v14-0001-Refactor-nbtree-insertion-scankeys.patch

From eed4ac506df1ee54faf82d820fde8b1943140f34 Mon Sep 17 00:00:00 2001
From: Peter Geoghegan <pg@bowt.ie>
Date: Sat, 29 Dec 2018 15:34:48 -0800
Subject: [PATCH v14 1/7] Refactor nbtree insertion scankeys.

Use dedicated struct to represent nbtree insertion scan keys.  Having a
dedicated struct makes the difference between search type scankeys and
insertion scankeys a lot clearer, and simplifies the signature of
several related functions.

Use the new struct to store mutable state about an in-progress binary
search, rather than having _bt_check_unique() callers cache
_bt_binsrch() effort in an ad-hoc manner.  This makes it easy to add a
new optimization: _bt_check_unique() now falls out of its loop
immediately in the common case where it's already clear that there
couldn't possibly be a duplicate.  More importantly, the new
_bt_check_unique() scheme makes it a lot easier to manage cached binary
search effort afterwards, from within _bt_findinsertloc().  This is
needed for the upcoming patch to make nbtree tuples unique by treating
heap TID as a final tie-breaker column.

Based on a suggestion by Andrey Lepikhov.
---
 contrib/amcheck/verify_nbtree.c       |  52 ++--
 src/backend/access/nbtree/nbtinsert.c | 334 ++++++++++++++++----------
 src/backend/access/nbtree/nbtpage.c   |  12 +-
 src/backend/access/nbtree/nbtsearch.c | 168 ++++++++-----
 src/backend/access/nbtree/nbtsort.c   |   8 +-
 src/backend/access/nbtree/nbtutils.c  |  98 +++-----
 src/backend/utils/sort/tuplesort.c    |  16 +-
 src/include/access/nbtree.h           |  61 ++++-
 8 files changed, 429 insertions(+), 320 deletions(-)

diff --git a/contrib/amcheck/verify_nbtree.c b/contrib/amcheck/verify_nbtree.c
index 964200a767..053ac9d192 100644
--- a/contrib/amcheck/verify_nbtree.c
+++ b/contrib/amcheck/verify_nbtree.c
@@ -126,9 +126,9 @@ static void bt_check_every_level(Relation rel, Relation heaprel,
 static BtreeLevel bt_check_level_from_leftmost(BtreeCheckState *state,
 							 BtreeLevel level);
 static void bt_target_page_check(BtreeCheckState *state);
-static ScanKey bt_right_page_check_scankey(BtreeCheckState *state);
-static void bt_downlink_check(BtreeCheckState *state, BlockNumber childblock,
-				  ScanKey targetkey);
+static BTScanInsert bt_right_page_check_scankey(BtreeCheckState *state);
+static void bt_downlink_check(BtreeCheckState *state, BTScanInsert targetkey,
+				  BlockNumber childblock);
 static void bt_downlink_missing_check(BtreeCheckState *state);
 static void bt_tuple_present_callback(Relation index, HeapTuple htup,
 						  Datum *values, bool *isnull,
@@ -138,14 +138,14 @@ static IndexTuple bt_normalize_tuple(BtreeCheckState *state,
 static inline bool offset_is_negative_infinity(BTPageOpaque opaque,
 							OffsetNumber offset);
 static inline bool invariant_leq_offset(BtreeCheckState *state,
-					 ScanKey key,
+					 BTScanInsert key,
 					 OffsetNumber upperbound);
 static inline bool invariant_geq_offset(BtreeCheckState *state,
-					 ScanKey key,
+					 BTScanInsert key,
 					 OffsetNumber lowerbound);
 static inline bool invariant_leq_nontarget_offset(BtreeCheckState *state,
-							   Page other,
-							   ScanKey key,
+							   BTScanInsert key,
+							   Page nontarget,
 							   OffsetNumber upperbound);
 static Page palloc_btree_page(BtreeCheckState *state, BlockNumber blocknum);
 
@@ -837,8 +837,8 @@ bt_target_page_check(BtreeCheckState *state)
 	{
 		ItemId		itemid;
 		IndexTuple	itup;
-		ScanKey		skey;
 		size_t		tupsize;
+		BTScanInsert skey;
 
 		CHECK_FOR_INTERRUPTS();
 
