegisterBufData(0, (char *) htup + htup->t_hoff, newlen);
5743
5744                 /* inplace updates aren't decoded atm, don't log the origin */
5745
5746                 recptr = XLogInsert(RM_HEAP_ID, XLOG_HEAP_INPLACE);
5747
5748                 PageSetLSN(page, recptr);
5749         }
5750
5751         END_CRIT_SECTION();
5752
5753         UnlockReleaseBuffer(buffer);
5754
5755         /*
5756          * Send out shared cache inval if necessary.  Note that because we only
5757          * pass the new version of the tuple, this mustn't be used for any
5758          * operations that could change catcache lookup keys.  But we aren't
5759          * bothering with index updates either, so that's true a fortiori.
5760          */
5761         if (!IsBootstrapProcessingMode())
5762                 CacheInvalidateHeapTuple(relation, tuple, NULL);
5763 }
5764
5765 #define         FRM_NOOP                                0x0001
5766 #define         FRM_INVALIDATE_XMAX             0x0002
5767 #define         FRM_RETURN_IS_XID               0x0004
5768 #define         FRM_RETURN_IS_MULTI             0x0008
5769 #define         FRM_MARK_COMMITTED              0x0010
5770
5771 /*
5772  * FreezeMultiXactId
5773  *              Determine what to do during freezing when a tuple is marked by a
5774  *              MultiXactId.
5775  *
5776  * NB -- this might have the side-effect of creating a new MultiXactId!
5777  *
5778  * "flags" is an output value; it's used to tell caller what to do on return.
5779  * Possible flags are:
5780  * FRM_NOOP
5781  *              don't do anything -- keep existing Xmax
5782  * FRM_INVALIDATE_XMAX
5783  *              mark Xmax as InvalidTransactionId and set XMAX_INVALID flag.
5784  * FRM_RETURN_IS_XID
5785  *              The Xid return value is a single update Xid to set as xmax.
5786  * FRM_MARK_COMMITTED
5787  *              Xmax can be marked as HEAP_XMAX_COMMITTED
5788  * FRM_RETURN_IS_MULTI
5789  *              The return value is a new MultiXactId to set as new Xmax.
5790  *              (caller must obtain proper infomask bits using GetMultiXactIdHintBits)
5791  */
5792 static TransactionId
5793 FreezeMultiXactId(MultiXactId multi, uint16 t_infomask,
5794                                   TransactionId relfrozenxid, TransactionId relminmxid,
5795                                   TransactionId cutoff_xid, MultiXactId cutoff_multi,
5796                                   uint16 *flags)
5797 {
5798         TransactionId xid = InvalidTransactionId;
5799         int                     i;
5800         MultiXactMember *members;
5801         int                     nmembers;
5802         bool            need_replace;
5803         int                     nnewmembers;
5804         MultiXactMember *newmembers;
5805         bool            has_lockers;
5806         TransactionId update_xid;
5807         bool            update_committed;
5808
5809         *flags = 0;
5810
5811         /* We should only be called in Multis */
5812         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
5813
5814         if (!MultiXactIdIsValid(multi) ||
5815                 HEAP_LOCKED_UPGRADED(t_infomask))
5816         {
5817                 /* Ensure infomask bits are appropriately set/reset */
5818                 *flags |= FRM_INVALIDATE_XMAX;
5819                 return InvalidTransactionId;
5820         }
5821         else if (MultiXactIdPrecedes(multi, relminmxid))
5822                 ereport(ERROR,
5823                                 (errcode(ERRCODE_DATA_CORRUPTED),
5824                                  errmsg_internal("found multixact %u from before relminmxid %u",
5825                                                                  multi, relminmxid)));
5826         else if (MultiXactIdPrecedes(multi, cutoff_multi))
5827         {
5828                 /*
5829                  * This old multi cannot possibly have members still running, but
5830                  * verify just in case.  If it was a locker only, it can be removed
5831                  * without any further consideration; but if it contained an update,
5832                  * we might need to preserve it.
5833                  */
5834                 if (MultiXactIdIsRunning(multi,
5835                                                                  HEAP_XMAX_IS_LOCKED_ONLY(t_infomask)))
5836                         ereport(ERROR,
5837                                         (errcode(ERRCODE_DATA_CORRUPTED),
5838                                          errmsg_internal("multixact %u from before cutoff %u found to be still running",
5839                                                                          multi, cutoff_multi)));
5840
5841                 if (HEAP_XMAX_IS_LOCKED_ONLY(t_infomask))
5842                 {
5843                         *flags |= FRM_INVALIDATE_XMAX;
5844                         xid = InvalidTransactionId; /* not strictly necessary */
5845                 }
5846                 else
5847                 {
5848                         /* replace multi by update xid */
5849                         xid = MultiXactIdGetUpdateXid(multi, t_infomask);
5850
5851                         /* wasn't only a lock, xid needs to be valid */
5852                         Assert(TransactionIdIsValid(xid));
5853
5854                         if (TransactionIdPrecedes(xid, relfrozenxid))
5855                                 ereport(ERROR,
5856                                                 (errcode(ERRCODE_DATA_CORRUPTED),
5857                                                  errmsg_internal("found update xid %u from before relfrozenxid %u",
5858                                                                                  xid, relfrozenxid)));
5859
5860                         /*
5861                          * If the xid is older than the cutoff, it has to have aborted,
5862                          * otherwise the tuple would have gotten pruned away.
5863                          */
5864                         if (TransactionIdPrecedes(xid, cutoff_xid))
5865                         {
5866                                 if (TransactionIdDidCommit(xid))
5867                                         ereport(ERROR,
5868                                                         (errcode(ERRCODE_DATA_CORRUPTED),
5869                                                          errmsg_internal("cannot freeze committed update xid %u", xid)));
5870                                 *flags |= FRM_INVALIDATE_XMAX;
5871                                 xid = InvalidTransactionId; /* not strictly necessary */
5872                         }
5873                         else
5874                         {
5875                                 *flags |= FRM_RETURN_IS_XID;
5876                         }
5877                 }
5878
5879                 return xid;
5880         }
5881
5882         /*
5883          * This multixact might have or might not have members still running, but
5884          * we know it's valid and is newer than the cutoff point for multis.
5885          * However, some member(s) of it may be below the cutoff for Xids, so we
5886          * need to walk the whole members array to figure out what to do, if
5887          * anything.
5888          */
5889
5890         nmembers =
5891                 GetMultiXactIdMembers(multi, &members, false,
5892                                                           HEAP_XMAX_IS_LOCKED_ONLY(t_infomask));
5893         if (nmembers <= 0)
5894         {
5895                 /* Nothing worth keeping */
5896                 *flags |= FRM_INVALIDATE_XMAX;
5897                 return InvalidTransactionId;
5898         }
5899
5900         /* is there anything older than the cutoff? */
5901         need_replace = false;
5902         for (i = 0; i < nmembers; i++)
5903         {
5904                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
5905                 {
5906                         need_replace = true;
5907                         break;
5908                 }
5909         }
5910
5911         /*
5912          * In the simplest case, there is no member older than the cutoff; we can
5913          * keep the existing MultiXactId as is.
5914          */
5915         if (!need_replace)
5916         {
5917                 *flags |= FRM_NOOP;
5918                 pfree(members);
5919                 return InvalidTransactionId;
5920         }
5921
5922         /*
5923          * If the multi needs to be updated, figure out which members do we need
5924          * to keep.
5925          */
5926         nnewmembers = 0;
5927         newmembers = palloc(sizeof(MultiXactMember) * nmembers);
5928         has_lockers = false;
5929         update_xid = InvalidTransactionId;
5930         update_committed = false;
5931
5932         for (i = 0; i < nmembers; i++)
5933         {
5934                 /*
5935                  * Determine whether to keep this member or ignore it.
5936                  */
5937                 if (ISUPDATE_from_mxstatus(members[i].status))
5938                 {
5939                         TransactionId xid = members[i].xid;
5940
5941                         Assert(TransactionIdIsValid(xid));
5942                         if (TransactionIdPrecedes(xid, relfrozenxid))
5943                                 ereport(ERROR,
5944                                                 (errcode(ERRCODE_DATA_CORRUPTED),
5945                                                  errmsg_internal("found update xid %u from before relfrozenxid %u",
5946                                                                                  xid, relfrozenxid)));
5947
5948                         /*
5949                          * It's an update; should we keep it?  If the transaction is known
5950                          * aborted or crashed then it's okay to ignore it, otherwise not.
5951                          * Note that an updater older than cutoff_xid cannot possibly be
5952                          * committed, because HeapTupleSatisfiesVacuum would have returned
5953                          * HEAPTUPLE_DEAD and we would not be trying to freeze the tuple.
5954                          *
5955                          * As with all tuple visibility routines, it's critical to test
5956                          * TransactionIdIsInProgress before TransactionIdDidCommit,
5957                          * because of race conditions explained in detail in
5958                          * heapam_visibility.c.
5959                          */
5960                         if (TransactionIdIsCurrentTransactionId(xid) ||
5961                                 TransactionIdIsInProgress(xid))
5962                         {
5963                                 Assert(!TransactionIdIsValid(update_xid));
5964                                 update_xid = xid;
5965                         }
5966                         else if (TransactionIdDidCommit(xid))
5967                         {
5968                                 /*
5969                                  * The transaction committed, so we can tell caller to set
5970                                  * HEAP_XMAX_COMMITTED.  (We can only do this because we know
5971                                  * the transaction is not running.)
5972                                  */
5973                                 Assert(!TransactionIdIsValid(update_xid));
5974                                 update_committed = true;
5975                                 update_xid = xid;
5976                         }
5977                         else
5978                         {
5979                                 /*
5980                                  * Not in progress, not committed -- must be aborted or
5981                                  * crashed; we can ignore it.
5982                                  */
5983                         }
5984
5985                         /*
5986                          * Since the tuple wasn't marked HEAPTUPLE_DEAD by vacuum, the
5987                          * update Xid cannot possibly be older than the xid cutoff. The
5988                          * presence of such a tuple would cause corruption, so be paranoid
5989                          * and check.
5990                          */
5991                         if (TransactionIdIsValid(update_xid) &&
5992                                 TransactionIdPrecedes(update_xid, cutoff_xid))
5993                                 ereport(ERROR,
5994                                                 (errcode(ERRCODE_DATA_CORRUPTED),
5995                                                  errmsg_internal("found update xid %u from before xid cutoff %u",
5996                                                                                  update_xid, cutoff_xid)));
5997
5998                         /*
5999                          * If we determined that it's an Xid corresponding to an update
6000                          * that must be retained, additionally add it to the list of
6001                          * members of the new Multi, in case we end up using that.  (We
6002                          * might still decide to use only an update Xid and not a multi,
6003                          * but it's easier to maintain the list as we walk the old members
6004                          * list.)
6005                          */
6006                         if (TransactionIdIsValid(update_xid))
6007                                 newmembers[nnewmembers++] = members[i];
6008                 }
6009                 else
6010                 {
6011                         /* We only keep lockers if they are still running */
6012                         if (TransactionIdIsCurrentTransactionId(members[i].xid) ||
6013                                 TransactionIdIsInProgress(members[i].xid))
6014                         {
6015                                 /* running locker cannot possibly be older than the cutoff */
6016                                 Assert(!TransactionIdPrecedes(members[i].xid, cutoff_xid));
6017                                 newmembers[nnewmembers++] = members[i];
6018                                 has_lockers = true;
6019                         }
6020                 }
6021         }
6022
6023         pfree(members);
6024
6025         if (nnewmembers == 0)
6026         {
6027                 /* nothing worth keeping!? Tell caller to remove the whole thing */
6028                 *flags |= FRM_INVALIDATE_XMAX;
6029                 xid = InvalidTransactionId;
6030         }
6031         else if (TransactionIdIsValid(update_xid) && !has_lockers)
6032         {
6033                 /*
6034                  * If there's a single member and it's an update, pass it back alone
6035                  * without creating a new Multi.  (XXX we could do this when there's a
6036                  * single remaining locker, too, but that would complicate the API too
6037                  * much; moreover, the case with the single updater is more
6038                  * interesting, because those are longer-lived.)
6039                  */
6040                 Assert(nnewmembers == 1);
6041                 *flags |= FRM_RETURN_IS_XID;
6042                 if (update_committed)
6043                         *flags |= FRM_MARK_COMMITTED;
6044                 xid = update_xid;
6045         }
6046         else
6047         {
6048                 /*
6049                  * Create a new multixact with the surviving members of the previous
6050                  * one, to set as new Xmax in the tuple.
6051                  */
6052                 xid = MultiXactIdCreateFromMembers(nnewmembers, newmembers);
6053                 *flags |= FRM_RETURN_IS_MULTI;
6054         }
6055
6056         pfree(newmembers);
6057
6058         return xid;
6059 }
6060
6061 /*
6062  * heap_prepare_freeze_tuple
6063  *
6064  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6065  * are older than the specified cutoff XID and cutoff MultiXactId.  If so,
6066  * setup enough state (in the *frz output argument) to later execute and
6067  * WAL-log what we would need to do, and return true.  Return false if nothing
6068  * is to be changed.  In addition, set *totally_frozen_p to true if the tuple
6069  * will be totally frozen after these operations are performed and false if
6070  * more freezing will eventually be required.
6071  *
6072  * Caller is responsible for setting the offset field, if appropriate.
6073  *
6074  * It is assumed that the caller has checked the tuple with
6075  * HeapTupleSatisfiesVacuum() and determined that it is not HEAPTUPLE_DEAD
6076  * (else we should be removing the tuple, not freezing it).
6077  *
6078  * NB: cutoff_xid *must* be <= the current global xmin, to ensure that any
6079  * XID older than it could neither be running nor seen as running by any
6080  * open transaction.  This ensures that the replacement will not change
6081  * anyone's idea of the tuple state.
6082  * Similarly, cutoff_multi must be less than or equal to the smallest
6083  * MultiXactId used by any transaction currently open.
6084  *
6085  * If the tuple is in a shared buffer, caller must hold an exclusive lock on
6086  * that buffer.
6087  *
6088  * NB: It is not enough to set hint bits to indicate something is
6089  * committed/invalid -- they might not be set on a standby, or after crash
6090  * recovery.  We really need to remove old xids.
6091  */
6092 bool
6093 heap_prepare_freeze_tuple(HeapTupleHeader tuple,
6094                                                   TransactionId relfrozenxid, TransactionId relminmxid,
6095                                                   TransactionId cutoff_xid, TransactionId cutoff_multi,
6096                                                   xl_heap_freeze_tuple *frz, bool *totally_frozen_p)
6097 {
6098         bool            changed = false;
6099         bool            xmax_already_frozen = false;
6100         bool            xmin_frozen;
6101         bool            freeze_xmax;
6102         TransactionId xid;
6103
6104         frz->frzflags = 0;
6105         frz->t_infomask2 = tuple->t_infomask2;
6106         frz->t_infomask = tuple->t_infomask;
6107         frz->xmax = HeapTupleHeaderGetRawXmax(tuple);
6108
6109         /*
6110          * Process xmin.  xmin_frozen has two slightly different meanings: in the
6111          * !XidIsNormal case, it means "the xmin doesn't need any freezing" (it's
6112          * already a permanent value), while in the block below it is set true to
6113          * mean "xmin won't need freezing after what we do to it here" (false
6114          * otherwise).  In both cases we're allowed to set totally_frozen, as far
6115          * as xmin is concerned.
6116          */
6117         xid = HeapTupleHeaderGetXmin(tuple);
6118         if (!TransactionIdIsNormal(xid))
6119                 xmin_frozen = true;
6120         else
6121         {
6122                 if (TransactionIdPrecedes(xid, relfrozenxid))
6123                         ereport(ERROR,
6124                                         (errcode(ERRCODE_DATA_CORRUPTED),
6125                                          errmsg_internal("found xmin %u from before relfrozenxid %u",
6126                                                                          xid, relfrozenxid)));
6127
6128                 xmin_frozen = TransactionIdPrecedes(xid, cutoff_xid);
6129                 if (xmin_frozen)
6130                 {
6131                         if (!TransactionIdDidCommit(xid))
6132                                 ereport(ERROR,
6133                                                 (errcode(ERRCODE_DATA_CORRUPTED),
6134                                                  errmsg_internal("uncommitted xmin %u from before xid cutoff %u needs to be frozen",
6135                                                                                  xid, cutoff_xid)));
6136
6137                         frz->t_infomask |= HEAP_XMIN_FROZEN;
6138                         changed = true;
6139                 }
6140         }
6141
6142         /*
6143          * Process xmax.  To thoroughly examine the current Xmax value we need to
6144          * resolve a MultiXactId to its member Xids, in case some of them are
6145          * below the given cutoff for Xids.  In that case, those values might need
6146          * freezing, too.  Also, if a multi needs freezing, we cannot simply take
6147          * it out --- if there's a live updater Xid, it needs to be kept.
