Fix possible duplicate tuples while GiST scan. Now page is processed

at once and ItemPointers are collected in memory.

Remove tuple's killing by killtuple() if tuple was moved to another
page - it could produce unaceptable overhead.

Backpatch up to 8.1 because the bug was introduced by GiST's concurrency support.

1 files changed, 93 insertions, 90 deletions

diff --git a/src/backend/access/gist/gistget.c b/src/backend/access/gist/gistget.c
index 1e02ec082f4..feeb84d38d7 100644
--- a/src/backend/access/gist/gistget.c
+++ b/src/backend/access/gist/gistget.c
@@ -8,7 +8,7 @@
  * Portions Copyright (c) 1994, Regents of the University of California
  *
  * IDENTIFICATION
- *	  $PostgreSQL: pgsql/src/backend/access/gist/gistget.c,v 1.52.2.1 2005/11/22 18:23:03 momjian Exp $
+ *	  $PostgreSQL: pgsql/src/backend/access/gist/gistget.c,v 1.52.2.2 2008/08/23 10:43:58 teodor Exp $
  *
  *-------------------------------------------------------------------------
  */
@@ -30,62 +30,38 @@ static bool gistindex_keytest(IndexTuple tuple, IndexScanDesc scan,
 static void
 killtuple(Relation r, GISTScanOpaque so, ItemPointer iptr)
 {
-	Buffer		buffer = so->curbuf;
+	Page        p;
+	OffsetNumber offset;
 
-	for (;;)
+	LockBuffer(so->curbuf, GIST_SHARE);
+	p = (Page) BufferGetPage(so->curbuf);
+ 
+	if (XLByteEQ(so->stack->lsn, PageGetLSN(p)))
 	{
-		Page		p;
-		BlockNumber blkno;
-		OffsetNumber offset,
-					maxoff;
-
-		LockBuffer(buffer, GIST_SHARE);
-		p = (Page) BufferGetPage(buffer);
-
-		if (buffer == so->curbuf && XLByteEQ(so->stack->lsn, PageGetLSN(p)))
-		{
-			/* page unchanged, so all is simple */
-			offset = ItemPointerGetOffsetNumber(iptr);
-			PageGetItemId(p, offset)->lp_flags |= LP_DELETE;
-			SetBufferCommitInfoNeedsSave(buffer);
-			LockBuffer(buffer, GIST_UNLOCK);
-			break;
-		}
-
-		maxoff = PageGetMaxOffsetNumber(p);
+		/* page unchanged, so all is simple */
+		offset = ItemPointerGetOffsetNumber(iptr);
+		PageGetItemId(p, offset)->lp_flags |= LP_DELETE;
+		SetBufferCommitInfoNeedsSave(so->curbuf);
+	}
+	else
+	{
+		OffsetNumber maxoff = PageGetMaxOffsetNumber(p);
 
 		for (offset = FirstOffsetNumber; offset <= maxoff; offset = OffsetNumberNext(offset))
 		{
-			IndexTuple	ituple = (IndexTuple) PageGetItem(p, PageGetItemId(p, offset));
+			IndexTuple  ituple = (IndexTuple) PageGetItem(p, PageGetItemId(p, offset));
 
 			if (ItemPointerEquals(&(ituple->t_tid), iptr))
 			{
 				/* found */
 				PageGetItemId(p, offset)->lp_flags |= LP_DELETE;
-				SetBufferCommitInfoNeedsSave(buffer);
-				LockBuffer(buffer, GIST_UNLOCK);
-				if (buffer != so->curbuf)
-					ReleaseBuffer(buffer);
-				return;
+				SetBufferCommitInfoNeedsSave(so->curbuf);
+				break;
 			}
 		}
-
-		/* follow right link */
-
-		/*
-		 * ??? is it good? if tuple dropped by concurrent vacuum, we will read
-		 * all leaf pages...
-		 */
-		blkno = GistPageGetOpaque(p)->rightlink;
-		LockBuffer(buffer, GIST_UNLOCK);
-		if (buffer != so->curbuf)
-			ReleaseBuffer(buffer);
-
-		if (blkno == InvalidBlockNumber)
-			/* can't found, dropped by somebody else */
-			return;
-		buffer = ReadBuffer(r, blkno);
 	}
+
+	LockBuffer(so->curbuf, GIST_UNLOCK);
 }
 
 /*
@@ -146,7 +122,6 @@ gistnext(IndexScanDesc scan, ScanDirection dir, ItemPointer tids, int maxtids, b
 	GISTSearchStack *stk;
 	IndexTuple	it;
 	GISTPageOpaque opaque;
-	bool		resetoffset = false;
 	int			ntids = 0;
 
 	so = (GISTScanOpaque) scan->opaque;
@@ -171,6 +146,42 @@ gistnext(IndexScanDesc scan, ScanDirection dir, ItemPointer tids, int maxtids, b
 		return 0;
 	}
 
