granicus.if.org Git - postgresql/blob - src/backend/executor/nodeTidscan.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * nodeTidscan.c
   4  *        Routines to support direct tid scans of relations
   5  *
   6  * Portions Copyright (c) 1996-2014, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        src/backend/executor/nodeTidscan.c
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15 /*
  16  * INTERFACE ROUTINES
  17  *
  18  *              ExecTidScan                     scans a relation using tids
  19  *              ExecInitTidScan         creates and initializes state info.
  20  *              ExecReScanTidScan       rescans the tid relation.
  21  *              ExecEndTidScan          releases all storage.
  22  */
  23 #include "postgres.h"
  24
  25 #include "access/sysattr.h"
  26 #include "catalog/pg_type.h"
  27 #include "executor/execdebug.h"
  28 #include "executor/nodeTidscan.h"
  29 #include "optimizer/clauses.h"
  30 #include "storage/bufmgr.h"
  31 #include "utils/array.h"
  32 #include "utils/rel.h"
  33
  34
  35 #define IsCTIDVar(node)  \
  36         ((node) != NULL && \
  37          IsA((node), Var) && \
  38          ((Var *) (node))->varattno == SelfItemPointerAttributeNumber && \
  39          ((Var *) (node))->varlevelsup == 0)
  40
  41 static void TidListCreate(TidScanState *tidstate);
  42 static int      itemptr_comparator(const void *a, const void *b);
  43 static TupleTableSlot *TidNext(TidScanState *node);
  44
  45
  46 /*
  47  * Compute the list of TIDs to be visited, by evaluating the expressions
  48  * for them.
  49  *
  50  * (The result is actually an array, not a list.)
  51  */
  52 static void
  53 TidListCreate(TidScanState *tidstate)
  54 {
  55         List       *evalList = tidstate->tss_tidquals;
  56         ExprContext *econtext = tidstate->ss.ps.ps_ExprContext;
  57         BlockNumber nblocks;
  58         ItemPointerData *tidList;
  59         int                     numAllocTids;
  60         int                     numTids;
  61         ListCell   *l;
  62
  63         /*
  64          * We silently discard any TIDs that are out of range at the time of scan
  65          * start.  (Since we hold at least AccessShareLock on the table, it won't
  66          * be possible for someone to truncate away the blocks we intend to
  67          * visit.)
  68          */
  69         nblocks = RelationGetNumberOfBlocks(tidstate->ss.ss_currentRelation);
  70
  71         /*
  72          * We initialize the array with enough slots for the case that all quals
  73          * are simple OpExprs or CurrentOfExprs.  If there are any
  74          * ScalarArrayOpExprs, we may have to enlarge the array.
  75          */
  76         numAllocTids = list_length(evalList);
  77         tidList = (ItemPointerData *)
  78                 palloc(numAllocTids * sizeof(ItemPointerData));
  79         numTids = 0;
  80         tidstate->tss_isCurrentOf = false;
  81
  82         foreach(l, evalList)
  83         {
  84                 ExprState  *exstate = (ExprState *) lfirst(l);
  85                 Expr       *expr = exstate->expr;
  86                 ItemPointer itemptr;
  87                 bool            isNull;
  88
  89                 if (is_opclause(expr))
  90                 {
  91                         FuncExprState *fexstate = (FuncExprState *) exstate;
  92                         Node       *arg1;
  93                         Node       *arg2;
  94
  95                         arg1 = get_leftop(expr);
  96                         arg2 = get_rightop(expr);
  97                         if (IsCTIDVar(arg1))
  98                                 exstate = (ExprState *) lsecond(fexstate->args);
  99                         else if (IsCTIDVar(arg2))
 100                                 exstate = (ExprState *) linitial(fexstate->args);
 101                         else
 102                                 elog(ERROR, "could not identify CTID variable");
 103
 104                         itemptr = (ItemPointer)
 105                                 DatumGetPointer(ExecEvalExprSwitchContext(exstate,
 106                                                                                                                   econtext,
 107                                                                                                                   &isNull,
 108                                                                                                                   NULL));
 109                         if (!isNull &&
 110                                 ItemPointerIsValid(itemptr) &&
 111                                 ItemPointerGetBlockNumber(itemptr) < nblocks)
 112                         {
 113                                 if (numTids >= numAllocTids)