6148          *
6149          * Make sure to keep heap_tuple_needs_freeze in sync with this.
6150          */
6151         xid = HeapTupleHeaderGetRawXmax(tuple);
6152
6153         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
6154         {
6155                 TransactionId newxmax;
6156                 uint16          flags;
6157
6158                 newxmax = FreezeMultiXactId(xid, tuple->t_infomask,
6159                                                                         relfrozenxid, relminmxid,
6160                                                                         cutoff_xid, cutoff_multi, &flags);
6161
6162                 freeze_xmax = (flags & FRM_INVALIDATE_XMAX);
6163
6164                 if (flags & FRM_RETURN_IS_XID)
6165                 {
6166                         /*
6167                          * NB -- some of these transformations are only valid because we
6168                          * know the return Xid is a tuple updater (i.e. not merely a
6169                          * locker.) Also note that the only reason we don't explicitly
6170                          * worry about HEAP_KEYS_UPDATED is because it lives in
6171                          * t_infomask2 rather than t_infomask.
6172                          */
6173                         frz->t_infomask &= ~HEAP_XMAX_BITS;
6174                         frz->xmax = newxmax;
6175                         if (flags & FRM_MARK_COMMITTED)
6176                                 frz->t_infomask |= HEAP_XMAX_COMMITTED;
6177                         changed = true;
6178                 }
6179                 else if (flags & FRM_RETURN_IS_MULTI)
6180                 {
6181                         uint16          newbits;
6182                         uint16          newbits2;
6183
6184                         /*
6185                          * We can't use GetMultiXactIdHintBits directly on the new multi
6186                          * here; that routine initializes the masks to all zeroes, which
6187                          * would lose other bits we need.  Doing it this way ensures all
6188                          * unrelated bits remain untouched.
6189                          */
6190                         frz->t_infomask &= ~HEAP_XMAX_BITS;
6191                         frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
6192                         GetMultiXactIdHintBits(newxmax, &newbits, &newbits2);
6193                         frz->t_infomask |= newbits;
6194                         frz->t_infomask2 |= newbits2;
6195
6196                         frz->xmax = newxmax;
6197
6198                         changed = true;
6199                 }
6200         }
6201         else if (TransactionIdIsNormal(xid))
6202         {
6203                 if (TransactionIdPrecedes(xid, relfrozenxid))
6204                         ereport(ERROR,
6205                                         (errcode(ERRCODE_DATA_CORRUPTED),
6206                                          errmsg_internal("found xmax %u from before relfrozenxid %u",
6207                                                                          xid, relfrozenxid)));
6208
6209                 if (TransactionIdPrecedes(xid, cutoff_xid))
6210                 {
6211                         /*
6212                          * If we freeze xmax, make absolutely sure that it's not an XID
6213                          * that is important.  (Note, a lock-only xmax can be removed
6214                          * independent of committedness, since a committed lock holder has
6215                          * released the lock).
6216                          */
6217                         if (!HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask) &&
6218                                 TransactionIdDidCommit(xid))
6219                                 ereport(ERROR,
6220                                                 (errcode(ERRCODE_DATA_CORRUPTED),
6221                                                  errmsg_internal("cannot freeze committed xmax %u",
6222                                                                                  xid)));
6223                         freeze_xmax = true;
6224                 }
6225                 else
6226                         freeze_xmax = false;
6227         }
6228         else if ((tuple->t_infomask & HEAP_XMAX_INVALID) ||
6229                          !TransactionIdIsValid(HeapTupleHeaderGetRawXmax(tuple)))
6230         {
6231                 freeze_xmax = false;
6232                 xmax_already_frozen = true;
6233         }
6234         else
6235                 ereport(ERROR,
6236                                 (errcode(ERRCODE_DATA_CORRUPTED),
6237                                  errmsg_internal("found xmax %u (infomask 0x%04x) not frozen, not multi, not normal",
6238                                                                  xid, tuple->t_infomask)));
6239
6240         if (freeze_xmax)
6241         {
6242                 Assert(!xmax_already_frozen);
6243
6244                 frz->xmax = InvalidTransactionId;
6245
6246                 /*
6247                  * The tuple might be marked either XMAX_INVALID or XMAX_COMMITTED +
6248                  * LOCKED.  Normalize to INVALID just to be sure no one gets confused.
6249                  * Also get rid of the HEAP_KEYS_UPDATED bit.
6250                  */
6251                 frz->t_infomask &= ~HEAP_XMAX_BITS;
6252                 frz->t_infomask |= HEAP_XMAX_INVALID;
6253                 frz->t_infomask2 &= ~HEAP_HOT_UPDATED;
6254                 frz->t_infomask2 &= ~HEAP_KEYS_UPDATED;
6255                 changed = true;
6256         }
6257
6258         /*
6259          * Old-style VACUUM FULL is gone, but we have to keep this code as long as
6260          * we support having MOVED_OFF/MOVED_IN tuples in the database.
6261          */
6262         if (tuple->t_infomask & HEAP_MOVED)
6263         {
6264                 xid = HeapTupleHeaderGetXvac(tuple);
6265
6266                 /*
6267                  * For Xvac, we ignore the cutoff_xid and just always perform the
6268                  * freeze operation.  The oldest release in which such a value can
6269                  * actually be set is PostgreSQL 8.4, because old-style VACUUM FULL
6270                  * was removed in PostgreSQL 9.0.  Note that if we were to respect
6271                  * cutoff_xid here, we'd need to make surely to clear totally_frozen
6272                  * when we skipped freezing on that basis.
6273                  */
6274                 if (TransactionIdIsNormal(xid))
6275                 {
6276                         /*
6277                          * If a MOVED_OFF tuple is not dead, the xvac transaction must
6278                          * have failed; whereas a non-dead MOVED_IN tuple must mean the
6279                          * xvac transaction succeeded.
6280                          */
6281                         if (tuple->t_infomask & HEAP_MOVED_OFF)
6282                                 frz->frzflags |= XLH_INVALID_XVAC;
6283                         else
6284                                 frz->frzflags |= XLH_FREEZE_XVAC;
6285
6286                         /*
6287                          * Might as well fix the hint bits too; usually XMIN_COMMITTED
6288                          * will already be set here, but there's a small chance not.
6289                          */
6290                         Assert(!(tuple->t_infomask & HEAP_XMIN_INVALID));
6291                         frz->t_infomask |= HEAP_XMIN_COMMITTED;
6292                         changed = true;
6293                 }
6294         }
6295
6296         *totally_frozen_p = (xmin_frozen &&
6297                                                  (freeze_xmax || xmax_already_frozen));
6298         return changed;
6299 }
6300
6301 /*
6302  * heap_execute_freeze_tuple
6303  *              Execute the prepared freezing of a tuple.
6304  *
6305  * Caller is responsible for ensuring that no other backend can access the
6306  * storage underlying this tuple, either by holding an exclusive lock on the
6307  * buffer containing it (which is what lazy VACUUM does), or by having it be
6308  * in private storage (which is what CLUSTER and friends do).
6309  *
6310  * Note: it might seem we could make the changes without exclusive lock, since
6311  * TransactionId read/write is assumed atomic anyway.  However there is a race
6312  * condition: someone who just fetched an old XID that we overwrite here could
6313  * conceivably not finish checking the XID against pg_xact before we finish
6314  * the VACUUM and perhaps truncate off the part of pg_xact he needs.  Getting
6315  * exclusive lock ensures no other backend is in process of checking the
6316  * tuple status.  Also, getting exclusive lock makes it safe to adjust the
6317  * infomask bits.
6318  *
6319  * NB: All code in here must be safe to execute during crash recovery!
6320  */
6321 void
6322 heap_execute_freeze_tuple(HeapTupleHeader tuple, xl_heap_freeze_tuple *frz)
6323 {
6324         HeapTupleHeaderSetXmax(tuple, frz->xmax);
6325
6326         if (frz->frzflags & XLH_FREEZE_XVAC)
6327                 HeapTupleHeaderSetXvac(tuple, FrozenTransactionId);
6328
6329         if (frz->frzflags & XLH_INVALID_XVAC)
6330                 HeapTupleHeaderSetXvac(tuple, InvalidTransactionId);
6331
6332         tuple->t_infomask = frz->t_infomask;
6333         tuple->t_infomask2 = frz->t_infomask2;
6334 }
6335
6336 /*
6337  * heap_freeze_tuple
6338  *              Freeze tuple in place, without WAL logging.
6339  *
6340  * Useful for callers like CLUSTER that perform their own WAL logging.
6341  */
6342 bool
6343 heap_freeze_tuple(HeapTupleHeader tuple,
6344                                   TransactionId relfrozenxid, TransactionId relminmxid,
6345                                   TransactionId cutoff_xid, TransactionId cutoff_multi)
6346 {
6347         xl_heap_freeze_tuple frz;
6348         bool            do_freeze;
6349         bool            tuple_totally_frozen;
6350
6351         do_freeze = heap_prepare_freeze_tuple(tuple,
6352                                                                                   relfrozenxid, relminmxid,
6353                                                                                   cutoff_xid, cutoff_multi,
6354                                                                                   &frz, &tuple_totally_frozen);
6355
6356         /*
6357          * Note that because this is not a WAL-logged operation, we don't need to
6358          * fill in the offset in the freeze record.
6359          */
6360
6361         if (do_freeze)
6362                 heap_execute_freeze_tuple(tuple, &frz);
6363         return do_freeze;
6364 }
6365
6366 /*
6367  * For a given MultiXactId, return the hint bits that should be set in the
6368  * tuple's infomask.
6369  *
6370  * Normally this should be called for a multixact that was just created, and
6371  * so is on our local cache, so the GetMembers call is fast.
6372  */
6373 static void
6374 GetMultiXactIdHintBits(MultiXactId multi, uint16 *new_infomask,
6375                                            uint16 *new_infomask2)
6376 {
6377         int                     nmembers;
6378         MultiXactMember *members;
6379         int                     i;
6380         uint16          bits = HEAP_XMAX_IS_MULTI;
6381         uint16          bits2 = 0;
6382         bool            has_update = false;
6383         LockTupleMode strongest = LockTupleKeyShare;
6384
6385         /*
6386          * We only use this in multis we just created, so they cannot be values
6387          * pre-pg_upgrade.
6388          */
6389         nmembers = GetMultiXactIdMembers(multi, &members, false, false);
6390
6391         for (i = 0; i < nmembers; i++)
6392         {
6393                 LockTupleMode mode;
6394
6395                 /*
6396                  * Remember the strongest lock mode held by any member of the
6397                  * multixact.
6398                  */
6399                 mode = TUPLOCK_from_mxstatus(members[i].status);
6400                 if (mode > strongest)
6401                         strongest = mode;
6402
6403                 /* See what other bits we need */
6404                 switch (members[i].status)
6405                 {
6406                         case MultiXactStatusForKeyShare:
6407                         case MultiXactStatusForShare:
6408                         case MultiXactStatusForNoKeyUpdate:
6409                                 break;
6410
6411                         case MultiXactStatusForUpdate:
6412                                 bits2 |= HEAP_KEYS_UPDATED;
6413                                 break;
6414
6415                         case MultiXactStatusNoKeyUpdate:
6416                                 has_update = true;
6417                                 break;
6418
6419                         case MultiXactStatusUpdate:
6420                                 bits2 |= HEAP_KEYS_UPDATED;
6421                                 has_update = true;
6422                                 break;
6423                 }
6424         }
6425
6426         if (strongest == LockTupleExclusive ||
6427                 strongest == LockTupleNoKeyExclusive)
6428                 bits |= HEAP_XMAX_EXCL_LOCK;
6429         else if (strongest == LockTupleShare)
6430                 bits |= HEAP_XMAX_SHR_LOCK;
6431         else if (strongest == LockTupleKeyShare)
6432                 bits |= HEAP_XMAX_KEYSHR_LOCK;
6433
6434         if (!has_update)
6435                 bits |= HEAP_XMAX_LOCK_ONLY;
6436
6437         if (nmembers > 0)
6438                 pfree(members);
6439
6440         *new_infomask = bits;
6441         *new_infomask2 = bits2;
6442 }
6443
6444 /*
6445  * MultiXactIdGetUpdateXid
6446  *
6447  * Given a multixact Xmax and corresponding infomask, which does not have the
6448  * HEAP_XMAX_LOCK_ONLY bit set, obtain and return the Xid of the updating
6449  * transaction.
6450  *
6451  * Caller is expected to check the status of the updating transaction, if
6452  * necessary.
6453  */
6454 static TransactionId
6455 MultiXactIdGetUpdateXid(TransactionId xmax, uint16 t_infomask)
6456 {
6457         TransactionId update_xact = InvalidTransactionId;
6458         MultiXactMember *members;
6459         int                     nmembers;
6460
6461         Assert(!(t_infomask & HEAP_XMAX_LOCK_ONLY));
6462         Assert(t_infomask & HEAP_XMAX_IS_MULTI);
6463
6464         /*
6465          * Since we know the LOCK_ONLY bit is not set, this cannot be a multi from
6466          * pre-pg_upgrade.
6467          */
6468         nmembers = GetMultiXactIdMembers(xmax, &members, false, false);
6469
6470         if (nmembers > 0)
6471         {
6472                 int                     i;
6473
6474                 for (i = 0; i < nmembers; i++)
6475                 {
6476                         /* Ignore lockers */
6477                         if (!ISUPDATE_from_mxstatus(members[i].status))
6478                                 continue;
6479
6480                         /* there can be at most one updater */
6481                         Assert(update_xact == InvalidTransactionId);
6482                         update_xact = members[i].xid;
6483 #ifndef USE_ASSERT_CHECKING
6484
6485                         /*
6486                          * in an assert-enabled build, walk the whole array to ensure
6487                          * there's no other updater.
6488                          */
6489                         break;
6490 #endif
6491                 }
6492
6493                 pfree(members);
6494         }
6495
6496         return update_xact;
6497 }
6498
6499 /*
6500  * HeapTupleGetUpdateXid
6501  *              As above, but use a HeapTupleHeader
6502  *
6503  * See also HeapTupleHeaderGetUpdateXid, which can be used without previously
6504  * checking the hint bits.
6505  */
6506 TransactionId
6507 HeapTupleGetUpdateXid(HeapTupleHeader tuple)
6508 {
6509         return MultiXactIdGetUpdateXid(HeapTupleHeaderGetRawXmax(tuple),
6510                                                                    tuple->t_infomask);
6511 }
6512
6513 /*
6514  * Does the given multixact conflict with the current transaction grabbing a
6515  * tuple lock of the given strength?
6516  *
6517  * The passed infomask pairs up with the given multixact in the tuple header.
6518  */
6519 static bool
6520 DoesMultiXactIdConflict(MultiXactId multi, uint16 infomask,
6521                                                 LockTupleMode lockmode)
6522 {
6523         int                     nmembers;
6524         MultiXactMember *members;
6525         bool            result = false;
6526         LOCKMODE        wanted = tupleLockExtraInfo[lockmode].hwlock;
6527
6528         if (HEAP_LOCKED_UPGRADED(infomask))
6529                 return false;
6530
6531         nmembers = GetMultiXactIdMembers(multi, &members, false,
6532                                                                          HEAP_XMAX_IS_LOCKED_ONLY(infomask));
6533         if (nmembers >= 0)
6534         {
6535                 int                     i;
6536
6537                 for (i = 0; i < nmembers; i++)
6538                 {
6539                         TransactionId memxid;
6540                         LOCKMODE        memlockmode;
6541
6542                         memlockmode = LOCKMODE_from_mxstatus(members[i].status);
6543
6544                         /* ignore members that don't conflict with the lock we want */
6545                         if (!DoLockModesConflict(memlockmode, wanted))
6546                                 continue;
6547
6548                         /* ignore members from current xact */
6549                         memxid = members[i].xid;
6550                         if (TransactionIdIsCurrentTransactionId(memxid))
6551                                 continue;
6552
6553                         if (ISUPDATE_from_mxstatus(members[i].status))
6554                         {
6555                                 /* ignore aborted updaters */
6556                                 if (TransactionIdDidAbort(memxid))
6557                                         continue;
6558                         }
6559                         else
6560                         {
6561                                 /* ignore lockers-only that are no longer in progress */
6562                                 if (!TransactionIdIsInProgress(memxid))
6563                                         continue;
6564                         }
6565
6566                         /*
6567                          * Whatever remains are either live lockers that conflict with our
6568                          * wanted lock, and updaters that are not aborted.  Those conflict
6569                          * with what we want, so return true.