@@ -1029,7 +1029,7 @@ bt_target_page_check(BtreeCheckState *state)
 		 */
 		else if (offset == max)
 		{
-			ScanKey		rightkey;
+			BTScanInsert	rightkey;
 
 			/* Get item in next/right page */
 			rightkey = bt_right_page_check_scankey(state);
@@ -1081,7 +1081,7 @@ bt_target_page_check(BtreeCheckState *state)
 		{
 			BlockNumber childblock = BTreeInnerTupleGetDownLink(itup);
 
-			bt_downlink_check(state, childblock, skey);
+			bt_downlink_check(state, skey, childblock);
 		}
 	}
 
@@ -1110,11 +1110,12 @@ bt_target_page_check(BtreeCheckState *state)
  * Note that !readonly callers must reverify that target page has not
  * been concurrently deleted.
  */
-static ScanKey
+static BTScanInsert
 bt_right_page_check_scankey(BtreeCheckState *state)
 {
 	BTPageOpaque opaque;
 	ItemId		rightitem;
+	IndexTuple	firstitup;
 	BlockNumber targetnext;
 	Page		rightpage;
 	OffsetNumber nline;
@@ -1302,8 +1303,8 @@ bt_right_page_check_scankey(BtreeCheckState *state)
 	 * Return first real item scankey.  Note that this relies on right page
 	 * memory remaining allocated.
 	 */
-	return _bt_mkscankey(state->rel,
-						 (IndexTuple) PageGetItem(rightpage, rightitem));
+	firstitup = (IndexTuple) PageGetItem(rightpage, rightitem);
+	return _bt_mkscankey(state->rel, firstitup);
 }
 
 /*
@@ -1316,8 +1317,8 @@ bt_right_page_check_scankey(BtreeCheckState *state)
  * verification this way around is much more practical.
  */
 static void
-bt_downlink_check(BtreeCheckState *state, BlockNumber childblock,
-				  ScanKey targetkey)
+bt_downlink_check(BtreeCheckState *state, BTScanInsert targetkey,
+				  BlockNumber childblock)
 {
 	OffsetNumber offset;
 	OffsetNumber maxoffset;
@@ -1422,8 +1423,7 @@ bt_downlink_check(BtreeCheckState *state, BlockNumber childblock,
 		if (offset_is_negative_infinity(copaque, offset))
 			continue;
 
-		if (!invariant_leq_nontarget_offset(state, child,
-											targetkey, offset))
+		if (!invariant_leq_nontarget_offset(state, targetkey, child, offset))
 			ereport(ERROR,
 					(errcode(ERRCODE_INDEX_CORRUPTED),
 					 errmsg("down-link lower bound invariant violated for index \"%s\"",
@@ -1863,13 +1863,12 @@ offset_is_negative_infinity(BTPageOpaque opaque, OffsetNumber offset)
  * to corruption.
  */
 static inline bool
-invariant_leq_offset(BtreeCheckState *state, ScanKey key,
+invariant_leq_offset(BtreeCheckState *state, BTScanInsert key,
 					 OffsetNumber upperbound)
 {
-	int16		nkeyatts = IndexRelationGetNumberOfKeyAttributes(state->rel);
 	int32		cmp;
 
-	cmp = _bt_compare(state->rel, nkeyatts, key, state->target, upperbound);
+	cmp = _bt_compare(state->rel, key, state->target, upperbound);
 
 	return cmp <= 0;
 }
@@ -1882,13 +1881,12 @@ invariant_leq_offset(BtreeCheckState *state, ScanKey key,
  * to corruption.
  */
 static inline bool
-invariant_geq_offset(BtreeCheckState *state, ScanKey key,
+invariant_geq_offset(BtreeCheckState *state, BTScanInsert key,
 					 OffsetNumber lowerbound)
 {
-	int16		nkeyatts = IndexRelationGetNumberOfKeyAttributes(state->rel);
 	int32		cmp;
 
-	cmp = _bt_compare(state->rel, nkeyatts, key, state->target, lowerbound);
+	cmp = _bt_compare(state->rel, key, state->target, lowerbound);
 
 	return cmp >= 0;
 }
@@ -1904,14 +1902,12 @@ invariant_geq_offset(BtreeCheckState *state, ScanKey key,
  * to corruption.
  */
 static inline bool
-invariant_leq_nontarget_offset(BtreeCheckState *state,
-							   Page nontarget, ScanKey key,
-							   OffsetNumber upperbound)
+invariant_leq_nontarget_offset(BtreeCheckState *state, BTScanInsert key,
+							   Page nontarget, OffsetNumber upperbound)
 {
-	int16		nkeyatts = IndexRelationGetNumberOfKeyAttributes(state->rel);
 	int32		cmp;
 