 114                                 {
 115                                         numAllocTids *= 2;
 116                                         tidList = (ItemPointerData *)
 117                                                 repalloc(tidList,
 118                                                                  numAllocTids * sizeof(ItemPointerData));
 119                                 }
 120                                 tidList[numTids++] = *itemptr;
 121                         }
 122                 }
 123                 else if (expr && IsA(expr, ScalarArrayOpExpr))
 124                 {
 125                         ScalarArrayOpExprState *saexstate = (ScalarArrayOpExprState *) exstate;
 126                         Datum           arraydatum;
 127                         ArrayType  *itemarray;
 128                         Datum      *ipdatums;
 129                         bool       *ipnulls;
 130                         int                     ndatums;
 131                         int                     i;
 132
 133                         exstate = (ExprState *) lsecond(saexstate->fxprstate.args);
 134                         arraydatum = ExecEvalExprSwitchContext(exstate,
 135                                                                                                    econtext,
 136                                                                                                    &isNull,
 137                                                                                                    NULL);
 138                         if (isNull)
 139                                 continue;
 140                         itemarray = DatumGetArrayTypeP(arraydatum);
 141                         deconstruct_array(itemarray,
 142                                                           TIDOID, SizeOfIptrData, false, 's',
 143                                                           &ipdatums, &ipnulls, &ndatums);
 144                         if (numTids + ndatums > numAllocTids)
 145                         {
 146                                 numAllocTids = numTids + ndatums;
 147                                 tidList = (ItemPointerData *)
 148                                         repalloc(tidList,
 149                                                          numAllocTids * sizeof(ItemPointerData));
 150                         }
 151                         for (i = 0; i < ndatums; i++)
 152                         {
 153                                 if (!ipnulls[i])
 154                                 {
 155                                         itemptr = (ItemPointer) DatumGetPointer(ipdatums[i]);
 156                                         if (ItemPointerIsValid(itemptr) &&
 157                                                 ItemPointerGetBlockNumber(itemptr) < nblocks)
 158                                                 tidList[numTids++] = *itemptr;
 159                                 }
 160                         }
 161                         pfree(ipdatums);
 162                         pfree(ipnulls);
 163                 }
 164                 else if (expr && IsA(expr, CurrentOfExpr))
 165                 {
 166                         CurrentOfExpr *cexpr = (CurrentOfExpr *) expr;
 167                         ItemPointerData cursor_tid;
 168
 169                         if (execCurrentOf(cexpr, econtext,
 170                                                    RelationGetRelid(tidstate->ss.ss_currentRelation),
 171                                                           &cursor_tid))
 172                         {
 173                                 if (numTids >= numAllocTids)
 174                                 {
 175                                         numAllocTids *= 2;
 176                                         tidList = (ItemPointerData *)
 177                                                 repalloc(tidList,
 178                                                                  numAllocTids * sizeof(ItemPointerData));
 179                                 }
 180                                 tidList[numTids++] = cursor_tid;
 181                                 tidstate->tss_isCurrentOf = true;
 182                         }
 183                 }
 184                 else
 185                         elog(ERROR, "could not identify CTID expression");
 186         }
 187
 188         /*
 189          * Sort the array of TIDs into order, and eliminate duplicates.
 190          * Eliminating duplicates is necessary since we want OR semantics across
 191          * the list.  Sorting makes it easier to detect duplicates, and as a bonus
 192          * ensures that we will visit the heap in the most efficient way.