6570                          */
6571                         result = true;
6572                         break;
6573                 }
6574                 pfree(members);
6575         }
6576
6577         return result;
6578 }
6579
6580 /*
6581  * Do_MultiXactIdWait
6582  *              Actual implementation for the two functions below.
6583  *
6584  * 'multi', 'status' and 'infomask' indicate what to sleep on (the status is
6585  * needed to ensure we only sleep on conflicting members, and the infomask is
6586  * used to optimize multixact access in case it's a lock-only multi); 'nowait'
6587  * indicates whether to use conditional lock acquisition, to allow callers to
6588  * fail if lock is unavailable.  'rel', 'ctid' and 'oper' are used to set up
6589  * context information for error messages.  'remaining', if not NULL, receives
6590  * the number of members that are still running, including any (non-aborted)
6591  * subtransactions of our own transaction.
6592  *
6593  * We do this by sleeping on each member using XactLockTableWait.  Any
6594  * members that belong to the current backend are *not* waited for, however;
6595  * this would not merely be useless but would lead to Assert failure inside
6596  * XactLockTableWait.  By the time this returns, it is certain that all
6597  * transactions *of other backends* that were members of the MultiXactId
6598  * that conflict with the requested status are dead (and no new ones can have
6599  * been added, since it is not legal to add members to an existing
6600  * MultiXactId).
6601  *
6602  * But by the time we finish sleeping, someone else may have changed the Xmax
6603  * of the containing tuple, so the caller needs to iterate on us somehow.
6604  *
6605  * Note that in case we return false, the number of remaining members is
6606  * not to be trusted.
6607  */
6608 static bool
6609 Do_MultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6610                                    uint16 infomask, bool nowait,
6611                                    Relation rel, ItemPointer ctid, XLTW_Oper oper,
6612                                    int *remaining)
6613 {
6614         bool            result = true;
6615         MultiXactMember *members;
6616         int                     nmembers;
6617         int                     remain = 0;
6618
6619         /* for pre-pg_upgrade tuples, no need to sleep at all */
6620         nmembers = HEAP_LOCKED_UPGRADED(infomask) ? -1 :
6621                 GetMultiXactIdMembers(multi, &members, false,
6622                                                           HEAP_XMAX_IS_LOCKED_ONLY(infomask));
6623
6624         if (nmembers >= 0)
6625         {
6626                 int                     i;
6627
6628                 for (i = 0; i < nmembers; i++)
6629                 {
6630                         TransactionId memxid = members[i].xid;
6631                         MultiXactStatus memstatus = members[i].status;
6632
6633                         if (TransactionIdIsCurrentTransactionId(memxid))
6634                         {
6635                                 remain++;
6636                                 continue;
6637                         }
6638
6639                         if (!DoLockModesConflict(LOCKMODE_from_mxstatus(memstatus),
6640                                                                          LOCKMODE_from_mxstatus(status)))
6641                         {
6642                                 if (remaining && TransactionIdIsInProgress(memxid))
6643                                         remain++;
6644                                 continue;
6645                         }
6646
6647                         /*
6648                          * This member conflicts with our multi, so we have to sleep (or
6649                          * return failure, if asked to avoid waiting.)
6650                          *
6651                          * Note that we don't set up an error context callback ourselves,
6652                          * but instead we pass the info down to XactLockTableWait.  This
6653                          * might seem a bit wasteful because the context is set up and
6654                          * tore down for each member of the multixact, but in reality it
6655                          * should be barely noticeable, and it avoids duplicate code.
6656                          */
6657                         if (nowait)
6658                         {
6659                                 result = ConditionalXactLockTableWait(memxid);
6660                                 if (!result)
6661                                         break;
6662                         }
6663                         else
6664                                 XactLockTableWait(memxid, rel, ctid, oper);
6665                 }
6666
6667                 pfree(members);
6668         }
6669
6670         if (remaining)
6671                 *remaining = remain;
6672
6673         return result;
6674 }
6675
6676 /*
6677  * MultiXactIdWait
6678  *              Sleep on a MultiXactId.
6679  *
6680  * By the time we finish sleeping, someone else may have changed the Xmax
6681  * of the containing tuple, so the caller needs to iterate on us somehow.
6682  *
6683  * We return (in *remaining, if not NULL) the number of members that are still
6684  * running, including any (non-aborted) subtransactions of our own transaction.
6685  */
6686 static void
6687 MultiXactIdWait(MultiXactId multi, MultiXactStatus status, uint16 infomask,
6688                                 Relation rel, ItemPointer ctid, XLTW_Oper oper,
6689                                 int *remaining)
6690 {
6691         (void) Do_MultiXactIdWait(multi, status, infomask, false,
6692                                                           rel, ctid, oper, remaining);
6693 }
6694
6695 /*
6696  * ConditionalMultiXactIdWait
6697  *              As above, but only lock if we can get the lock without blocking.
6698  *
6699  * By the time we finish sleeping, someone else may have changed the Xmax
6700  * of the containing tuple, so the caller needs to iterate on us somehow.
6701  *
6702  * If the multixact is now all gone, return true.  Returns false if some
6703  * transactions might still be running.
6704  *
6705  * We return (in *remaining, if not NULL) the number of members that are still
6706  * running, including any (non-aborted) subtransactions of our own transaction.
6707  */
6708 static bool
6709 ConditionalMultiXactIdWait(MultiXactId multi, MultiXactStatus status,
6710                                                    uint16 infomask, Relation rel, int *remaining)
6711 {
6712         return Do_MultiXactIdWait(multi, status, infomask, true,
6713                                                           rel, NULL, XLTW_None, remaining);
6714 }
6715
6716 /*
6717  * heap_tuple_needs_eventual_freeze
6718  *
6719  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6720  * will eventually require freezing.  Similar to heap_tuple_needs_freeze,
6721  * but there's no cutoff, since we're trying to figure out whether freezing
6722  * will ever be needed, not whether it's needed now.
6723  */
6724 bool
6725 heap_tuple_needs_eventual_freeze(HeapTupleHeader tuple)
6726 {
6727         TransactionId xid;
6728
6729         /*
6730          * If xmin is a normal transaction ID, this tuple is definitely not
6731          * frozen.
6732          */
6733         xid = HeapTupleHeaderGetXmin(tuple);
6734         if (TransactionIdIsNormal(xid))
6735                 return true;
6736
6737         /*
6738          * If xmax is a valid xact or multixact, this tuple is also not frozen.
6739          */
6740         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
6741         {
6742                 MultiXactId multi;
6743
6744                 multi = HeapTupleHeaderGetRawXmax(tuple);
6745                 if (MultiXactIdIsValid(multi))
6746                         return true;
6747         }
6748         else
6749         {
6750                 xid = HeapTupleHeaderGetRawXmax(tuple);
6751                 if (TransactionIdIsNormal(xid))
6752                         return true;
6753         }
6754
6755         if (tuple->t_infomask & HEAP_MOVED)
6756         {
6757                 xid = HeapTupleHeaderGetXvac(tuple);
6758                 if (TransactionIdIsNormal(xid))
6759                         return true;
6760         }
6761
6762         return false;
6763 }
6764
6765 /*
6766  * heap_tuple_needs_freeze
6767  *
6768  * Check to see whether any of the XID fields of a tuple (xmin, xmax, xvac)
6769  * are older than the specified cutoff XID or MultiXactId.  If so, return true.
6770  *
6771  * It doesn't matter whether the tuple is alive or dead, we are checking
6772  * to see if a tuple needs to be removed or frozen to avoid wraparound.
6773  *
6774  * NB: Cannot rely on hint bits here, they might not be set after a crash or
6775  * on a standby.
6776  */
6777 bool
6778 heap_tuple_needs_freeze(HeapTupleHeader tuple, TransactionId cutoff_xid,
6779                                                 MultiXactId cutoff_multi, Buffer buf)
6780 {
6781         TransactionId xid;
6782
6783         xid = HeapTupleHeaderGetXmin(tuple);
6784         if (TransactionIdIsNormal(xid) &&
6785                 TransactionIdPrecedes(xid, cutoff_xid))
6786                 return true;
6787
6788         /*
6789          * The considerations for multixacts are complicated; look at
6790          * heap_prepare_freeze_tuple for justifications.  This routine had better
6791          * be in sync with that one!
6792          */
6793         if (tuple->t_infomask & HEAP_XMAX_IS_MULTI)
6794         {
6795                 MultiXactId multi;
6796
6797                 multi = HeapTupleHeaderGetRawXmax(tuple);
6798                 if (!MultiXactIdIsValid(multi))
6799                 {
6800                         /* no xmax set, ignore */
6801                         ;
6802                 }
6803                 else if (HEAP_LOCKED_UPGRADED(tuple->t_infomask))
6804                         return true;
6805                 else if (MultiXactIdPrecedes(multi, cutoff_multi))
6806                         return true;
6807                 else
6808                 {
6809                         MultiXactMember *members;
6810                         int                     nmembers;
6811                         int                     i;
6812
6813                         /* need to check whether any member of the mxact is too old */
6814
6815                         nmembers = GetMultiXactIdMembers(multi, &members, false,
6816                                                                                          HEAP_XMAX_IS_LOCKED_ONLY(tuple->t_infomask));
6817
6818                         for (i = 0; i < nmembers; i++)
6819                         {
6820                                 if (TransactionIdPrecedes(members[i].xid, cutoff_xid))
6821                                 {
6822                                         pfree(members);
6823                                         return true;
6824                                 }
6825                         }
6826                         if (nmembers > 0)
6827                                 pfree(members);
6828                 }
6829         }
6830         else
6831         {
6832                 xid = HeapTupleHeaderGetRawXmax(tuple);
6833                 if (TransactionIdIsNormal(xid) &&
6834                         TransactionIdPrecedes(xid, cutoff_xid))
6835                         return true;
6836         }
6837
6838         if (tuple->t_infomask & HEAP_MOVED)
6839         {
6840                 xid = HeapTupleHeaderGetXvac(tuple);
6841                 if (TransactionIdIsNormal(xid) &&
6842                         TransactionIdPrecedes(xid, cutoff_xid))
6843                         return true;
6844         }
6845
6846         return false;
6847 }
6848
6849 /*
6850  * If 'tuple' contains any visible XID greater than latestRemovedXid,
6851  * ratchet forwards latestRemovedXid to the greatest one found.
6852  * This is used as the basis for generating Hot Standby conflicts, so
6853  * if a tuple was never visible then removing it should not conflict
6854  * with queries.
6855  */
6856 void
6857 HeapTupleHeaderAdvanceLatestRemovedXid(HeapTupleHeader tuple,
6858                                                                            TransactionId *latestRemovedXid)
6859 {
6860         TransactionId xmin = HeapTupleHeaderGetXmin(tuple);
6861         TransactionId xmax = HeapTupleHeaderGetUpdateXid(tuple);
6862         TransactionId xvac = HeapTupleHeaderGetXvac(tuple);
6863
6864         if (tuple->t_infomask & HEAP_MOVED)
6865         {
6866                 if (TransactionIdPrecedes(*latestRemovedXid, xvac))
6867                         *latestRemovedXid = xvac;
6868         }
6869
6870         /*
6871          * Ignore tuples inserted by an aborted transaction or if the tuple was
6872          * updated/deleted by the inserting transaction.
6873          *
6874          * Look for a committed hint bit, or if no xmin bit is set, check clog.
6875          * This needs to work on both master and standby, where it is used to
6876          * assess btree delete records.
6877          */
6878         if (HeapTupleHeaderXminCommitted(tuple) ||
6879                 (!HeapTupleHeaderXminInvalid(tuple) && TransactionIdDidCommit(xmin)))
6880         {
6881                 if (xmax != xmin &&
6882                         TransactionIdFollows(xmax, *latestRemovedXid))
6883                         *latestRemovedXid = xmax;
6884         }
6885
6886         /* *latestRemovedXid may still be invalid at end */
6887 }
6888
6889 #ifdef USE_PREFETCH
6890 /*
6891  * Helper function for heap_compute_xid_horizon_for_tuples.  Issue prefetch
6892  * requests for the number of buffers indicated by prefetch_count.  The
6893  * prefetch_state keeps track of all the buffers that we can prefetch and
6894  * which ones have already been prefetched; each call to this function picks
6895  * up where the previous call left off.
6896  */
6897 static void
6898 xid_horizon_prefetch_buffer(Relation rel,
6899                                                         XidHorizonPrefetchState *prefetch_state,
6900                                                         int prefetch_count)
6901 {
6902         BlockNumber cur_hblkno = prefetch_state->cur_hblkno;
6903         int                     count = 0;
6904         int                     i;
6905         int                     nitems = prefetch_state->nitems;
6906         ItemPointerData *tids = prefetch_state->tids;
6907
6908         for (i = prefetch_state->next_item;
6909                  i < nitems && count < prefetch_count;
6910                  i++)
6911         {
6912                 ItemPointer htid = &tids[i];
6913
6914                 if (cur_hblkno == InvalidBlockNumber ||
6915                         ItemPointerGetBlockNumber(htid) != cur_hblkno)
6916                 {
6917                         cur_hblkno = ItemPointerGetBlockNumber(htid);
6918                         PrefetchBuffer(rel, MAIN_FORKNUM, cur_hblkno);
6919                         count++;
6920                 }
6921         }
6922
6923         /*
6924          * Save the prefetch position so that next time we can continue from that
6925          * position.
6926          */
6927         prefetch_state->next_item = i;
6928         prefetch_state->cur_hblkno = cur_hblkno;
6929 }
6930 #endif
6931
6932 /*
6933  * Get the latestRemovedXid from the heap pages pointed at by the index
6934  * tuples being deleted.
6935  *
6936  * We used to do this during recovery rather than on the primary, but that
6937  * approach now appears inferior.  It meant that the master could generate
6938  * a lot of work for the standby without any back-pressure to slow down the
6939  * master, and it required the standby to have reached consistency, whereas
6940  * we want to have correct information available even before that point.
6941  *
6942  * It's possible for this to generate a fair amount of I/O, since we may be
6943  * deleting hundreds of tuples from a single index block.  To amortize that
6944  * cost to some degree, this uses prefetching and combines repeat accesses to
6945  * the same block.
6946  */
6947 TransactionId
6948 heap_compute_xid_horizon_for_tuples(Relation rel,
6949                                                                         ItemPointerData *tids,
6950                                                                         int nitems)
6951 {
6952         TransactionId latestRemovedXid = InvalidTransactionId;
6953         BlockNumber hblkno;
6954         Buffer          buf = InvalidBuffer;
6955         Page            hpage;
6956 #ifdef USE_PREFETCH
6957         XidHorizonPrefetchState prefetch_state;
6958         int                     io_concurrency;
6959         int                     prefetch_distance;
6960 #endif
6961
6962         /*
6963          * Sort to avoid repeated lookups for the same page, and to make it more
6964          * likely to access items in an efficient order. In particular, this
6965          * ensures that if there are multiple pointers to the same page, they all
6966          * get processed looking up and locking the page just once.
6967          */
6968         qsort((void *) tids, nitems, sizeof(ItemPointerData),
6969                   (int (*) (const void *, const void *)) ItemPointerCompare);
6970
6971 #ifdef USE_PREFETCH
6972         /* Initialize prefetch state. */
6973         prefetch_state.cur_hblkno = InvalidBlockNumber;
6974         prefetch_state.next_item = 0;
6975         prefetch_state.nitems = nitems;
6976         prefetch_state.tids = tids;
6977
6978         /*
6979          * Compute the prefetch distance that we will attempt to maintain.
6980          *
6981          * We don't use the regular formula to determine how much to prefetch
6982          * here, but instead just add a constant to effective_io_concurrency.