+	/*
+	 * check stored pointers from last visit 
+	 */
+	if ( so->nPageData > 0 ) 
+	{
+		while( ntids < maxtids && so->curPageData < so->nPageData )
+		{
+			tids[ ntids ] = scan->xs_ctup.t_self = so->pageData[ so->curPageData ];
+				
+			so->curPageData ++;
+			ntids++;
+		}
+
+		if ( ntids == maxtids )
+			return ntids;
+		
+		/*
+		 * Go to the next page
+		 */
+		stk = so->stack->next;
+		pfree(so->stack);
+		so->stack = stk;
+
+		/* If we're out of stack entries, we're done */
+		if (so->stack == NULL)
+		{
+			ReleaseBuffer(so->curbuf);
+			so->curbuf = InvalidBuffer;
+			return ntids;
+		}
+
+		so->curbuf = ReleaseAndReadBuffer(so->curbuf,
+										  scan->indexRelation,
+										  stk->block);
+	}
+
 	for (;;)
 	{
 		/* First of all, we need lock buffer */
@@ -178,30 +189,25 @@ gistnext(IndexScanDesc scan, ScanDirection dir, ItemPointer tids, int maxtids, b
 		LockBuffer(so->curbuf, GIST_SHARE);
 		p = BufferGetPage(so->curbuf);
 		opaque = GistPageGetOpaque(p);
-		resetoffset = false;
 
-		if (XLogRecPtrIsInvalid(so->stack->lsn) || !XLByteEQ(so->stack->lsn, PageGetLSN(p)))
-		{
-			/* page changed from last visit or visit first time , reset offset */
-			so->stack->lsn = PageGetLSN(p);
-			resetoffset = true;
-
-			/* check page split, occured from last visit or visit to parent */
-			if (!XLogRecPtrIsInvalid(so->stack->parentlsn) &&
-				XLByteLT(so->stack->parentlsn, opaque->nsn) &&
-				opaque->rightlink != InvalidBlockNumber /* sanity check */ &&
-				(so->stack->next == NULL || so->stack->next->block != opaque->rightlink)		/* check if already
-					added */ )
-			{
-				/* detect page split, follow right link to add pages */
+		/* remember lsn to identify page changed for tuple's killing */
+		so->stack->lsn = PageGetLSN(p);
 
-				stk = (GISTSearchStack *) palloc(sizeof(GISTSearchStack));
-				stk->next = so->stack->next;
-				stk->block = opaque->rightlink;
-				stk->parentlsn = so->stack->parentlsn;
-				memset(&(stk->lsn), 0, sizeof(GistNSN));
-				so->stack->next = stk;
-			}
+		/* check page split, occured from last visit or visit to parent */
+		if (!XLogRecPtrIsInvalid(so->stack->parentlsn) &&
+			XLByteLT(so->stack->parentlsn, opaque->nsn) &&
+			opaque->rightlink != InvalidBlockNumber /* sanity check */ &&
+			(so->stack->next == NULL || so->stack->next->block != opaque->rightlink)		/* check if already
+					added */ )
+		{
+			/* detect page split, follow right link to add pages */
+
+			stk = (GISTSearchStack *) palloc(sizeof(GISTSearchStack));
+			stk->next = so->stack->next;
+			stk->block = opaque->rightlink;
+			stk->parentlsn = so->stack->parentlsn;
+			memset(&(stk->lsn), 0, sizeof(GistNSN));
+			so->stack->next = stk;
 		}
 
 		/* if page is empty, then just skip it */
@@ -224,24 +230,13 @@ gistnext(IndexScanDesc scan, ScanDirection dir, ItemPointer tids, int maxtids, b
 			continue;
 		}
 
-		if (!GistPageIsLeaf(p) || resetoffset || ItemPointerIsValid(&scan->currentItemData) == false)
-		{
-			if (ScanDirectionIsBackward(dir))
-				n = PageGetMaxOffsetNumber(p);
-			else
-				n = FirstOffsetNumber;
-		}
+		if (ScanDirectionIsBackward(dir))
+			n = PageGetMaxOffsetNumber(p);
 		else
-		{
-			n = ItemPointerGetOffsetNumber(&(scan->currentItemData));
-
-			if (ScanDirectionIsBackward(dir))
-				n = OffsetNumberPrev(n);
-			else
-				n = OffsetNumberNext(n);
-		}
+			n = FirstOffsetNumber;
 
 		/* wonderfull, we can look at page */
+		so->nPageData = so->curPageData = 0;
 
 		for (;;)
 		{
@@ -249,6 +244,20 @@ gistnext(IndexScanDesc scan, ScanDirection dir, ItemPointer tids, int maxtids, b
 
 			if (!OffsetNumberIsValid(n))
 			{
+				while( ntids < maxtids && so->curPageData < so->nPageData )
+				{
+					tids[ ntids ] = scan->xs_ctup.t_self = so->pageData[ so->curPageData ];
+				
+					so->curPageData ++;
+					ntids++;
+				}
+
+				if ( ntids == maxtids )
+				{
+					LockBuffer(so->curbuf, GIST_UNLOCK);
+					return ntids;
+				}
+
 				/*
 				 * We ran out of matching index entries on the current page,
 				 * so pop the top stack entry and use it to continue the
@@ -288,14 +297,8 @@ gistnext(IndexScanDesc scan, ScanDirection dir, ItemPointer tids, int maxtids, b
 				if (!(ignore_killed_tuples && ItemIdDeleted(PageGetItemId(p, n))))
 				{
 					it = (IndexTuple) PageGetItem(p, PageGetItemId(p, n));
-					tids[ntids] = scan->xs_ctup.t_self = it->t_tid;
-					ntids++;
-
-					if (ntids == maxtids)
-					{
-						LockBuffer(so->curbuf, GIST_UNLOCK);
-						return ntids;
-					}
+					so->pageData[ so->nPageData ] = it->t_tid;
+					so->nPageData ++;
 				}
 			}
 			else