-	cmp = _bt_compare(state->rel, nkeyatts, key, nontarget, upperbound);
+	cmp = _bt_compare(state->rel, key, nontarget, upperbound);
 
 	return cmp <= 0;
 }
diff --git a/src/backend/access/nbtree/nbtinsert.c b/src/backend/access/nbtree/nbtinsert.c
index 2b18028823..f143ea8be2 100644
--- a/src/backend/access/nbtree/nbtinsert.c
+++ b/src/backend/access/nbtree/nbtinsert.c
@@ -51,19 +51,19 @@ typedef struct
 
 static Buffer _bt_newroot(Relation rel, Buffer lbuf, Buffer rbuf);
 
-static TransactionId _bt_check_unique(Relation rel, IndexTuple itup,
-				 Relation heapRel, Buffer buf, OffsetNumber offset,
-				 ScanKey itup_scankey,
+static TransactionId _bt_check_unique(Relation rel, BTScanInsert itup_key,
+				 IndexTuple itup, Relation heapRel, Buffer buf,
 				 IndexUniqueCheck checkUnique, bool *is_unique,
 				 uint32 *speculativeToken);
-static void _bt_findinsertloc(Relation rel,
+static OffsetNumber _bt_findinsertloc(Relation rel,
+				  BTScanInsert itup_key,
 				  Buffer *bufptr,
-				  OffsetNumber *offsetptr,
-				  int keysz,
-				  ScanKey scankey,
+				  bool checkingunique,
 				  IndexTuple newtup,
 				  BTStack stack,
 				  Relation heapRel);
+static bool _bt_useduplicatepage(Relation rel, Relation heapRel, Buffer buf,
+					 bool *restorebinsrch, Size itemsz);
 static void _bt_insertonpg(Relation rel, Buffer buf, Buffer cbuf,
 			   BTStack stack,
 			   IndexTuple itup,
@@ -83,8 +83,8 @@ static void _bt_checksplitloc(FindSplitData *state,
 				  int dataitemstoleft, Size firstoldonrightsz);
 static bool _bt_pgaddtup(Page page, Size itemsize, IndexTuple itup,
 			 OffsetNumber itup_off);
-static bool _bt_isequal(TupleDesc itupdesc, Page page, OffsetNumber offnum,
-			int keysz, ScanKey scankey);
+static bool _bt_isequal(TupleDesc itupdesc, BTScanInsert itup_key,
+			Page page, OffsetNumber offnum);
 static void _bt_vacuum_one_page(Relation rel, Buffer buffer, Relation heapRel);
 
 /*
@@ -97,7 +97,9 @@ static void _bt_vacuum_one_page(Relation rel, Buffer buffer, Relation heapRel);
  *		will allow duplicates.  Otherwise (UNIQUE_CHECK_YES or
  *		UNIQUE_CHECK_EXISTING) it will throw error for a duplicate.
  *		For UNIQUE_CHECK_EXISTING we merely run the duplicate check, and
- *		don't actually insert.
+ *		don't actually insert.  If rel is a unique index, then every call
+ *		here is a checkingunique call (i.e. every call does a duplicate
+ *		check, though perhaps only a tentative check).
  *
  *		The result value is only significant for UNIQUE_CHECK_PARTIAL:
  *		it must be true if the entry is known unique, else false.
@@ -110,18 +112,14 @@ _bt_doinsert(Relation rel, IndexTuple itup,
 			 IndexUniqueCheck checkUnique, Relation heapRel)
 {
 	bool		is_unique = false;
-	int			indnkeyatts;
-	ScanKey		itup_scankey;
+	BTScanInsert itup_key;
 	BTStack		stack = NULL;
 	Buffer		buf;
-	OffsetNumber offset;
 	bool		fastpath;
-
-	indnkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
-	Assert(indnkeyatts != 0);
+	bool		checkingunique = (checkUnique != UNIQUE_CHECK_NO);
 