6983          * That's because it seems best to do some prefetching here even when
6984          * effective_io_concurrency is set to 0, but if the DBA thinks it's OK to
6985          * do more prefetching for other operations, then it's probably OK to do
6986          * more prefetching in this case, too. It may be that this formula is too
6987          * simplistic, but at the moment there is no evidence of that or any idea
6988          * about what would work better.
6989          *
6990          * Since the caller holds a buffer lock somewhere in rel, we'd better make
6991          * sure that isn't a catalog relation before we call code that does
6992          * syscache lookups, to avoid risk of deadlock.
6993          */
6994         if (IsCatalogRelation(rel))
6995                 io_concurrency = effective_io_concurrency;
6996         else
6997                 io_concurrency = get_tablespace_io_concurrency(rel->rd_rel->reltablespace);
6998         prefetch_distance = Min((io_concurrency) + 10, MAX_IO_CONCURRENCY);
6999
7000         /* Start prefetching. */
7001         xid_horizon_prefetch_buffer(rel, &prefetch_state, prefetch_distance);
7002 #endif
7003
7004         /* Iterate over all tids, and check their horizon */
7005         hblkno = InvalidBlockNumber;
7006         hpage = NULL;
7007         for (int i = 0; i < nitems; i++)
7008         {
7009                 ItemPointer htid = &tids[i];
7010                 ItemId          hitemid;
7011                 OffsetNumber hoffnum;
7012
7013                 /*
7014                  * Read heap buffer, but avoid refetching if it's the same block as
7015                  * required for the last tid.
7016                  */
7017                 if (hblkno == InvalidBlockNumber ||
7018                         ItemPointerGetBlockNumber(htid) != hblkno)
7019                 {
7020                         /* release old buffer */
7021                         if (BufferIsValid(buf))
7022                         {
7023                                 LockBuffer(buf, BUFFER_LOCK_UNLOCK);
7024                                 ReleaseBuffer(buf);
7025                         }
7026
7027                         hblkno = ItemPointerGetBlockNumber(htid);
7028
7029                         buf = ReadBuffer(rel, hblkno);
7030
7031 #ifdef USE_PREFETCH
7032
7033                         /*
7034                          * To maintain the prefetch distance, prefetch one more page for
7035                          * each page we read.
7036                          */
7037                         xid_horizon_prefetch_buffer(rel, &prefetch_state, 1);
7038 #endif
7039
7040                         hpage = BufferGetPage(buf);
7041
7042                         LockBuffer(buf, BUFFER_LOCK_SHARE);
7043                 }
7044
7045                 hoffnum = ItemPointerGetOffsetNumber(htid);
7046                 hitemid = PageGetItemId(hpage, hoffnum);
7047
7048                 /*
7049                  * Follow any redirections until we find something useful.
7050                  */
7051                 while (ItemIdIsRedirected(hitemid))
7052                 {
7053                         hoffnum = ItemIdGetRedirect(hitemid);
7054                         hitemid = PageGetItemId(hpage, hoffnum);
7055                         CHECK_FOR_INTERRUPTS();
7056                 }
7057
7058                 /*
7059                  * If the heap item has storage, then read the header and use that to
7060                  * set latestRemovedXid.
7061                  *
7062                  * Some LP_DEAD items may not be accessible, so we ignore them.
7063                  */
7064                 if (ItemIdHasStorage(hitemid))
7065                 {
7066                         HeapTupleHeader htuphdr;
7067
7068                         htuphdr = (HeapTupleHeader) PageGetItem(hpage, hitemid);
7069
7070                         HeapTupleHeaderAdvanceLatestRemovedXid(htuphdr, &latestRemovedXid);
7071                 }
7072                 else if (ItemIdIsDead(hitemid))
7073                 {
7074                         /*
7075                          * Conjecture: if hitemid is dead then it had xids before the xids
7076                          * marked on LP_NORMAL items. So we just ignore this item and move
7077                          * onto the next, for the purposes of calculating
7078                          * latestRemovedxids.
7079                          */
7080                 }
7081                 else
7082                         Assert(!ItemIdIsUsed(hitemid));
7083
7084         }
7085
7086         if (BufferIsValid(buf))
7087         {
7088                 LockBuffer(buf, BUFFER_LOCK_UNLOCK);
7089                 ReleaseBuffer(buf);
7090         }
7091
7092         /*
7093          * If all heap tuples were LP_DEAD then we will be returning
7094          * InvalidTransactionId here, which avoids conflicts. This matches
7095          * existing logic which assumes that LP_DEAD tuples must already be older
7096          * than the latestRemovedXid on the cleanup record that set them as
7097          * LP_DEAD, hence must already have generated a conflict.
7098          */
7099
7100         return latestRemovedXid;
7101 }
7102
7103 /*
7104  * Perform XLogInsert to register a heap cleanup info message. These
7105  * messages are sent once per VACUUM and are required because
7106  * of the phasing of removal operations during a lazy VACUUM.
7107  * see comments for vacuum_log_cleanup_info().
7108  */
7109 XLogRecPtr
7110 log_heap_cleanup_info(RelFileNode rnode, TransactionId latestRemovedXid)
7111 {
7112         xl_heap_cleanup_info xlrec;
7113         XLogRecPtr      recptr;
7114
7115         xlrec.node = rnode;
7116         xlrec.latestRemovedXid = latestRemovedXid;
7117
7118         XLogBeginInsert();
7119         XLogRegisterData((char *) &xlrec, SizeOfHeapCleanupInfo);
7120
7121         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_CLEANUP_INFO);
7122
7123         return recptr;
7124 }
7125
7126 /*
7127  * Perform XLogInsert for a heap-clean operation.  Caller must already
7128  * have modified the buffer and marked it dirty.
7129  *
7130  * Note: prior to Postgres 8.3, the entries in the nowunused[] array were
7131  * zero-based tuple indexes.  Now they are one-based like other uses
7132  * of OffsetNumber.
7133  *
7134  * We also include latestRemovedXid, which is the greatest XID present in
7135  * the removed tuples. That allows recovery processing to cancel or wait
7136  * for long standby queries that can still see these tuples.
7137  */
7138 XLogRecPtr
7139 log_heap_clean(Relation reln, Buffer buffer,
7140                            OffsetNumber *redirected, int nredirected,
7141                            OffsetNumber *nowdead, int ndead,
7142                            OffsetNumber *nowunused, int nunused,
7143                            TransactionId latestRemovedXid)
7144 {
7145         xl_heap_clean xlrec;
7146         XLogRecPtr      recptr;
7147
7148         /* Caller should not call me on a non-WAL-logged relation */
7149         Assert(RelationNeedsWAL(reln));
7150
7151         xlrec.latestRemovedXid = latestRemovedXid;
7152         xlrec.nredirected = nredirected;
7153         xlrec.ndead = ndead;
7154
7155         XLogBeginInsert();
7156         XLogRegisterData((char *) &xlrec, SizeOfHeapClean);
7157
7158         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
7159
7160         /*
7161          * The OffsetNumber arrays are not actually in the buffer, but we pretend
7162          * that they are.  When XLogInsert stores the whole buffer, the offset
7163          * arrays need not be stored too.  Note that even if all three arrays are
7164          * empty, we want to expose the buffer as a candidate for whole-page
7165          * storage, since this record type implies a defragmentation operation
7166          * even if no line pointers changed state.
7167          */
7168         if (nredirected > 0)
7169                 XLogRegisterBufData(0, (char *) redirected,
7170                                                         nredirected * sizeof(OffsetNumber) * 2);
7171
7172         if (ndead > 0)
7173                 XLogRegisterBufData(0, (char *) nowdead,
7174                                                         ndead * sizeof(OffsetNumber));
7175
7176         if (nunused > 0)
7177                 XLogRegisterBufData(0, (char *) nowunused,
7178                                                         nunused * sizeof(OffsetNumber));
7179
7180         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_CLEAN);
7181
7182         return recptr;
7183 }
7184
7185 /*
7186  * Perform XLogInsert for a heap-freeze operation.  Caller must have already
7187  * modified the buffer and marked it dirty.
7188  */
7189 XLogRecPtr
7190 log_heap_freeze(Relation reln, Buffer buffer, TransactionId cutoff_xid,
7191                                 xl_heap_freeze_tuple *tuples, int ntuples)
7192 {
7193         xl_heap_freeze_page xlrec;
7194         XLogRecPtr      recptr;
7195
7196         /* Caller should not call me on a non-WAL-logged relation */
7197         Assert(RelationNeedsWAL(reln));
7198         /* nor when there are no tuples to freeze */
7199         Assert(ntuples > 0);
7200
7201         xlrec.cutoff_xid = cutoff_xid;
7202         xlrec.ntuples = ntuples;
7203
7204         XLogBeginInsert();
7205         XLogRegisterData((char *) &xlrec, SizeOfHeapFreezePage);
7206
7207         /*
7208          * The freeze plan array is not actually in the buffer, but pretend that
7209          * it is.  When XLogInsert stores the whole buffer, the freeze plan need
7210          * not be stored too.
7211          */
7212         XLogRegisterBuffer(0, buffer, REGBUF_STANDARD);
7213         XLogRegisterBufData(0, (char *) tuples,
7214                                                 ntuples * sizeof(xl_heap_freeze_tuple));
7215
7216         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_FREEZE_PAGE);
7217
7218         return recptr;
7219 }
7220
7221 /*
7222  * Perform XLogInsert for a heap-visible operation.  'block' is the block
7223  * being marked all-visible, and vm_buffer is the buffer containing the
7224  * corresponding visibility map block.  Both should have already been modified
7225  * and dirtied.
7226  *
7227  * If checksums are enabled, we also generate a full-page image of
7228  * heap_buffer, if necessary.
7229  */
7230 XLogRecPtr
7231 log_heap_visible(RelFileNode rnode, Buffer heap_buffer, Buffer vm_buffer,
7232                                  TransactionId cutoff_xid, uint8 vmflags)
7233 {
7234         xl_heap_visible xlrec;
7235         XLogRecPtr      recptr;
7236         uint8           flags;
7237
7238         Assert(BufferIsValid(heap_buffer));
7239         Assert(BufferIsValid(vm_buffer));
7240
7241         xlrec.cutoff_xid = cutoff_xid;
7242         xlrec.flags = vmflags;
7243         XLogBeginInsert();
7244         XLogRegisterData((char *) &xlrec, SizeOfHeapVisible);
7245
7246         XLogRegisterBuffer(0, vm_buffer, 0);
7247
7248         flags = REGBUF_STANDARD;
7249         if (!XLogHintBitIsNeeded())
7250                 flags |= REGBUF_NO_IMAGE;
7251         XLogRegisterBuffer(1, heap_buffer, flags);
7252
7253         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_VISIBLE);
7254
7255         return recptr;
7256 }
7257
7258 /*
7259  * Perform XLogInsert for a heap-update operation.  Caller must already
7260  * have modified the buffer(s) and marked them dirty.
7261  */
7262 static XLogRecPtr
7263 log_heap_update(Relation reln, Buffer oldbuf,
7264                                 Buffer newbuf, HeapTuple oldtup, HeapTuple newtup,
7265                                 HeapTuple old_key_tuple,
7266                                 bool all_visible_cleared, bool new_all_visible_cleared)
7267 {
7268         xl_heap_update xlrec;
7269         xl_heap_header xlhdr;
7270         xl_heap_header xlhdr_idx;
7271         uint8           info;
7272         uint16          prefix_suffix[2];
7273         uint16          prefixlen = 0,
7274                                 suffixlen = 0;
7275         XLogRecPtr      recptr;
7276         Page            page = BufferGetPage(newbuf);
7277         bool            need_tuple_data = RelationIsLogicallyLogged(reln);
7278         bool            init;
7279         int                     bufflags;
7280
7281         /* Caller should not call me on a non-WAL-logged relation */
7282         Assert(RelationNeedsWAL(reln));
7283
7284         XLogBeginInsert();
7285
7286         if (HeapTupleIsHeapOnly(newtup))
7287                 info = XLOG_HEAP_HOT_UPDATE;
7288         else
7289                 info = XLOG_HEAP_UPDATE;
7290
7291         /*
7292          * If the old and new tuple are on the same page, we only need to log the
7293          * parts of the new tuple that were changed.  That saves on the amount of
7294          * WAL we need to write.  Currently, we just count any unchanged bytes in
7295          * the beginning and end of the tuple.  That's quick to check, and
7296          * perfectly covers the common case that only one field is updated.
7297          *
7298          * We could do this even if the old and new tuple are on different pages,
7299          * but only if we don't make a full-page image of the old page, which is
7300          * difficult to know in advance.  Also, if the old tuple is corrupt for
7301          * some reason, it would allow the corruption to propagate the new page,
7302          * so it seems best to avoid.  Under the general assumption that most
7303          * updates tend to create the new tuple version on the same page, there
7304          * isn't much to be gained by doing this across pages anyway.
7305          *
7306          * Skip this if we're taking a full-page image of the new page, as we
7307          * don't include the new tuple in the WAL record in that case.  Also
7308          * disable if wal_level='logical', as logical decoding needs to be able to
7309          * read the new tuple in whole from the WAL record alone.
7310          */
7311         if (oldbuf == newbuf && !need_tuple_data &&
7312                 !XLogCheckBufferNeedsBackup(newbuf))
7313         {
7314                 char       *oldp = (char *) oldtup->t_data + oldtup->t_data->t_hoff;
7315                 char       *newp = (char *) newtup->t_data + newtup->t_data->t_hoff;
7316                 int                     oldlen = oldtup->t_len - oldtup->t_data->t_hoff;
7317                 int                     newlen = newtup->t_len - newtup->t_data->t_hoff;
7318
7319                 /* Check for common prefix between old and new tuple */
7320                 for (prefixlen = 0; prefixlen < Min(oldlen, newlen); prefixlen++)
7321                 {
7322                         if (newp[prefixlen] != oldp[prefixlen])
7323                                 break;
7324                 }
7325
7326                 /*
7327                  * Storing the length of the prefix takes 2 bytes, so we need to save
7328                  * at least 3 bytes or there's no point.
7329                  */
7330                 if (prefixlen < 3)
7331                         prefixlen = 0;
7332
7333                 /* Same for suffix */
7334                 for (suffixlen = 0; suffixlen < Min(oldlen, newlen) - prefixlen; suffixlen++)
7335                 {
7336                         if (newp[newlen - suffixlen - 1] != oldp[oldlen - suffixlen - 1])
7337                                 break;
7338                 }
7339                 if (suffixlen < 3)
7340                         suffixlen = 0;
7341         }
7342
7343         /* Prepare main WAL data chain */
7344         xlrec.flags = 0;
7345         if (all_visible_cleared)
7346                 xlrec.flags |= XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED;
7347         if (new_all_visible_cleared)
7348                 xlrec.flags |= XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED;
7349         if (prefixlen > 0)
7350                 xlrec.flags |= XLH_UPDATE_PREFIX_FROM_OLD;
7351         if (suffixlen > 0)
7352                 xlrec.flags |= XLH_UPDATE_SUFFIX_FROM_OLD;
7353         if (need_tuple_data)
7354         {
7355                 xlrec.flags |= XLH_UPDATE_CONTAINS_NEW_TUPLE;
7356                 if (old_key_tuple)
7357                 {
7358                         if (reln->rd_rel->relreplident == REPLICA_IDENTITY_FULL)
7359                                 xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_TUPLE;
7360                         else
7361                                 xlrec.flags |= XLH_UPDATE_CONTAINS_OLD_KEY;
7362                 }
7363         }
7364
7365         /* If new tuple is the single and first tuple on page... */
7366         if (ItemPointerGetOffsetNumber(&(newtup->t_self)) == FirstOffsetNumber &&
7367                 PageGetMaxOffsetNumber(page) == FirstOffsetNumber)
7368         {
7369                 info |= XLOG_HEAP_INIT_PAGE;
7370                 init = true;
7371         }
7372         else
7373                 init = false;
7374
7375         /* Prepare WAL data for the old page */
7376         xlrec.old_offnum = ItemPointerGetOffsetNumber(&oldtup->t_self);
7377         xlrec.old_xmax = HeapTupleHeaderGetRawXmax(oldtup->t_data);
7378         xlrec.old_infobits_set = compute_infobits(oldtup->t_data->t_infomask,
7379                                                                                           oldtup->t_data->t_infomask2);
7380
7381         /* Prepare WAL data for the new page */
7382         xlrec.new_offnum = ItemPointerGetOffsetNumber(&newtup->t_self);
7383         xlrec.new_xmax = HeapTupleHeaderGetRawXmax(newtup->t_data);
7384
7385         bufflags = REGBUF_STANDARD;
7386         if (init)
7387                 bufflags |= REGBUF_WILL_INIT;
7388         if (need_tuple_data)
7389                 bufflags |= REGBUF_KEEP_DATA;
7390
7391         XLogRegisterBuffer(0, newbuf, bufflags);
7392         if (oldbuf != newbuf)
7393                 XLogRegisterBuffer(1, oldbuf, REGBUF_STANDARD);
7394
7395         XLogRegisterData((char *) &xlrec, SizeOfHeapUpdate);
7396
7397         /*
7398          * Prepare WAL data for the new tuple.