 193          */
 194         if (numTids > 1)
 195         {
 196                 int                     lastTid;
 197                 int                     i;
 198
 199                 /* CurrentOfExpr could never appear OR'd with something else */
 200                 Assert(!tidstate->tss_isCurrentOf);
 201
 202                 qsort((void *) tidList, numTids, sizeof(ItemPointerData),
 203                           itemptr_comparator);
 204                 lastTid = 0;
 205                 for (i = 1; i < numTids; i++)
 206                 {
 207                         if (!ItemPointerEquals(&tidList[lastTid], &tidList[i]))
 208                                 tidList[++lastTid] = tidList[i];
 209                 }
 210                 numTids = lastTid + 1;
 211         }
 212
 213         tidstate->tss_TidList = tidList;
 214         tidstate->tss_NumTids = numTids;
 215         tidstate->tss_TidPtr = -1;
 216 }
 217
 218 /*
 219  * qsort comparator for ItemPointerData items
 220  */
 221 static int
 222 itemptr_comparator(const void *a, const void *b)
 223 {
 224         const ItemPointerData *ipa = (const ItemPointerData *) a;
 225         const ItemPointerData *ipb = (const ItemPointerData *) b;
 226         BlockNumber ba = ItemPointerGetBlockNumber(ipa);
 227         BlockNumber bb = ItemPointerGetBlockNumber(ipb);
 228         OffsetNumber oa = ItemPointerGetOffsetNumber(ipa);
 229         OffsetNumber ob = ItemPointerGetOffsetNumber(ipb);
 230
 231         if (ba < bb)
 232                 return -1;
 233         if (ba > bb)
 234                 return 1;
 235         if (oa < ob)
 236                 return -1;
 237         if (oa > ob)
 238                 return 1;
 239         return 0;
 240 }
 241
 242 /* ----------------------------------------------------------------
 243  *              TidNext
 244  *
 245  *              Retrieve a tuple from the TidScan node's currentRelation
 246  *              using the tids in the TidScanState information.
 247  *
 248  * ----------------------------------------------------------------
 249  */
 250 static TupleTableSlot *
 251 TidNext(TidScanState *node)
 252 {
 253         EState     *estate;
 254         ScanDirection direction;
 255         Snapshot        snapshot;
 256         Relation        heapRelation;
 257         HeapTuple       tuple;
 258         TupleTableSlot *slot;
 259         Buffer          buffer = InvalidBuffer;
 260         ItemPointerData *tidList;
 261         int                     numTids;
 262         bool            bBackward;
 263
 264         /*
 265          * extract necessary information from tid scan node
 266          */
 267         estate = node->ss.ps.state;
 268         direction = estate->es_direction;
 269         snapshot = estate->es_snapshot;
 270         heapRelation = node->ss.ss_currentRelation;
 271         slot = node->ss.ss_ScanTupleSlot;
 272
 273         /*
 274          * First time through, compute the list of TIDs to be visited
 275          */
 276         if (node->tss_TidList == NULL)
 277                 TidListCreate(node);
 278
 279         tidList = node->tss_TidList;
 280         numTids = node->tss_NumTids;
 281
 282         tuple = &(node->tss_htup);
 283
 284         /*
 285          * Initialize or advance scan position, depending on direction.
 286          */
 287         bBackward = ScanDirectionIsBackward(direction);
 288         if (bBackward)
 289         {
 290                 if (node->tss_TidPtr < 0)
 291                 {
 292                         /* initialize for backward scan */
 293                         node->tss_TidPtr = numTids - 1;
 294                 }
 295                 else
 296                         node->tss_TidPtr--;
 297         }
 298         else
 299         {
 300                 if (node->tss_TidPtr < 0)
 301                 {
 302                         /* initialize for forward scan */
 303                         node->tss_TidPtr = 0;
 304                 }
 305                 else
 306                         node->tss_TidPtr++;
 307         }
 308
 309         while (node->tss_TidPtr >= 0 && node->tss_TidPtr < numTids)
 310         {
 311                 tuple->t_self = tidList[node->tss_TidPtr];
 312
 313                 /*
 314                  * For WHERE CURRENT OF, the tuple retrieved from the cursor might
 315                  * since have been updated; if so, we should fetch the version that is
 316                  * current according to our snapshot.