 	/* we need an insertion scan key to do our search, so build one */
-	itup_scankey = _bt_mkscankey(rel, itup);
+	itup_key = _bt_mkscankey(rel, itup);
 
 	/*
 	 * It's very common to have an index on an auto-incremented or
@@ -144,7 +142,6 @@ _bt_doinsert(Relation rel, IndexTuple itup,
 	 */
 top:
 	fastpath = false;
-	offset = InvalidOffsetNumber;
 	if (RelationGetTargetBlock(rel) != InvalidBlockNumber)
 	{
 		Size		itemsz;
@@ -179,8 +176,7 @@ top:
 				!P_IGNORE(lpageop) &&
 				(PageGetFreeSpace(page) > itemsz) &&
 				PageGetMaxOffsetNumber(page) >= P_FIRSTDATAKEY(lpageop) &&
-				_bt_compare(rel, indnkeyatts, itup_scankey, page,
-							P_FIRSTDATAKEY(lpageop)) > 0)
+				_bt_compare(rel, itup_key, page, P_FIRSTDATAKEY(lpageop)) > 0)
 			{
 				/*
 				 * The right-most block should never have an incomplete split.
@@ -219,8 +215,7 @@ top:
 		 * Find the first page containing this key.  Buffer returned by
 		 * _bt_search() is locked in exclusive mode.
 		 */
-		stack = _bt_search(rel, indnkeyatts, itup_scankey, false, &buf, BT_WRITE,
-						   NULL);
+		stack = _bt_search(rel, itup_key, &buf, BT_WRITE, NULL);
 	}
 
 	/*
@@ -244,13 +239,12 @@ top:
 	 * let the tuple in and return false for possibly non-unique, or true for
 	 * definitely unique.
 	 */
-	if (checkUnique != UNIQUE_CHECK_NO)
+	if (checkingunique)
 	{
 		TransactionId xwait;
 		uint32		speculativeToken;
 
-		offset = _bt_binsrch(rel, buf, indnkeyatts, itup_scankey, false);
-		xwait = _bt_check_unique(rel, itup, heapRel, buf, offset, itup_scankey,
+		xwait = _bt_check_unique(rel, itup_key, itup, heapRel, buf,
 								 checkUnique, &is_unique, &speculativeToken);
 
 		if (TransactionIdIsValid(xwait))
@@ -277,6 +271,8 @@ top:
 
 	if (checkUnique != UNIQUE_CHECK_EXISTING)
 	{
+		OffsetNumber newitemoff;
+
 		/*
 		 * The only conflict predicate locking cares about for indexes is when
 		 * an index tuple insert conflicts with an existing lock.  Since the
@@ -287,10 +283,16 @@ top:
 		 * attributes are not considered part of the key space.
 		 */
 		CheckForSerializableConflictIn(rel, NULL, buf);
-		/* do the insertion */
-		_bt_findinsertloc(rel, &buf, &offset, indnkeyatts, itup_scankey, itup,
-						  stack, heapRel);
-		_bt_insertonpg(rel, buf, InvalidBuffer, stack, itup, offset, false);
+
+		/*
+		 * Do the insertion.  Note that itup_key contains mutable state used
+		 * by _bt_check_unique to help _bt_findinsertloc avoid repeating its
+		 * binary search.  !checkingunique case must start own binary search.
+		 */
+		newitemoff = _bt_findinsertloc(rel, itup_key, &buf, checkingunique,
+									   itup, stack, heapRel);
+		_bt_insertonpg(rel, buf, InvalidBuffer, stack, itup, newitemoff,
+					   false);
 	}
 	else
 	{
@@ -301,7 +303,7 @@ top:
 	/* be tidy */
 	if (stack)
 		_bt_freestack(stack);
-	_bt_freeskey(itup_scankey);
+	pfree(itup_key);
 