7399          */
7400         if (prefixlen > 0 || suffixlen > 0)
7401         {
7402                 if (prefixlen > 0 && suffixlen > 0)
7403                 {
7404                         prefix_suffix[0] = prefixlen;
7405                         prefix_suffix[1] = suffixlen;
7406                         XLogRegisterBufData(0, (char *) &prefix_suffix, sizeof(uint16) * 2);
7407                 }
7408                 else if (prefixlen > 0)
7409                 {
7410                         XLogRegisterBufData(0, (char *) &prefixlen, sizeof(uint16));
7411                 }
7412                 else
7413                 {
7414                         XLogRegisterBufData(0, (char *) &suffixlen, sizeof(uint16));
7415                 }
7416         }
7417
7418         xlhdr.t_infomask2 = newtup->t_data->t_infomask2;
7419         xlhdr.t_infomask = newtup->t_data->t_infomask;
7420         xlhdr.t_hoff = newtup->t_data->t_hoff;
7421         Assert(SizeofHeapTupleHeader + prefixlen + suffixlen <= newtup->t_len);
7422
7423         /*
7424          * PG73FORMAT: write bitmap [+ padding] [+ oid] + data
7425          *
7426          * The 'data' doesn't include the common prefix or suffix.
7427          */
7428         XLogRegisterBufData(0, (char *) &xlhdr, SizeOfHeapHeader);
7429         if (prefixlen == 0)
7430         {
7431                 XLogRegisterBufData(0,
7432                                                         ((char *) newtup->t_data) + SizeofHeapTupleHeader,
7433                                                         newtup->t_len - SizeofHeapTupleHeader - suffixlen);
7434         }
7435         else
7436         {
7437                 /*
7438                  * Have to write the null bitmap and data after the common prefix as
7439                  * two separate rdata entries.
7440                  */
7441                 /* bitmap [+ padding] [+ oid] */
7442                 if (newtup->t_data->t_hoff - SizeofHeapTupleHeader > 0)
7443                 {
7444                         XLogRegisterBufData(0,
7445                                                                 ((char *) newtup->t_data) + SizeofHeapTupleHeader,
7446                                                                 newtup->t_data->t_hoff - SizeofHeapTupleHeader);
7447                 }
7448
7449                 /* data after common prefix */
7450                 XLogRegisterBufData(0,
7451                                                         ((char *) newtup->t_data) + newtup->t_data->t_hoff + prefixlen,
7452                                                         newtup->t_len - newtup->t_data->t_hoff - prefixlen - suffixlen);
7453         }
7454
7455         /* We need to log a tuple identity */
7456         if (need_tuple_data && old_key_tuple)
7457         {
7458                 /* don't really need this, but its more comfy to decode */
7459                 xlhdr_idx.t_infomask2 = old_key_tuple->t_data->t_infomask2;
7460                 xlhdr_idx.t_infomask = old_key_tuple->t_data->t_infomask;
7461                 xlhdr_idx.t_hoff = old_key_tuple->t_data->t_hoff;
7462
7463                 XLogRegisterData((char *) &xlhdr_idx, SizeOfHeapHeader);
7464
7465                 /* PG73FORMAT: write bitmap [+ padding] [+ oid] + data */
7466                 XLogRegisterData((char *) old_key_tuple->t_data + SizeofHeapTupleHeader,
7467                                                  old_key_tuple->t_len - SizeofHeapTupleHeader);
7468         }
7469
7470         /* filtering by origin on a row level is much more efficient */
7471         XLogSetRecordFlags(XLOG_INCLUDE_ORIGIN);
7472
7473         recptr = XLogInsert(RM_HEAP_ID, info);
7474
7475         return recptr;
7476 }
7477
7478 /*
7479  * Perform XLogInsert of an XLOG_HEAP2_NEW_CID record
7480  *
7481  * This is only used in wal_level >= WAL_LEVEL_LOGICAL, and only for catalog
7482  * tuples.
7483  */
7484 static XLogRecPtr
7485 log_heap_new_cid(Relation relation, HeapTuple tup)
7486 {
7487         xl_heap_new_cid xlrec;
7488
7489         XLogRecPtr      recptr;
7490         HeapTupleHeader hdr = tup->t_data;
7491
7492         Assert(ItemPointerIsValid(&tup->t_self));
7493         Assert(tup->t_tableOid != InvalidOid);
7494
7495         xlrec.top_xid = GetTopTransactionId();
7496         xlrec.target_node = relation->rd_node;
7497         xlrec.target_tid = tup->t_self;
7498
7499         /*
7500          * If the tuple got inserted & deleted in the same TX we definitely have a
7501          * combocid, set cmin and cmax.
7502          */
7503         if (hdr->t_infomask & HEAP_COMBOCID)
7504         {
7505                 Assert(!(hdr->t_infomask & HEAP_XMAX_INVALID));
7506                 Assert(!HeapTupleHeaderXminInvalid(hdr));
7507                 xlrec.cmin = HeapTupleHeaderGetCmin(hdr);
7508                 xlrec.cmax = HeapTupleHeaderGetCmax(hdr);
7509                 xlrec.combocid = HeapTupleHeaderGetRawCommandId(hdr);
7510         }
7511         /* No combocid, so only cmin or cmax can be set by this TX */
7512         else
7513         {
7514                 /*
7515                  * Tuple inserted.
7516                  *
7517                  * We need to check for LOCK ONLY because multixacts might be
7518                  * transferred to the new tuple in case of FOR KEY SHARE updates in
7519                  * which case there will be an xmax, although the tuple just got
7520                  * inserted.
7521                  */
7522                 if (hdr->t_infomask & HEAP_XMAX_INVALID ||
7523                         HEAP_XMAX_IS_LOCKED_ONLY(hdr->t_infomask))
7524                 {
7525                         xlrec.cmin = HeapTupleHeaderGetRawCommandId(hdr);
7526                         xlrec.cmax = InvalidCommandId;
7527                 }
7528                 /* Tuple from a different tx updated or deleted. */
7529                 else
7530                 {
7531                         xlrec.cmin = InvalidCommandId;
7532                         xlrec.cmax = HeapTupleHeaderGetRawCommandId(hdr);
7533
7534                 }
7535                 xlrec.combocid = InvalidCommandId;
7536         }
7537
7538         /*
7539          * Note that we don't need to register the buffer here, because this
7540          * operation does not modify the page. The insert/update/delete that
7541          * called us certainly did, but that's WAL-logged separately.
7542          */
7543         XLogBeginInsert();
7544         XLogRegisterData((char *) &xlrec, SizeOfHeapNewCid);
7545
7546         /* will be looked at irrespective of origin */
7547
7548         recptr = XLogInsert(RM_HEAP2_ID, XLOG_HEAP2_NEW_CID);
7549
7550         return recptr;
7551 }
7552
7553 /*
7554  * Build a heap tuple representing the configured REPLICA IDENTITY to represent
7555  * the old tuple in a UPDATE or DELETE.
7556  *
7557  * Returns NULL if there's no need to log an identity or if there's no suitable
7558  * key in the Relation relation.
7559  */
7560 static HeapTuple
7561 ExtractReplicaIdentity(Relation relation, HeapTuple tp, bool key_changed, bool *copy)
7562 {
7563         TupleDesc       desc = RelationGetDescr(relation);
7564         Oid                     replidindex;
7565         Relation        idx_rel;
7566         char            replident = relation->rd_rel->relreplident;
7567         HeapTuple       key_tuple = NULL;
7568         bool            nulls[MaxHeapAttributeNumber];
7569         Datum           values[MaxHeapAttributeNumber];
7570         int                     natt;
7571
7572         *copy = false;
7573
7574         if (!RelationIsLogicallyLogged(relation))
7575                 return NULL;
7576
7577         if (replident == REPLICA_IDENTITY_NOTHING)
7578                 return NULL;
7579
7580         if (replident == REPLICA_IDENTITY_FULL)
7581         {
7582                 /*
7583                  * When logging the entire old tuple, it very well could contain
7584                  * toasted columns. If so, force them to be inlined.
7585                  */
7586                 if (HeapTupleHasExternal(tp))
7587                 {
7588                         *copy = true;
7589                         tp = toast_flatten_tuple(tp, RelationGetDescr(relation));
7590                 }
7591                 return tp;
7592         }
7593
7594         /* if the key hasn't changed and we're only logging the key, we're done */
7595         if (!key_changed)
7596                 return NULL;
7597
7598         /* find the replica identity index */
7599         replidindex = RelationGetReplicaIndex(relation);
7600         if (!OidIsValid(replidindex))
7601         {
7602                 elog(DEBUG4, "could not find configured replica identity for table \"%s\"",
7603                          RelationGetRelationName(relation));
7604                 return NULL;
7605         }
7606
7607         idx_rel = RelationIdGetRelation(replidindex);
7608
7609         Assert(CheckRelationLockedByMe(idx_rel, AccessShareLock, true));
7610
7611         /* deform tuple, so we have fast access to columns */
7612         heap_deform_tuple(tp, desc, values, nulls);
7613
7614         /* set all columns to NULL, regardless of whether they actually are */
7615         memset(nulls, 1, sizeof(nulls));
7616
7617         /*
7618          * Now set all columns contained in the index to NOT NULL, they cannot
7619          * currently be NULL.
7620          */
7621         for (natt = 0; natt < IndexRelationGetNumberOfKeyAttributes(idx_rel); natt++)
7622         {
7623                 int                     attno = idx_rel->rd_index->indkey.values[natt];
7624
7625                 if (attno < 0)
7626                         elog(ERROR, "system column in index");
7627                 nulls[attno - 1] = false;
7628         }
7629
7630         key_tuple = heap_form_tuple(desc, values, nulls);
7631         *copy = true;
7632         RelationClose(idx_rel);
7633
7634         /*
7635          * If the tuple, which by here only contains indexed columns, still has
7636          * toasted columns, force them to be inlined. This is somewhat unlikely
7637          * since there's limits on the size of indexed columns, so we don't
7638          * duplicate toast_flatten_tuple()s functionality in the above loop over
7639          * the indexed columns, even if it would be more efficient.
7640          */
7641         if (HeapTupleHasExternal(key_tuple))
7642         {
7643                 HeapTuple       oldtup = key_tuple;
7644
7645                 key_tuple = toast_flatten_tuple(oldtup, RelationGetDescr(relation));
7646                 heap_freetuple(oldtup);
7647         }
7648
7649         return key_tuple;
7650 }
7651
7652 /*
7653  * Handles CLEANUP_INFO
7654  */
7655 static void
7656 heap_xlog_cleanup_info(XLogReaderState *record)
7657 {
7658         xl_heap_cleanup_info *xlrec = (xl_heap_cleanup_info *) XLogRecGetData(record);
7659
7660         if (InHotStandby)
7661                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, xlrec->node);
7662
7663         /*
7664          * Actual operation is a no-op. Record type exists to provide a means for
7665          * conflict processing to occur before we begin index vacuum actions. see
7666          * vacuumlazy.c and also comments in btvacuumpage()
7667          */
7668
7669         /* Backup blocks are not used in cleanup_info records */
7670         Assert(!XLogRecHasAnyBlockRefs(record));
7671 }
7672
7673 /*
7674  * Handles XLOG_HEAP2_CLEAN record type
7675  */
7676 static void
7677 heap_xlog_clean(XLogReaderState *record)
7678 {
7679         XLogRecPtr      lsn = record->EndRecPtr;
7680         xl_heap_clean *xlrec = (xl_heap_clean *) XLogRecGetData(record);
7681         Buffer          buffer;
7682         RelFileNode rnode;
7683         BlockNumber blkno;
7684         XLogRedoAction action;
7685
7686         XLogRecGetBlockTag(record, 0, &rnode, NULL, &blkno);
7687
7688         /*
7689          * We're about to remove tuples. In Hot Standby mode, ensure that there's
7690          * no queries running for which the removed tuples are still visible.
7691          *
7692          * Not all HEAP2_CLEAN records remove tuples with xids, so we only want to
7693          * conflict on the records that cause MVCC failures for user queries. If
7694          * latestRemovedXid is invalid, skip conflict processing.
7695          */
7696         if (InHotStandby && TransactionIdIsValid(xlrec->latestRemovedXid))
7697                 ResolveRecoveryConflictWithSnapshot(xlrec->latestRemovedXid, rnode);
7698
7699         /*
7700          * If we have a full-page image, restore it (using a cleanup lock) and
7701          * we're done.
7702          */
7703         action = XLogReadBufferForRedoExtended(record, 0, RBM_NORMAL, true,
7704                                                                                    &buffer);
7705         if (action == BLK_NEEDS_REDO)
7706         {
7707                 Page            page = (Page) BufferGetPage(buffer);
7708                 OffsetNumber *end;
7709                 OffsetNumber *redirected;
7710                 OffsetNumber *nowdead;
7711                 OffsetNumber *nowunused;
7712                 int                     nredirected;
7713                 int                     ndead;
7714                 int                     nunused;
7715                 Size            datalen;
7716
7717                 redirected = (OffsetNumber *) XLogRecGetBlockData(record, 0, &datalen);
7718
7719                 nredirected = xlrec->nredirected;
7720                 ndead = xlrec->ndead;
7721                 end = (OffsetNumber *) ((char *) redirected + datalen);
7722                 nowdead = redirected + (nredirected * 2);
7723                 nowunused = nowdead + ndead;
7724                 nunused = (end - nowunused);
7725                 Assert(nunused >= 0);
7726
7727                 /* Update all line pointers per the record, and repair fragmentation */
7728                 heap_page_prune_execute(buffer,
7729                                                                 redirected, nredirected,
7730                                                                 nowdead, ndead,
7731                                                                 nowunused, nunused);
7732
7733                 /*
7734                  * Note: we don't worry about updating the page's prunability hints.
7735                  * At worst this will cause an extra prune cycle to occur soon.
7736                  */
7737
7738                 PageSetLSN(page, lsn);
7739                 MarkBufferDirty(buffer);
7740         }
7741
7742         if (BufferIsValid(buffer))
7743         {
7744                 Size            freespace = PageGetHeapFreeSpace(BufferGetPage(buffer));
7745
7746                 UnlockReleaseBuffer(buffer);
7747
7748                 /*
7749                  * After cleaning records from a page, it's useful to update the FSM
7750                  * about it, as it may cause the page become target for insertions
7751                  * later even if vacuum decides not to visit it (which is possible if
7752                  * gets marked all-visible.)
7753                  *
7754                  * Do this regardless of a full-page image being applied, since the
7755                  * FSM data is not in the page anyway.
7756                  */
7757                 XLogRecordPageWithFreeSpace(rnode, blkno, freespace);
7758         }
7759 }
7760
7761 /*
7762  * Replay XLOG_HEAP2_VISIBLE record.
7763  *
7764  * The critical integrity requirement here is that we must never end up with
7765  * a situation where the visibility map bit is set, and the page-level
7766  * PD_ALL_VISIBLE bit is clear.  If that were to occur, then a subsequent
7767  * page modification would fail to clear the visibility map bit.
7768  */
7769 static void
7770 heap_xlog_visible(XLogReaderState *record)
7771 {
7772         XLogRecPtr      lsn = record->EndRecPtr;
7773         xl_heap_visible *xlrec = (xl_heap_visible *) XLogRecGetData(record);
7774         Buffer          vmbuffer = InvalidBuffer;
7775         Buffer          buffer;
7776         Page            page;
7777         RelFileNode rnode;
7778         BlockNumber blkno;
7779         XLogRedoAction action;
7780
7781         XLogRecGetBlockTag(record, 1, &rnode, NULL, &blkno);
7782
7783         /*
7784          * If there are any Hot Standby transactions running that have an xmin
7785          * horizon old enough that this page isn't all-visible for them, they
7786          * might incorrectly decide that an index-only scan can skip a heap fetch.