 317                  */
 318                 if (node->tss_isCurrentOf)
 319                         heap_get_latest_tid(heapRelation, snapshot, &tuple->t_self);
 320
 321                 if (heap_fetch(heapRelation, snapshot, tuple, &buffer, false, NULL))
 322                 {
 323                         /*
 324                          * store the scanned tuple in the scan tuple slot of the scan
 325                          * state.  Eventually we will only do this and not return a tuple.
 326                          * Note: we pass 'false' because tuples returned by amgetnext are
 327                          * pointers onto disk pages and were not created with palloc() and
 328                          * so should not be pfree()'d.
 329                          */
 330                         ExecStoreTuple(tuple,           /* tuple to store */
 331                                                    slot,        /* slot to store in */
 332                                                    buffer,              /* buffer associated with tuple  */
 333                                                    false);              /* don't pfree */
 334
 335                         /*
 336                          * At this point we have an extra pin on the buffer, because
 337                          * ExecStoreTuple incremented the pin count. Drop our local pin.
 338                          */
 339                         ReleaseBuffer(buffer);
 340
 341                         return slot;
 342                 }
 343                 /* Bad TID or failed snapshot qual; try next */
 344                 if (bBackward)
 345                         node->tss_TidPtr--;
 346                 else
 347                         node->tss_TidPtr++;
 348         }
 349
 350         /*
 351          * if we get here it means the tid scan failed so we are at the end of the
 352          * scan..
 353          */
 354         return ExecClearTuple(slot);
 355 }
 356
 357 /*
 358  * TidRecheck -- access method routine to recheck a tuple in EvalPlanQual
 359  */
 360 static bool
 361 TidRecheck(TidScanState *node, TupleTableSlot *slot)
 362 {
 363         /*
 364          * XXX shouldn't we check here to make sure tuple matches TID list? In
 365          * runtime-key case this is not certain, is it?  However, in the WHERE
 366          * CURRENT OF case it might not match anyway ...
 367          */
 368         return true;
 369 }
 370
 371
 372 /* ----------------------------------------------------------------
 373  *              ExecTidScan(node)
 374  *
 375  *              Scans the relation using tids and returns
 376  *                 the next qualifying tuple in the direction specified.
 377  *              We call the ExecScan() routine and pass it the appropriate
 378  *              access method functions.
 379  *
 380  *              Conditions:
 381  *                -- the "cursor" maintained by the AMI is positioned at the tuple
 382  *                       returned previously.
 383  *
 384  *              Initial States:
 385  *                -- the relation indicated is opened for scanning so that the
 386  *                       "cursor" is positioned before the first qualifying tuple.
 387  *                -- tidPtr is -1.
 388  * ----------------------------------------------------------------
 389  */
 390 TupleTableSlot *
 391 ExecTidScan(TidScanState *node)
 392 {
 393         return ExecScan(&node->ss,
 394                                         (ExecScanAccessMtd) TidNext,
 395                                         (ExecScanRecheckMtd) TidRecheck);
 396 }
 397
 398 /* ----------------------------------------------------------------
 399  *              ExecReScanTidScan(node)
 400  * ----------------------------------------------------------------
 401  */
 402 void
 403 ExecReScanTidScan(TidScanState *node)
 404 {
 405         if (node->tss_TidList)
 406                 pfree(node->tss_TidList);
 407         node->tss_TidList = NULL;
 408         node->tss_NumTids = 0;
 409         node->tss_TidPtr = -1;
 410
 411         ExecScanReScan(&node->ss);
 412 }
 413
 414 /* ----------------------------------------------------------------
 415  *              ExecEndTidScan
 416  *
 417  *              Releases any storage allocated through C routines.