 	return is_unique;
 }
@@ -309,9 +311,9 @@ top:
 /*
  *	_bt_check_unique() -- Check for violation of unique index constraint
  *
- * offset points to the first possible item that could conflict. It can
- * also point to end-of-page, which means that the first tuple to check
- * is the first tuple on the next page.
+ * Sets state in itup_key sufficient for later _bt_findinsertloc() call to
+ * reuse most of the work of our initial binary search to find conflicting
+ * tuples.
  *
  * Returns InvalidTransactionId if there is no conflict, else an xact ID
  * we must wait for to see if it commits a conflicting tuple.   If an actual
@@ -326,14 +328,14 @@ top:
  * core code must redo the uniqueness check later.
  */
 static TransactionId
-_bt_check_unique(Relation rel, IndexTuple itup, Relation heapRel,
-				 Buffer buf, OffsetNumber offset, ScanKey itup_scankey,
+_bt_check_unique(Relation rel, BTScanInsert itup_key,
+				 IndexTuple itup, Relation heapRel, Buffer buf,
 				 IndexUniqueCheck checkUnique, bool *is_unique,
 				 uint32 *speculativeToken)
 {
 	TupleDesc	itupdesc = RelationGetDescr(rel);
-	int			indnkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
 	SnapshotData SnapshotDirty;
+	OffsetNumber offset;
 	OffsetNumber maxoff;
 	Page		page;
 	BTPageOpaque opaque;
@@ -349,9 +351,17 @@ _bt_check_unique(Relation rel, IndexTuple itup, Relation heapRel,
 	opaque = (BTPageOpaque) PageGetSpecialPointer(page);
 	maxoff = PageGetMaxOffsetNumber(page);
 
+	/*
+	 * Save binary search bounds.  Note that this is also used within
+	 * _bt_findinsertloc() later.
+	 */
+	itup_key->savebinsrch = true;
+	offset = _bt_binsrch(rel, itup_key, buf);
+
 	/*
 	 * Scan over all equal tuples, looking for live conflicts.
 	 */
+	Assert(itup_key->low == offset);
 	for (;;)
 	{
 		ItemId		curitemid;
@@ -364,6 +374,26 @@ _bt_check_unique(Relation rel, IndexTuple itup, Relation heapRel,
 		 */
 		if (offset <= maxoff)
 		{
+			/*
+			 * Fastpath: _bt_binsrch() search bounds can be used to limit our
+			 * consideration to items that are definitely duplicates in most
+			 * cases (though not when original page is empty, or when initial
+			 * offset is past the end of the original page, which may indicate
+			 * that we'll have to examine a second or subsequent page).
+			 *
+			 * Note that this optimization avoids calling _bt_isequal()
+			 * entirely when there are no duplicates, provided initial offset
+			 * isn't past end of the initial page (and provided page has at
+			 * least one item).
+			 */
+			if (nbuf == InvalidBuffer && offset == itup_key->stricthigh)
+			{
+				Assert(itup_key->low >= P_FIRSTDATAKEY(opaque));
+				Assert(itup_key->low <= itup_key->stricthigh);
+				Assert(!_bt_isequal(itupdesc, itup_key, page, offset));
+				break;
+			}
+
 			curitemid = PageGetItemId(page, offset);
 
 			/*
@@ -378,7 +408,7 @@ _bt_check_unique(Relation rel, IndexTuple itup, Relation heapRel,
 			 * first, so that we didn't actually get to exit any sooner
 			 * anyway. So now we just advance over killed items as quickly as
 			 * we can. We only apply _bt_isequal() when we get to a non-killed
-			 * item or the end of the page.
+			 * item.
 			 */
 			if (!ItemIdIsDead(curitemid))
 			{
@@ -391,7 +421,7 @@ _bt_check_unique(Relation rel, IndexTuple itup, Relation heapRel,
 				 * in real comparison, but only for ordering/finding items on
 				 * pages. - vadim 03/24/97
 				 */
-				if (!_bt_isequal(itupdesc, page, offset, indnkeyatts, itup_scankey))
+				if (!_bt_isequal(itupdesc, itup_key, page, offset))
 					break;		/* we're past all the equal tuples */
 