7787          *
7788          * NB: It might be better to throw some kind of "soft" conflict here that
7789          * forces any index-only scan that is in flight to perform heap fetches,
7790          * rather than killing the transaction outright.
7791          */
7792         if (InHotStandby)
7793                 ResolveRecoveryConflictWithSnapshot(xlrec->cutoff_xid, rnode);
7794
7795         /*
7796          * Read the heap page, if it still exists. If the heap file has dropped or
7797          * truncated later in recovery, we don't need to update the page, but we'd
7798          * better still update the visibility map.
7799          */
7800         action = XLogReadBufferForRedo(record, 1, &buffer);
7801         if (action == BLK_NEEDS_REDO)
7802         {
7803                 /*
7804                  * We don't bump the LSN of the heap page when setting the visibility
7805                  * map bit (unless checksums or wal_hint_bits is enabled, in which
7806                  * case we must), because that would generate an unworkable volume of
7807                  * full-page writes.  This exposes us to torn page hazards, but since
7808                  * we're not inspecting the existing page contents in any way, we
7809                  * don't care.
7810                  *
7811                  * However, all operations that clear the visibility map bit *do* bump
7812                  * the LSN, and those operations will only be replayed if the XLOG LSN
7813                  * follows the page LSN.  Thus, if the page LSN has advanced past our
7814                  * XLOG record's LSN, we mustn't mark the page all-visible, because
7815                  * the subsequent update won't be replayed to clear the flag.
7816                  */
7817                 page = BufferGetPage(buffer);
7818
7819                 PageSetAllVisible(page);
7820
7821                 MarkBufferDirty(buffer);
7822         }
7823         else if (action == BLK_RESTORED)
7824         {
7825                 /*
7826                  * If heap block was backed up, we already restored it and there's
7827                  * nothing more to do. (This can only happen with checksums or
7828                  * wal_log_hints enabled.)
7829                  */
7830         }
7831
7832         if (BufferIsValid(buffer))
7833         {
7834                 Size            space = PageGetFreeSpace(BufferGetPage(buffer));
7835
7836                 UnlockReleaseBuffer(buffer);
7837
7838                 /*
7839                  * Since FSM is not WAL-logged and only updated heuristically, it
7840                  * easily becomes stale in standbys.  If the standby is later promoted
7841                  * and runs VACUUM, it will skip updating individual free space
7842                  * figures for pages that became all-visible (or all-frozen, depending
7843                  * on the vacuum mode,) which is troublesome when FreeSpaceMapVacuum
7844                  * propagates too optimistic free space values to upper FSM layers;
7845                  * later inserters try to use such pages only to find out that they
7846                  * are unusable.  This can cause long stalls when there are many such
7847                  * pages.
7848                  *
7849                  * Forestall those problems by updating FSM's idea about a page that
7850                  * is becoming all-visible or all-frozen.
7851                  *
7852                  * Do this regardless of a full-page image being applied, since the
7853                  * FSM data is not in the page anyway.
7854                  */
7855                 if (xlrec->flags & VISIBILITYMAP_VALID_BITS)
7856                         XLogRecordPageWithFreeSpace(rnode, blkno, space);
7857         }
7858
7859         /*
7860          * Even if we skipped the heap page update due to the LSN interlock, it's
7861          * still safe to update the visibility map.  Any WAL record that clears
7862          * the visibility map bit does so before checking the page LSN, so any
7863          * bits that need to be cleared will still be cleared.
7864          */
7865         if (XLogReadBufferForRedoExtended(record, 0, RBM_ZERO_ON_ERROR, false,
7866                                                                           &vmbuffer) == BLK_NEEDS_REDO)
7867         {
7868                 Page            vmpage = BufferGetPage(vmbuffer);
7869                 Relation        reln;
7870
7871                 /* initialize the page if it was read as zeros */
7872                 if (PageIsNew(vmpage))
7873                         PageInit(vmpage, BLCKSZ, 0);
7874
7875                 /*
7876                  * XLogReadBufferForRedoExtended locked the buffer. But
7877                  * visibilitymap_set will handle locking itself.
7878                  */
7879                 LockBuffer(vmbuffer, BUFFER_LOCK_UNLOCK);
7880
7881                 reln = CreateFakeRelcacheEntry(rnode);
7882                 visibilitymap_pin(reln, blkno, &vmbuffer);
7883
7884                 /*
7885                  * Don't set the bit if replay has already passed this point.
7886                  *
7887                  * It might be safe to do this unconditionally; if replay has passed
7888                  * this point, we'll replay at least as far this time as we did
7889                  * before, and if this bit needs to be cleared, the record responsible
7890                  * for doing so should be again replayed, and clear it.  For right
7891                  * now, out of an abundance of conservatism, we use the same test here
7892                  * we did for the heap page.  If this results in a dropped bit, no
7893                  * real harm is done; and the next VACUUM will fix it.
7894                  */
7895                 if (lsn > PageGetLSN(vmpage))
7896                         visibilitymap_set(reln, blkno, InvalidBuffer, lsn, vmbuffer,
7897                                                           xlrec->cutoff_xid, xlrec->flags);
7898
7899                 ReleaseBuffer(vmbuffer);
7900                 FreeFakeRelcacheEntry(reln);
7901         }
7902         else if (BufferIsValid(vmbuffer))
7903                 UnlockReleaseBuffer(vmbuffer);
7904 }
7905
7906 /*
7907  * Replay XLOG_HEAP2_FREEZE_PAGE records
7908  */
7909 static void
7910 heap_xlog_freeze_page(XLogReaderState *record)
7911 {
7912         XLogRecPtr      lsn = record->EndRecPtr;
7913         xl_heap_freeze_page *xlrec = (xl_heap_freeze_page *) XLogRecGetData(record);
7914         TransactionId cutoff_xid = xlrec->cutoff_xid;
7915         Buffer          buffer;
7916         int                     ntup;
7917
7918         /*
7919          * In Hot Standby mode, ensure that there's no queries running which still
7920          * consider the frozen xids as running.
7921          */
7922         if (InHotStandby)
7923         {
7924                 RelFileNode rnode;
7925                 TransactionId latestRemovedXid = cutoff_xid;
7926
7927                 TransactionIdRetreat(latestRemovedXid);
7928
7929                 XLogRecGetBlockTag(record, 0, &rnode, NULL, NULL);
7930                 ResolveRecoveryConflictWithSnapshot(latestRemovedXid, rnode);
7931         }
7932
7933         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
7934         {
7935                 Page            page = BufferGetPage(buffer);
7936                 xl_heap_freeze_tuple *tuples;
7937
7938                 tuples = (xl_heap_freeze_tuple *) XLogRecGetBlockData(record, 0, NULL);
7939
7940                 /* now execute freeze plan for each frozen tuple */
7941                 for (ntup = 0; ntup < xlrec->ntuples; ntup++)
7942                 {
7943                         xl_heap_freeze_tuple *xlrec_tp;
7944                         ItemId          lp;
7945                         HeapTupleHeader tuple;
7946
7947                         xlrec_tp = &tuples[ntup];
7948                         lp = PageGetItemId(page, xlrec_tp->offset); /* offsets are one-based */
7949                         tuple = (HeapTupleHeader) PageGetItem(page, lp);
7950
7951                         heap_execute_freeze_tuple(tuple, xlrec_tp);
7952                 }
7953
7954                 PageSetLSN(page, lsn);
7955                 MarkBufferDirty(buffer);
7956         }
7957         if (BufferIsValid(buffer))
7958                 UnlockReleaseBuffer(buffer);
7959 }
7960
7961 /*
7962  * Given an "infobits" field from an XLog record, set the correct bits in the
7963  * given infomask and infomask2 for the tuple touched by the record.
7964  *
7965  * (This is the reverse of compute_infobits).
7966  */
7967 static void
7968 fix_infomask_from_infobits(uint8 infobits, uint16 *infomask, uint16 *infomask2)
7969 {
7970         *infomask &= ~(HEAP_XMAX_IS_MULTI | HEAP_XMAX_LOCK_ONLY |
7971                                    HEAP_XMAX_KEYSHR_LOCK | HEAP_XMAX_EXCL_LOCK);
7972         *infomask2 &= ~HEAP_KEYS_UPDATED;
7973
7974         if (infobits & XLHL_XMAX_IS_MULTI)
7975                 *infomask |= HEAP_XMAX_IS_MULTI;
7976         if (infobits & XLHL_XMAX_LOCK_ONLY)
7977                 *infomask |= HEAP_XMAX_LOCK_ONLY;
7978         if (infobits & XLHL_XMAX_EXCL_LOCK)
7979                 *infomask |= HEAP_XMAX_EXCL_LOCK;
7980         /* note HEAP_XMAX_SHR_LOCK isn't considered here */
7981         if (infobits & XLHL_XMAX_KEYSHR_LOCK)
7982                 *infomask |= HEAP_XMAX_KEYSHR_LOCK;
7983
7984         if (infobits & XLHL_KEYS_UPDATED)
7985                 *infomask2 |= HEAP_KEYS_UPDATED;
7986 }
7987
7988 static void
7989 heap_xlog_delete(XLogReaderState *record)
7990 {
7991         XLogRecPtr      lsn = record->EndRecPtr;
7992         xl_heap_delete *xlrec = (xl_heap_delete *) XLogRecGetData(record);
7993         Buffer          buffer;
7994         Page            page;
7995         ItemId          lp = NULL;
7996         HeapTupleHeader htup;
7997         BlockNumber blkno;
7998         RelFileNode target_node;
7999         ItemPointerData target_tid;
8000
8001         XLogRecGetBlockTag(record, 0, &target_node, NULL, &blkno);
8002         ItemPointerSetBlockNumber(&target_tid, blkno);
8003         ItemPointerSetOffsetNumber(&target_tid, xlrec->offnum);
8004
8005         /*
8006          * The visibility map may need to be fixed even if the heap page is
8007          * already up-to-date.
8008          */
8009         if (xlrec->flags & XLH_DELETE_ALL_VISIBLE_CLEARED)
8010         {
8011                 Relation        reln = CreateFakeRelcacheEntry(target_node);
8012                 Buffer          vmbuffer = InvalidBuffer;
8013
8014                 visibilitymap_pin(reln, blkno, &vmbuffer);
8015                 visibilitymap_clear(reln, blkno, vmbuffer, VISIBILITYMAP_VALID_BITS);
8016                 ReleaseBuffer(vmbuffer);
8017                 FreeFakeRelcacheEntry(reln);
8018         }
8019
8020         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8021         {
8022                 page = BufferGetPage(buffer);
8023
8024                 if (PageGetMaxOffsetNumber(page) >= xlrec->offnum)
8025                         lp = PageGetItemId(page, xlrec->offnum);
8026
8027                 if (PageGetMaxOffsetNumber(page) < xlrec->offnum || !ItemIdIsNormal(lp))
8028                         elog(PANIC, "invalid lp");
8029
8030                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8031
8032                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
8033                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
8034                 HeapTupleHeaderClearHotUpdated(htup);
8035                 fix_infomask_from_infobits(xlrec->infobits_set,
8036                                                                    &htup->t_infomask, &htup->t_infomask2);
8037                 if (!(xlrec->flags & XLH_DELETE_IS_SUPER))
8038                         HeapTupleHeaderSetXmax(htup, xlrec->xmax);
8039                 else
8040                         HeapTupleHeaderSetXmin(htup, InvalidTransactionId);
8041                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
8042
8043                 /* Mark the page as a candidate for pruning */
8044                 PageSetPrunable(page, XLogRecGetXid(record));
8045
8046                 if (xlrec->flags & XLH_DELETE_ALL_VISIBLE_CLEARED)
8047                         PageClearAllVisible(page);
8048
8049                 /* Make sure t_ctid is set correctly */
8050                 if (xlrec->flags & XLH_DELETE_IS_PARTITION_MOVE)
8051                         HeapTupleHeaderSetMovedPartitions(htup);
8052                 else
8053                         htup->t_ctid = target_tid;
8054                 PageSetLSN(page, lsn);
8055                 MarkBufferDirty(buffer);
8056         }
8057         if (BufferIsValid(buffer))
8058                 UnlockReleaseBuffer(buffer);
8059 }
8060
8061 static void
8062 heap_xlog_insert(XLogReaderState *record)
8063 {
8064         XLogRecPtr      lsn = record->EndRecPtr;
8065         xl_heap_insert *xlrec = (xl_heap_insert *) XLogRecGetData(record);
8066         Buffer          buffer;
8067         Page            page;
8068         union
8069         {
8070                 HeapTupleHeaderData hdr;
8071                 char            data[MaxHeapTupleSize];
8072         }                       tbuf;
8073         HeapTupleHeader htup;
8074         xl_heap_header xlhdr;
8075         uint32          newlen;
8076         Size            freespace = 0;
8077         RelFileNode target_node;
8078         BlockNumber blkno;
8079         ItemPointerData target_tid;
8080         XLogRedoAction action;
8081
8082         XLogRecGetBlockTag(record, 0, &target_node, NULL, &blkno);
8083         ItemPointerSetBlockNumber(&target_tid, blkno);
8084         ItemPointerSetOffsetNumber(&target_tid, xlrec->offnum);
8085
8086         /*
8087          * The visibility map may need to be fixed even if the heap page is
8088          * already up-to-date.
8089          */
8090         if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8091         {
8092                 Relation        reln = CreateFakeRelcacheEntry(target_node);
8093                 Buffer          vmbuffer = InvalidBuffer;
8094
8095                 visibilitymap_pin(reln, blkno, &vmbuffer);
8096                 visibilitymap_clear(reln, blkno, vmbuffer, VISIBILITYMAP_VALID_BITS);
8097                 ReleaseBuffer(vmbuffer);
8098                 FreeFakeRelcacheEntry(reln);
8099         }
8100
8101         /*
8102          * If we inserted the first and only tuple on the page, re-initialize the
8103          * page from scratch.
8104          */
8105         if (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE)
8106         {
8107                 buffer = XLogInitBufferForRedo(record, 0);
8108                 page = BufferGetPage(buffer);
8109                 PageInit(page, BufferGetPageSize(buffer), 0);
8110                 action = BLK_NEEDS_REDO;
8111         }
8112         else
8113                 action = XLogReadBufferForRedo(record, 0, &buffer);
8114         if (action == BLK_NEEDS_REDO)
8115         {
8116                 Size            datalen;
8117                 char       *data;
8118
8119                 page = BufferGetPage(buffer);
8120
8121                 if (PageGetMaxOffsetNumber(page) + 1 < xlrec->offnum)
8122                         elog(PANIC, "invalid max offset number");
8123
8124                 data = XLogRecGetBlockData(record, 0, &datalen);
8125
8126                 newlen = datalen - SizeOfHeapHeader;
8127                 Assert(datalen > SizeOfHeapHeader && newlen <= MaxHeapTupleSize);
8128                 memcpy((char *) &xlhdr, data, SizeOfHeapHeader);
8129                 data += SizeOfHeapHeader;
8130
8131                 htup = &tbuf.hdr;
8132                 MemSet((char *) htup, 0, SizeofHeapTupleHeader);
8133                 /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
8134                 memcpy((char *) htup + SizeofHeapTupleHeader,
8135                            data,
8136                            newlen);
8137                 newlen += SizeofHeapTupleHeader;
8138                 htup->t_infomask2 = xlhdr.t_infomask2;
8139                 htup->t_infomask = xlhdr.t_infomask;
8140                 htup->t_hoff = xlhdr.t_hoff;
8141                 HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
8142                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
8143                 htup->t_ctid = target_tid;
8144
8145                 if (PageAddItem(page, (Item) htup, newlen, xlrec->offnum,
8146                                                 true, true) == InvalidOffsetNumber)
8147                         elog(PANIC, "failed to add tuple");
8148
8149                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
8150
8151                 PageSetLSN(page, lsn);
8152
8153                 if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8154                         PageClearAllVisible(page);
8155
8156                 MarkBufferDirty(buffer);
8157         }
8158         if (BufferIsValid(buffer))
8159                 UnlockReleaseBuffer(buffer);
8160
8161         /*
8162          * If the page is running low on free space, update the FSM as well.
8163          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
8164          * better than that without knowing the fill-factor for the table.
8165          *
8166          * XXX: Don't do this if the page was restored from full page image. We
8167          * don't bother to update the FSM in that case, it doesn't need to be
8168          * totally accurate anyway.