 418  *              Returns nothing.
 419  * ----------------------------------------------------------------
 420  */
 421 void
 422 ExecEndTidScan(TidScanState *node)
 423 {
 424         /*
 425          * Free the exprcontext
 426          */
 427         ExecFreeExprContext(&node->ss.ps);
 428
 429         /*
 430          * clear out tuple table slots
 431          */
 432         ExecClearTuple(node->ss.ps.ps_ResultTupleSlot);
 433         ExecClearTuple(node->ss.ss_ScanTupleSlot);
 434
 435         /*
 436          * close the heap relation.
 437          */
 438         ExecCloseScanRelation(node->ss.ss_currentRelation);
 439 }
 440
 441 /* ----------------------------------------------------------------
 442  *              ExecInitTidScan
 443  *
 444  *              Initializes the tid scan's state information, creates
 445  *              scan keys, and opens the base and tid relations.
 446  *
 447  *              Parameters:
 448  *                node: TidNode node produced by the planner.
 449  *                estate: the execution state initialized in InitPlan.
 450  * ----------------------------------------------------------------
 451  */
 452 TidScanState *
 453 ExecInitTidScan(TidScan *node, EState *estate, int eflags)
 454 {
 455         TidScanState *tidstate;
 456         Relation        currentRelation;
 457
 458         /*
 459          * create state structure
 460          */
 461         tidstate = makeNode(TidScanState);
 462         tidstate->ss.ps.plan = (Plan *) node;
 463         tidstate->ss.ps.state = estate;
 464
 465         /*
 466          * Miscellaneous initialization
 467          *
 468          * create expression context for node
 469          */
 470         ExecAssignExprContext(estate, &tidstate->ss.ps);
 471
 472         tidstate->ss.ps.ps_TupFromTlist = false;
 473
 474         /*
 475          * initialize child expressions
 476          */
 477         tidstate->ss.ps.targetlist = (List *)
 478                 ExecInitExpr((Expr *) node->scan.plan.targetlist,
 479                                          (PlanState *) tidstate);
 480         tidstate->ss.ps.qual = (List *)
 481                 ExecInitExpr((Expr *) node->scan.plan.qual,
 482                                          (PlanState *) tidstate);
 483
 484         tidstate->tss_tidquals = (List *)
 485                 ExecInitExpr((Expr *) node->tidquals,
 486                                          (PlanState *) tidstate);
 487
 488         /*
 489          * tuple table initialization
 490          */
 491         ExecInitResultTupleSlot(estate, &tidstate->ss.ps);
 492         ExecInitScanTupleSlot(estate, &tidstate->ss);
 493
 494         /*
 495          * mark tid list as not computed yet
 496          */
 497         tidstate->tss_TidList = NULL;
 498         tidstate->tss_NumTids = 0;
 499         tidstate->tss_TidPtr = -1;
 500
 501         /*
 502          * open the base relation and acquire appropriate lock on it.
 503          */
 504         currentRelation = ExecOpenScanRelation(estate, node->scan.scanrelid, eflags);
 505
 506         tidstate->ss.ss_currentRelation = currentRelation;
 507         tidstate->ss.ss_currentScanDesc = NULL;         /* no heap scan here */
 508
 509         /*
 510          * get the scan type from the relation descriptor.
 511          */
 512         ExecAssignScanType(&tidstate->ss, RelationGetDescr(currentRelation));
 513
 514         /*
 515          * Initialize result tuple type and projection info.
 516          */
 517         ExecAssignResultTypeFromTL(&tidstate->ss.ps);
 518         ExecAssignScanProjectionInfo(&tidstate->ss);
 519
 520         /*
 521          * all done.
 522          */
 523         return tidstate;
 524 }