 				/* okay, we gotta fetch the heap tuple ... */
@@ -552,11 +582,14 @@ _bt_check_unique(Relation rel, IndexTuple itup, Relation heapRel,
 			offset = OffsetNumberNext(offset);
 		else
 		{
+			int			highkeycmp;
+
 			/* If scankey == hikey we gotta check the next page too */
 			if (P_RIGHTMOST(opaque))
 				break;
-			if (!_bt_isequal(itupdesc, page, P_HIKEY,
-							 indnkeyatts, itup_scankey))
+			highkeycmp = _bt_compare(rel, itup_key, page, P_HIKEY);
+			Assert(highkeycmp <= 0);
+			if (highkeycmp != 0)
 				break;
 			/* Advance to next non-dead page --- there must be one */
 			for (;;)
@@ -611,39 +644,40 @@ _bt_check_unique(Relation rel, IndexTuple itup, Relation heapRel,
  *		Once we have chosen the page to put the key on, we'll insert it before
  *		any existing equal keys because of the way _bt_binsrch() works.
  *
- *		If there's not enough room in the space, we try to make room by
- *		removing any LP_DEAD tuples.
+ *		_bt_check_unique() callers arrange for their insertion scan key to
+ *		save the progress of the last binary search performed.  No additional
+ *		binary search comparisons occur in the common case where there was no
+ *		existing duplicate tuple, though we may occasionally still not be able
+ *		to reuse their work for our own reasons.  Even when there are garbage
+ *		duplicates, very few binary search comparisons will be performed
+ *		without being strictly necessary.
  *
- *		On entry, *bufptr and *offsetptr point to the first legal position
- *		where the new tuple could be inserted.  The caller should hold an
- *		exclusive lock on *bufptr.  *offsetptr can also be set to
- *		InvalidOffsetNumber, in which case the function will search for the
- *		right location within the page if needed.  On exit, they point to the
- *		chosen insert location.  If _bt_findinsertloc decides to move right,
- *		the lock and pin on the original page will be released and the new
- *		page returned to the caller is exclusively locked instead.
+ *		The caller should hold an exclusive lock on *bufptr in all cases.  On
+ *		exit,  bufptr points to the chosen insert location in all cases.  If
+ *		we have to move right, the lock and pin on the original page will be
+ *		released, and the new page returned to the caller is exclusively
+ *		locked instead.  In any case, we return the offset that caller should
+ *		use to insert into the buffer pointed to by bufptr on return.
  *
- *		newtup is the new tuple we're inserting, and scankey is an insertion
- *		type scan key for it.
+ *		This is also where opportunistic microvacuuming of LP_DEAD tuples
+ *		occurs.  It has to happen here, since it may invalidate a
+ *		_bt_check_unique() caller's cached binary search work.
  */
-static void
+static OffsetNumber
 _bt_findinsertloc(Relation rel,
+				  BTScanInsert itup_key,
 				  Buffer *bufptr,
-				  OffsetNumber *offsetptr,
-				  int keysz,
-				  ScanKey scankey,
+				  bool checkingunique,
 				  IndexTuple newtup,
 				  BTStack stack,
 				  Relation heapRel)
 {
 	Buffer		buf = *bufptr;
 	Page		page = BufferGetPage(buf);
+	bool		restorebinsrch = checkingunique;
 	Size		itemsz;
 	BTPageOpaque lpageop;
-	bool		movedright,
-				vacuumed;
 	OffsetNumber newitemoff;
-	OffsetNumber firstlegaloff = *offsetptr;
 
 	lpageop = (BTPageOpaque) PageGetSpecialPointer(page);
 
@@ -672,55 +706,36 @@ _bt_findinsertloc(Relation rel,
 				 errtableconstraint(heapRel,
 									RelationGetRelationName(rel))));
 
-	/*----------
-	 * If we will need to split the page to put the item on this page,
-	 * check whether we can put the tuple somewhere to the right,
-	 * instead.  Keep scanning right until we
-	 *		(a) find a page with enough free space,
-	 *		(b) reach the last page where the tuple can legally go, or
-	 *		(c) get tired of searching.
-	 * (c) is not flippant; it is important because if there are many
-	 * pages' worth of equal keys, it's better to split one of the early
-	 * pages than to scan all the way to the end of the run of equal keys
-	 * on every insert.  We implement "get tired" as a random choice,
-	 * since stopping after scanning a fixed number of pages wouldn't work
-	 * well (we'd never reach the right-hand side of previously split
-	 * pages).  Currently the probability of moving right is set at 0.99,
-	 * which may seem too high to change the behavior much, but it does an
-	 * excellent job of preventing O(N^2) behavior with many equal keys.
-	 *----------
-	 */
-	movedright = false;
-	vacuumed = false;
-	while (PageGetFreeSpace(page) < itemsz)
+	Assert(P_ISLEAF(lpageop) && !P_INCOMPLETE_SPLIT(lpageop));
+	for (;;)
 	{
+		int			cmpval;
 		Buffer		rbuf;
 		BlockNumber rblkno;
 