8169          */
8170         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
8171                 XLogRecordPageWithFreeSpace(target_node, blkno, freespace);
8172 }
8173
8174 /*
8175  * Handles MULTI_INSERT record type.
8176  */
8177 static void
8178 heap_xlog_multi_insert(XLogReaderState *record)
8179 {
8180         XLogRecPtr      lsn = record->EndRecPtr;
8181         xl_heap_multi_insert *xlrec;
8182         RelFileNode rnode;
8183         BlockNumber blkno;
8184         Buffer          buffer;
8185         Page            page;
8186         union
8187         {
8188                 HeapTupleHeaderData hdr;
8189                 char            data[MaxHeapTupleSize];
8190         }                       tbuf;
8191         HeapTupleHeader htup;
8192         uint32          newlen;
8193         Size            freespace = 0;
8194         int                     i;
8195         bool            isinit = (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE) != 0;
8196         XLogRedoAction action;
8197
8198         /*
8199          * Insertion doesn't overwrite MVCC data, so no conflict processing is
8200          * required.
8201          */
8202         xlrec = (xl_heap_multi_insert *) XLogRecGetData(record);
8203
8204         XLogRecGetBlockTag(record, 0, &rnode, NULL, &blkno);
8205
8206         /*
8207          * The visibility map may need to be fixed even if the heap page is
8208          * already up-to-date.
8209          */
8210         if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8211         {
8212                 Relation        reln = CreateFakeRelcacheEntry(rnode);
8213                 Buffer          vmbuffer = InvalidBuffer;
8214
8215                 visibilitymap_pin(reln, blkno, &vmbuffer);
8216                 visibilitymap_clear(reln, blkno, vmbuffer, VISIBILITYMAP_VALID_BITS);
8217                 ReleaseBuffer(vmbuffer);
8218                 FreeFakeRelcacheEntry(reln);
8219         }
8220
8221         if (isinit)
8222         {
8223                 buffer = XLogInitBufferForRedo(record, 0);
8224                 page = BufferGetPage(buffer);
8225                 PageInit(page, BufferGetPageSize(buffer), 0);
8226                 action = BLK_NEEDS_REDO;
8227         }
8228         else
8229                 action = XLogReadBufferForRedo(record, 0, &buffer);
8230         if (action == BLK_NEEDS_REDO)
8231         {
8232                 char       *tupdata;
8233                 char       *endptr;
8234                 Size            len;
8235
8236                 /* Tuples are stored as block data */
8237                 tupdata = XLogRecGetBlockData(record, 0, &len);
8238                 endptr = tupdata + len;
8239
8240                 page = (Page) BufferGetPage(buffer);
8241
8242                 for (i = 0; i < xlrec->ntuples; i++)
8243                 {
8244                         OffsetNumber offnum;
8245                         xl_multi_insert_tuple *xlhdr;
8246
8247                         /*
8248                          * If we're reinitializing the page, the tuples are stored in
8249                          * order from FirstOffsetNumber. Otherwise there's an array of
8250                          * offsets in the WAL record, and the tuples come after that.
8251                          */
8252                         if (isinit)
8253                                 offnum = FirstOffsetNumber + i;
8254                         else
8255                                 offnum = xlrec->offsets[i];
8256                         if (PageGetMaxOffsetNumber(page) + 1 < offnum)
8257                                 elog(PANIC, "invalid max offset number");
8258
8259                         xlhdr = (xl_multi_insert_tuple *) SHORTALIGN(tupdata);
8260                         tupdata = ((char *) xlhdr) + SizeOfMultiInsertTuple;
8261
8262                         newlen = xlhdr->datalen;
8263                         Assert(newlen <= MaxHeapTupleSize);
8264                         htup = &tbuf.hdr;
8265                         MemSet((char *) htup, 0, SizeofHeapTupleHeader);
8266                         /* PG73FORMAT: get bitmap [+ padding] [+ oid] + data */
8267                         memcpy((char *) htup + SizeofHeapTupleHeader,
8268                                    (char *) tupdata,
8269                                    newlen);
8270                         tupdata += newlen;
8271
8272                         newlen += SizeofHeapTupleHeader;
8273                         htup->t_infomask2 = xlhdr->t_infomask2;
8274                         htup->t_infomask = xlhdr->t_infomask;
8275                         htup->t_hoff = xlhdr->t_hoff;
8276                         HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
8277                         HeapTupleHeaderSetCmin(htup, FirstCommandId);
8278                         ItemPointerSetBlockNumber(&htup->t_ctid, blkno);
8279                         ItemPointerSetOffsetNumber(&htup->t_ctid, offnum);
8280
8281                         offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
8282                         if (offnum == InvalidOffsetNumber)
8283                                 elog(PANIC, "failed to add tuple");
8284                 }
8285                 if (tupdata != endptr)
8286                         elog(PANIC, "total tuple length mismatch");
8287
8288                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
8289
8290                 PageSetLSN(page, lsn);
8291
8292                 if (xlrec->flags & XLH_INSERT_ALL_VISIBLE_CLEARED)
8293                         PageClearAllVisible(page);
8294
8295                 MarkBufferDirty(buffer);
8296         }
8297         if (BufferIsValid(buffer))
8298                 UnlockReleaseBuffer(buffer);
8299
8300         /*
8301          * If the page is running low on free space, update the FSM as well.
8302          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
8303          * better than that without knowing the fill-factor for the table.
8304          *
8305          * XXX: Don't do this if the page was restored from full page image. We
8306          * don't bother to update the FSM in that case, it doesn't need to be
8307          * totally accurate anyway.
8308          */
8309         if (action == BLK_NEEDS_REDO && freespace < BLCKSZ / 5)
8310                 XLogRecordPageWithFreeSpace(rnode, blkno, freespace);
8311 }
8312
8313 /*
8314  * Handles UPDATE and HOT_UPDATE
8315  */
8316 static void
8317 heap_xlog_update(XLogReaderState *record, bool hot_update)
8318 {
8319         XLogRecPtr      lsn = record->EndRecPtr;
8320         xl_heap_update *xlrec = (xl_heap_update *) XLogRecGetData(record);
8321         RelFileNode rnode;
8322         BlockNumber oldblk;
8323         BlockNumber newblk;
8324         ItemPointerData newtid;
8325         Buffer          obuffer,
8326                                 nbuffer;
8327         Page            page;
8328         OffsetNumber offnum;
8329         ItemId          lp = NULL;
8330         HeapTupleData oldtup;
8331         HeapTupleHeader htup;
8332         uint16          prefixlen = 0,
8333                                 suffixlen = 0;
8334         char       *newp;
8335         union
8336         {
8337                 HeapTupleHeaderData hdr;
8338                 char            data[MaxHeapTupleSize];
8339         }                       tbuf;
8340         xl_heap_header xlhdr;
8341         uint32          newlen;
8342         Size            freespace = 0;
8343         XLogRedoAction oldaction;
8344         XLogRedoAction newaction;
8345
8346         /* initialize to keep the compiler quiet */
8347         oldtup.t_data = NULL;
8348         oldtup.t_len = 0;
8349
8350         XLogRecGetBlockTag(record, 0, &rnode, NULL, &newblk);
8351         if (XLogRecGetBlockTag(record, 1, NULL, NULL, &oldblk))
8352         {
8353                 /* HOT updates are never done across pages */
8354                 Assert(!hot_update);
8355         }
8356         else
8357                 oldblk = newblk;
8358
8359         ItemPointerSet(&newtid, newblk, xlrec->new_offnum);
8360
8361         /*
8362          * The visibility map may need to be fixed even if the heap page is
8363          * already up-to-date.
8364          */
8365         if (xlrec->flags & XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED)
8366         {
8367                 Relation        reln = CreateFakeRelcacheEntry(rnode);
8368                 Buffer          vmbuffer = InvalidBuffer;
8369
8370                 visibilitymap_pin(reln, oldblk, &vmbuffer);
8371                 visibilitymap_clear(reln, oldblk, vmbuffer, VISIBILITYMAP_VALID_BITS);
8372                 ReleaseBuffer(vmbuffer);
8373                 FreeFakeRelcacheEntry(reln);
8374         }
8375
8376         /*
8377          * In normal operation, it is important to lock the two pages in
8378          * page-number order, to avoid possible deadlocks against other update
8379          * operations going the other way.  However, during WAL replay there can
8380          * be no other update happening, so we don't need to worry about that. But
8381          * we *do* need to worry that we don't expose an inconsistent state to Hot
8382          * Standby queries --- so the original page can't be unlocked before we've
8383          * added the new tuple to the new page.
8384          */
8385
8386         /* Deal with old tuple version */
8387         oldaction = XLogReadBufferForRedo(record, (oldblk == newblk) ? 0 : 1,
8388                                                                           &obuffer);
8389         if (oldaction == BLK_NEEDS_REDO)
8390         {
8391                 page = BufferGetPage(obuffer);
8392                 offnum = xlrec->old_offnum;
8393                 if (PageGetMaxOffsetNumber(page) >= offnum)
8394                         lp = PageGetItemId(page, offnum);
8395
8396                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8397                         elog(PANIC, "invalid lp");
8398
8399                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8400
8401                 oldtup.t_data = htup;
8402                 oldtup.t_len = ItemIdGetLength(lp);
8403
8404                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
8405                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
8406                 if (hot_update)
8407                         HeapTupleHeaderSetHotUpdated(htup);
8408                 else
8409                         HeapTupleHeaderClearHotUpdated(htup);
8410                 fix_infomask_from_infobits(xlrec->old_infobits_set, &htup->t_infomask,
8411                                                                    &htup->t_infomask2);
8412                 HeapTupleHeaderSetXmax(htup, xlrec->old_xmax);
8413                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
8414                 /* Set forward chain link in t_ctid */
8415                 htup->t_ctid = newtid;
8416
8417                 /* Mark the page as a candidate for pruning */
8418                 PageSetPrunable(page, XLogRecGetXid(record));
8419
8420                 if (xlrec->flags & XLH_UPDATE_OLD_ALL_VISIBLE_CLEARED)
8421                         PageClearAllVisible(page);
8422
8423                 PageSetLSN(page, lsn);
8424                 MarkBufferDirty(obuffer);
8425         }
8426
8427         /*
8428          * Read the page the new tuple goes into, if different from old.
8429          */
8430         if (oldblk == newblk)
8431         {
8432                 nbuffer = obuffer;
8433                 newaction = oldaction;
8434         }
8435         else if (XLogRecGetInfo(record) & XLOG_HEAP_INIT_PAGE)
8436         {
8437                 nbuffer = XLogInitBufferForRedo(record, 0);
8438                 page = (Page) BufferGetPage(nbuffer);
8439                 PageInit(page, BufferGetPageSize(nbuffer), 0);
8440                 newaction = BLK_NEEDS_REDO;
8441         }
8442         else
8443                 newaction = XLogReadBufferForRedo(record, 0, &nbuffer);
8444
8445         /*
8446          * The visibility map may need to be fixed even if the heap page is
8447          * already up-to-date.
8448          */
8449         if (xlrec->flags & XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED)
8450         {
8451                 Relation        reln = CreateFakeRelcacheEntry(rnode);
8452                 Buffer          vmbuffer = InvalidBuffer;
8453
8454                 visibilitymap_pin(reln, newblk, &vmbuffer);
8455                 visibilitymap_clear(reln, newblk, vmbuffer, VISIBILITYMAP_VALID_BITS);
8456                 ReleaseBuffer(vmbuffer);
8457                 FreeFakeRelcacheEntry(reln);
8458         }
8459
8460         /* Deal with new tuple */
8461         if (newaction == BLK_NEEDS_REDO)
8462         {
8463                 char       *recdata;
8464                 char       *recdata_end;
8465                 Size            datalen;
8466                 Size            tuplen;
8467
8468                 recdata = XLogRecGetBlockData(record, 0, &datalen);
8469                 recdata_end = recdata + datalen;
8470
8471                 page = BufferGetPage(nbuffer);
8472
8473                 offnum = xlrec->new_offnum;
8474                 if (PageGetMaxOffsetNumber(page) + 1 < offnum)
8475                         elog(PANIC, "invalid max offset number");
8476
8477                 if (xlrec->flags & XLH_UPDATE_PREFIX_FROM_OLD)
8478                 {
8479                         Assert(newblk == oldblk);
8480                         memcpy(&prefixlen, recdata, sizeof(uint16));
8481                         recdata += sizeof(uint16);
8482                 }
8483                 if (xlrec->flags & XLH_UPDATE_SUFFIX_FROM_OLD)
8484                 {
8485                         Assert(newblk == oldblk);
8486                         memcpy(&suffixlen, recdata, sizeof(uint16));
8487                         recdata += sizeof(uint16);
8488                 }
8489
8490                 memcpy((char *) &xlhdr, recdata, SizeOfHeapHeader);
8491                 recdata += SizeOfHeapHeader;
8492
8493                 tuplen = recdata_end - recdata;
8494                 Assert(tuplen <= MaxHeapTupleSize);
8495
8496                 htup = &tbuf.hdr;
8497                 MemSet((char *) htup, 0, SizeofHeapTupleHeader);
8498
8499                 /*
8500                  * Reconstruct the new tuple using the prefix and/or suffix from the
8501                  * old tuple, and the data stored in the WAL record.
8502                  */
8503                 newp = (char *) htup + SizeofHeapTupleHeader;
8504                 if (prefixlen > 0)
8505                 {
8506                         int                     len;
8507
8508                         /* copy bitmap [+ padding] [+ oid] from WAL record */
8509                         len = xlhdr.t_hoff - SizeofHeapTupleHeader;
8510                         memcpy(newp, recdata, len);
8511                         recdata += len;
8512                         newp += len;
8513
8514                         /* copy prefix from old tuple */
8515                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_data->t_hoff, prefixlen);
8516                         newp += prefixlen;
8517
8518                         /* copy new tuple data from WAL record */
8519                         len = tuplen - (xlhdr.t_hoff - SizeofHeapTupleHeader);
8520                         memcpy(newp, recdata, len);
8521                         recdata += len;
8522                         newp += len;
8523                 }
8524                 else
8525                 {
8526                         /*
8527                          * copy bitmap [+ padding] [+ oid] + data from record, all in one
8528                          * go
8529                          */
8530                         memcpy(newp, recdata, tuplen);
8531                         recdata += tuplen;
8532                         newp += tuplen;
8533                 }
8534                 Assert(recdata == recdata_end);
8535
8536                 /* copy suffix from old tuple */
8537                 if (suffixlen > 0)
8538                         memcpy(newp, (char *) oldtup.t_data + oldtup.t_len - suffixlen, suffixlen);
8539
8540                 newlen = SizeofHeapTupleHeader + tuplen + prefixlen + suffixlen;
8541                 htup->t_infomask2 = xlhdr.t_infomask2;
8542                 htup->t_infomask = xlhdr.t_infomask;
8543                 htup->t_hoff = xlhdr.t_hoff;
8544
8545                 HeapTupleHeaderSetXmin(htup, XLogRecGetXid(record));
8546                 HeapTupleHeaderSetCmin(htup, FirstCommandId);
8547                 HeapTupleHeaderSetXmax(htup, xlrec->new_xmax);
8548                 /* Make sure there is no forward chain link in t_ctid */
8549                 htup->t_ctid = newtid;
8550
8551                 offnum = PageAddItem(page, (Item) htup, newlen, offnum, true, true);
8552                 if (offnum == InvalidOffsetNumber)
8553                         elog(PANIC, "failed to add tuple");
8554
8555                 if (xlrec->flags & XLH_UPDATE_NEW_ALL_VISIBLE_CLEARED)
8556                         PageClearAllVisible(page);
8557
8558                 freespace = PageGetHeapFreeSpace(page); /* needed to update FSM below */
8559
8560                 PageSetLSN(page, lsn);
8561                 MarkBufferDirty(nbuffer);
8562         }
8563
8564         if (BufferIsValid(nbuffer) && nbuffer != obuffer)
8565                 UnlockReleaseBuffer(nbuffer);
8566         if (BufferIsValid(obuffer))
8567                 UnlockReleaseBuffer(obuffer);
8568
8569         /*
8570          * If the new page is running low on free space, update the FSM as well.
8571          * Arbitrarily, our definition of "low" is less than 20%. We can't do much
8572          * better than that without knowing the fill-factor for the table.
8573          *
8574          * However, don't update the FSM on HOT updates, because after crash
8575          * recovery, either the old or the new tuple will certainly be dead and
8576          * prunable. After pruning, the page will have roughly as much free space
8577          * as it did before the update, assuming the new tuple is about the same
8578          * size as the old one.
8579          *
8580          * XXX: Don't do this if the page was restored from full page image. We
8581          * don't bother to update the FSM in that case, it doesn't need to be
8582          * totally accurate anyway.
8583          */
8584         if (newaction == BLK_NEEDS_REDO && !hot_update && freespace < BLCKSZ / 5)
8585                 XLogRecordPageWithFreeSpace(rnode, newblk, freespace);
8586 }
8587
8588 static void
8589 heap_xlog_confirm(XLogReaderState *record)
8590 {
8591         XLogRecPtr      lsn = record->EndRecPtr;
8592         xl_heap_confirm *xlrec = (xl_heap_confirm *) XLogRecGetData(record);
8593         Buffer          buffer;
8594         Page            page;
8595         OffsetNumber offnum;
8596         ItemId          lp = NULL;
8597         HeapTupleHeader htup;
8598
8599         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8600         {
8601                 page = BufferGetPage(buffer);
8602
8603                 offnum = xlrec->offnum;
8604                 if (PageGetMaxOffsetNumber(page) >= offnum)
8605                         lp = PageGetItemId(page, offnum);
8606
8607                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8608                         elog(PANIC, "invalid lp");
8609
8610                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8611
8612                 /*
8613                  * Confirm tuple as actually inserted
8614                  */
8615                 ItemPointerSet(&htup->t_ctid, BufferGetBlockNumber(buffer), offnum);
8616
8617                 PageSetLSN(page, lsn);
8618                 MarkBufferDirty(buffer);
8619         }
8620         if (BufferIsValid(buffer))
8621                 UnlockReleaseBuffer(buffer);
8622 }
8623
8624 static void
8625 heap_xlog_lock(XLogReaderState *record)
8626 {
8627         XLogRecPtr      lsn = record->EndRecPtr;
8628         xl_heap_lock *xlrec = (xl_heap_lock *) XLogRecGetData(record);
8629         Buffer          buffer;
8630         Page            page;
8631         OffsetNumber offnum;
8632         ItemId          lp = NULL;
8633         HeapTupleHeader htup;
8634
8635         /*
8636          * The visibility map may need to be fixed even if the heap page is
8637          * already up-to-date.
8638          */
8639         if (xlrec->flags & XLH_LOCK_ALL_FROZEN_CLEARED)
8640         {
8641                 RelFileNode rnode;
8642                 Buffer          vmbuffer = InvalidBuffer;
8643                 BlockNumber block;
8644                 Relation        reln;
8645
8646                 XLogRecGetBlockTag(record, 0, &rnode, NULL, &block);
8647                 reln = CreateFakeRelcacheEntry(rnode);
8648
8649                 visibilitymap_pin(reln, block, &vmbuffer);
8650                 visibilitymap_clear(reln, block, vmbuffer, VISIBILITYMAP_ALL_FROZEN);
8651
8652                 ReleaseBuffer(vmbuffer);
8653                 FreeFakeRelcacheEntry(reln);
8654         }
8655
8656         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8657         {
8658                 page = (Page) BufferGetPage(buffer);
8659
8660                 offnum = xlrec->offnum;
8661                 if (PageGetMaxOffsetNumber(page) >= offnum)
8662                         lp = PageGetItemId(page, offnum);
8663
8664                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8665                         elog(PANIC, "invalid lp");
8666
8667                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8668
8669                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
8670                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
8671                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
8672                                                                    &htup->t_infomask2);
8673
8674                 /*
8675                  * Clear relevant update flags, but only if the modified infomask says
8676                  * there's no update.
8677                  */
8678                 if (HEAP_XMAX_IS_LOCKED_ONLY(htup->t_infomask))
8679                 {
8680                         HeapTupleHeaderClearHotUpdated(htup);
8681                         /* Make sure there is no forward chain link in t_ctid */
8682                         ItemPointerSet(&htup->t_ctid,
8683                                                    BufferGetBlockNumber(buffer),
8684                                                    offnum);
8685                 }
8686                 HeapTupleHeaderSetXmax(htup, xlrec->locking_xid);
8687                 HeapTupleHeaderSetCmax(htup, FirstCommandId, false);
8688                 PageSetLSN(page, lsn);
8689                 MarkBufferDirty(buffer);
8690         }
8691         if (BufferIsValid(buffer))
8692                 UnlockReleaseBuffer(buffer);
8693 }
8694
8695 static void
8696 heap_xlog_lock_updated(XLogReaderState *record)
8697 {
8698         XLogRecPtr      lsn = record->EndRecPtr;
8699         xl_heap_lock_updated *xlrec;
8700         Buffer          buffer;
8701         Page            page;
8702         OffsetNumber offnum;
8703         ItemId          lp = NULL;
8704         HeapTupleHeader htup;
8705
8706         xlrec = (xl_heap_lock_updated *) XLogRecGetData(record);
8707
8708         /*
8709          * The visibility map may need to be fixed even if the heap page is
8710          * already up-to-date.
8711          */
8712         if (xlrec->flags & XLH_LOCK_ALL_FROZEN_CLEARED)
8713         {
8714                 RelFileNode rnode;
8715                 Buffer          vmbuffer = InvalidBuffer;
8716                 BlockNumber block;
8717                 Relation        reln;
8718
8719                 XLogRecGetBlockTag(record, 0, &rnode, NULL, &block);
8720                 reln = CreateFakeRelcacheEntry(rnode);
8721
8722                 visibilitymap_pin(reln, block, &vmbuffer);
8723                 visibilitymap_clear(reln, block, vmbuffer, VISIBILITYMAP_ALL_FROZEN);
8724
8725                 ReleaseBuffer(vmbuffer);
8726                 FreeFakeRelcacheEntry(reln);
8727         }
8728
8729         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8730         {
8731                 page = BufferGetPage(buffer);
8732
8733                 offnum = xlrec->offnum;
8734                 if (PageGetMaxOffsetNumber(page) >= offnum)
8735                         lp = PageGetItemId(page, offnum);
8736
8737                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8738                         elog(PANIC, "invalid lp");
8739
8740                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8741
8742                 htup->t_infomask &= ~(HEAP_XMAX_BITS | HEAP_MOVED);
8743                 htup->t_infomask2 &= ~HEAP_KEYS_UPDATED;
8744                 fix_infomask_from_infobits(xlrec->infobits_set, &htup->t_infomask,
8745                                                                    &htup->t_infomask2);
8746                 HeapTupleHeaderSetXmax(htup, xlrec->xmax);
8747
8748                 PageSetLSN(page, lsn);
8749                 MarkBufferDirty(buffer);
8750         }
8751         if (BufferIsValid(buffer))
8752                 UnlockReleaseBuffer(buffer);
8753 }
8754
8755 static void
8756 heap_xlog_inplace(XLogReaderState *record)
8757 {
8758         XLogRecPtr      lsn = record->EndRecPtr;
8759         xl_heap_inplace *xlrec = (xl_heap_inplace *) XLogRecGetData(record);
8760         Buffer          buffer;
8761         Page            page;
8762         OffsetNumber offnum;
8763         ItemId          lp = NULL;
8764         HeapTupleHeader htup;
8765         uint32          oldlen;
8766         Size            newlen;
8767
8768         if (XLogReadBufferForRedo(record, 0, &buffer) == BLK_NEEDS_REDO)
8769         {
8770                 char       *newtup = XLogRecGetBlockData(record, 0, &newlen);
8771
8772                 page = BufferGetPage(buffer);
8773
8774                 offnum = xlrec->offnum;
8775                 if (PageGetMaxOffsetNumber(page) >= offnum)
8776                         lp = PageGetItemId(page, offnum);
8777
8778                 if (PageGetMaxOffsetNumber(page) < offnum || !ItemIdIsNormal(lp))
8779                         elog(PANIC, "invalid lp");
8780
8781                 htup = (HeapTupleHeader) PageGetItem(page, lp);
8782
8783                 oldlen = ItemIdGetLength(lp) - htup->t_hoff;
8784                 if (oldlen != newlen)
8785                         elog(PANIC, "wrong tuple length");
8786
8787                 memcpy((char *) htup + htup->t_hoff, newtup, newlen);
8788
8789                 PageSetLSN(page, lsn);
8790                 MarkBufferDirty(buffer);
8791         }
8792         if (BufferIsValid(buffer))
8793                 UnlockReleaseBuffer(buffer);
8794 }
8795
8796 void
8797 heap_redo(XLogReaderState *record)
8798 {
8799         uint8           info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8800
8801         /*
8802          * These operations don't overwrite MVCC data so no conflict processing is
8803          * required. The ones in heap2 rmgr do.
8804          */
8805
8806         switch (info & XLOG_HEAP_OPMASK)
8807         {
8808                 case XLOG_HEAP_INSERT:
8809                         heap_xlog_insert(record);
8810                         break;
8811                 case XLOG_HEAP_DELETE:
8812                         heap_xlog_delete(record);
8813                         break;
8814                 case XLOG_HEAP_UPDATE:
8815                         heap_xlog_update(record, false);
8816                         break;
8817                 case XLOG_HEAP_TRUNCATE:
8818
8819                         /*
8820                          * TRUNCATE is a no-op because the actions are already logged as
8821                          * SMGR WAL records.  TRUNCATE WAL record only exists for logical
8822                          * decoding.
8823                          */
8824                         break;
8825                 case XLOG_HEAP_HOT_UPDATE:
8826                         heap_xlog_update(record, true);
8827                         break;
8828                 case XLOG_HEAP_CONFIRM:
8829                         heap_xlog_confirm(record);
8830                         break;
8831                 case XLOG_HEAP_LOCK:
8832                         heap_xlog_lock(record);
8833                         break;
8834                 case XLOG_HEAP_INPLACE:
8835                         heap_xlog_inplace(record);
8836                         break;
8837                 default:
8838                         elog(PANIC, "heap_redo: unknown op code %u", info);
8839         }
8840 }
8841
8842 void
8843 heap2_redo(XLogReaderState *record)
8844 {
8845         uint8           info = XLogRecGetInfo(record) & ~XLR_INFO_MASK;
8846
8847         switch (info & XLOG_HEAP_OPMASK)
8848         {
8849                 case XLOG_HEAP2_CLEAN:
8850                         heap_xlog_clean(record);
8851                         break;
8852                 case XLOG_HEAP2_FREEZE_PAGE:
8853                         heap_xlog_freeze_page(record);
8854                         break;
8855                 case XLOG_HEAP2_CLEANUP_INFO:
8856                         heap_xlog_cleanup_info(record);
8857                         break;
8858                 case XLOG_HEAP2_VISIBLE:
8859                         heap_xlog_visible(record);
8860                         break;
8861                 case XLOG_HEAP2_MULTI_INSERT:
8862                         heap_xlog_multi_insert(record);
8863                         break;
8864                 case XLOG_HEAP2_LOCK_UPDATED:
8865                         heap_xlog_lock_updated(record);
8866                         break;
8867                 case XLOG_HEAP2_NEW_CID:
8868
8869                         /*
8870                          * Nothing to do on a real replay, only used during logical
8871                          * decoding.
8872                          */
8873                         break;
8874                 case XLOG_HEAP2_REWRITE:
8875                         heap_xlog_logical_rewrite(record);
8876                         break;
8877                 default:
8878                         elog(PANIC, "heap2_redo: unknown op code %u", info);
8879         }
8880 }
8881
8882 /*
8883  *      heap_sync               - sync a heap, for use when no WAL has been written
8884  *
8885  * This forces the heap contents (including TOAST heap if any) down to disk.
8886  * If we skipped using WAL, and WAL is otherwise needed, we must force the
8887  * relation down to disk before it's safe to commit the transaction.  This
8888  * requires writing out any dirty buffers and then doing a forced fsync.
8889  *
8890  * Indexes are not touched.  (Currently, index operations associated with
8891  * the commands that use this are WAL-logged and so do not need fsync.
8892  * That behavior might change someday, but in any case it's likely that
8893  * any fsync decisions required would be per-index and hence not appropriate
8894  * to be done here.)
8895  */
8896 void
8897 heap_sync(Relation rel)
8898 {
8899         /* non-WAL-logged tables never need fsync */
8900         if (!RelationNeedsWAL(rel))
8901                 return;
8902
8903         /* main heap */
8904         FlushRelationBuffers(rel);
8905         /* FlushRelationBuffers will have opened rd_smgr */
8906         smgrimmedsync(rel->rd_smgr, MAIN_FORKNUM);
8907
8908         /* FSM is not critical, don't bother syncing it */
8909
8910         /* toast heap, if any */
8911         if (OidIsValid(rel->rd_rel->reltoastrelid))
8912         {
8913                 Relation        toastrel;
8914
8915                 toastrel = table_open(rel->rd_rel->reltoastrelid, AccessShareLock);
8916                 FlushRelationBuffers(toastrel);
8917                 smgrimmedsync(toastrel->rd_smgr, MAIN_FORKNUM);
8918                 table_close(toastrel, AccessShareLock);
8919         }
8920 }
8921
8922 /*
8923  * Mask a heap page before performing consistency checks on it.
8924  */
8925 void
8926 heap_mask(char *pagedata, BlockNumber blkno)
8927 {
8928         Page            page = (Page) pagedata;
8929         OffsetNumber off;
8930
8931         mask_page_lsn_and_checksum(page);
8932
8933         mask_page_hint_bits(page);
8934         mask_unused_space(page);
8935
8936         for (off = 1; off <= PageGetMaxOffsetNumber(page); off++)
8937         {
8938                 ItemId          iid = PageGetItemId(page, off);
8939                 char       *page_item;
8940
8941                 page_item = (char *) (page + ItemIdGetOffset(iid));
8942
8943                 if (ItemIdIsNormal(iid))
8944                 {
8945                         HeapTupleHeader page_htup = (HeapTupleHeader) page_item;
8946
8947                         /*
8948                          * If xmin of a tuple is not yet frozen, we should ignore
8949                          * differences in hint bits, since they can be set without
8950                          * emitting WAL.
8951                          */
8952                         if (!HeapTupleHeaderXminFrozen(page_htup))
8953                                 page_htup->t_infomask &= ~HEAP_XACT_MASK;
8954                         else
8955                         {
8956                                 /* Still we need to mask xmax hint bits. */
8957                                 page_htup->t_infomask &= ~HEAP_XMAX_INVALID;
8958                                 page_htup->t_infomask &= ~HEAP_XMAX_COMMITTED;
8959                         }
8960
8961                         /*
8962                          * During replay, we set Command Id to FirstCommandId. Hence, mask
8963                          * it. See heap_xlog_insert() for details.
8964                          */
8965                         page_htup->t_choice.t_heap.t_field3.t_cid = MASK_MARKER;
8966
8967                         /*
8968                          * For a speculative tuple, heap_insert() does not set ctid in the
8969                          * caller-passed heap tuple itself, leaving the ctid field to
8970                          * contain a speculative token value - a per-backend monotonically
8971                          * increasing identifier. Besides, it does not WAL-log ctid under
8972                          * any circumstances.
8973                          *
8974                          * During redo, heap_xlog_insert() sets t_ctid to current block
8975                          * number and self offset number. It doesn't care about any
8976                          * speculative insertions in master. Hence, we set t_ctid to
8977                          * current block number and self offset number to ignore any
8978                          * inconsistency.
8979                          */
8980                         if (HeapTupleHeaderIsSpeculative(page_htup))
8981                                 ItemPointerSet(&page_htup->t_ctid, blkno, off);
8982
8983                         /*
8984                          * NB: Not ignoring ctid changes due to the tuple having moved
8985                          * (i.e. HeapTupleHeaderIndicatesMovedPartitions), because that's
8986                          * important information that needs to be in-sync between primary
8987                          * and standby, and thus is WAL logged.
8988                          */
8989                 }
8990
8991                 /*
8992                  * Ignore any padding bytes after the tuple, when the length of the
8993                  * item is not MAXALIGNed.
8994                  */
8995                 if (ItemIdHasStorage(iid))
8996                 {
8997                         int                     len = ItemIdGetLength(iid);
8998                         int                     padlen = MAXALIGN(len) - len;
8999
9000                         if (padlen > 0)
9001                                 memset(page_item + len, MASK_MARKER, padlen);
9002                 }
9003         }
9004 }