 		/*
-		 * before considering moving right, see if we can obtain enough space
-		 * by erasing LP_DEAD items
+		 * The checkingunique (restorebinsrch) case may well have established
+		 * bounds within _bt_check_unique()'s binary search that preclude the
+		 * need for a further high key check.  This fastpath isn't used when
+		 * there are no items on the existing page (other than high key), or
+		 * when it looks like the new item belongs last on the page, but it
+		 * might go on a later page instead.
 		 */
-		if (P_ISLEAF(lpageop) && P_HAS_GARBAGE(lpageop))
-		{
-			_bt_vacuum_one_page(rel, buf, heapRel);
+		if (restorebinsrch && itup_key->low <= itup_key->stricthigh &&
+			itup_key->stricthigh <= PageGetMaxOffsetNumber(page))
+			break;
 
-			/*
-			 * remember that we vacuumed this page, because that makes the
-			 * hint supplied by the caller invalid
-			 */
-			vacuumed = true;
-
-			if (PageGetFreeSpace(page) >= itemsz)
-				break;			/* OK, now we have enough space */
-		}
+		if (P_RIGHTMOST(lpageop))
+			break;
+		cmpval = _bt_compare(rel, itup_key, page, P_HIKEY);
 
 		/*
-		 * nope, so check conditions (b) and (c) enumerated above
+		 * May have to handle case where there is a choice of which page to
+		 * place new tuple on, and we must balance space utilization as best
+		 * we can.
 		 */
-		if (P_RIGHTMOST(lpageop) ||
-			_bt_compare(rel, keysz, scankey, page, P_HIKEY) != 0 ||
-			random() <= (MAX_RANDOM_VALUE / 100))
+		if (cmpval != 0 || _bt_useduplicatepage(rel, heapRel, buf,
+												&restorebinsrch, itemsz))
 			break;
 
 		/*
@@ -763,27 +778,98 @@ _bt_findinsertloc(Relation rel,
 		}
 		_bt_relbuf(rel, buf);
 		buf = rbuf;
-		movedright = true;
-		vacuumed = false;
+		restorebinsrch = false;
+	}
+
+	/* Loop should not break until correct page located */
+	Assert(P_RIGHTMOST(lpageop) ||
+		   _bt_compare(rel, itup_key, page, P_HIKEY) <= 0);
+
+	/*
+	 * Perform microvacuuming of the page we're about to insert tuple on if it
+	 * looks like it has LP_DEAD items.  Only microvacuum when it's likely to
+	 * forestall a page split, though.
+	 */
+	if (P_HAS_GARBAGE(lpageop) && PageGetFreeSpace(page) < itemsz)
+	{
+		_bt_vacuum_one_page(rel, buf, heapRel);
+
+		restorebinsrch = false;
 	}
 
 	/*
-	 * Now we are on the right page, so find the insert position. If we moved
-	 * right at all, we know we should insert at the start of the page. If we
-	 * didn't move right, we can use the firstlegaloff hint if the caller
-	 * supplied one, unless we vacuumed the page which might have moved tuples
-	 * around making the hint invalid. If we didn't move right or can't use
-	 * the hint, find the position by searching.
+	 * Reuse binary search bounds established within _bt_check_unique if
+	 * caller is checkingunique caller, and first page locked is also where
+	 * new tuple should be inserted
 	 */
-	if (movedright)
-		newitemoff = P_FIRSTDATAKEY(lpageop);
-	else if (firstlegaloff != InvalidOffsetNumber && !vacuumed)
-		newitemoff = firstlegaloff;
-	else
-		newitemoff = _bt_binsrch(rel, buf, keysz, scankey, false);
+	itup_key->restorebinsrch = restorebinsrch;
+	newitemoff = _bt_binsrch(rel, itup_key, buf);
+	Assert(!itup_key->restorebinsrch);
+	Assert(!restorebinsrch || newitemoff == _bt_binsrch(rel, itup_key, buf));
 
 	*bufptr = buf;
-	*offsetptr = newitemoff;
+	return newitemoff;
+}
+
+/*
+ *	_bt_useduplicatepage() -- Settle for this page of duplicates?
+ *
+ *		This function handles the question of whether or not an insertion
+ *		of a duplicate into a pg_upgrade'd !heapkeyspace index should
+ *		insert on the page contained in buf when a choice must be made.
+ *		Preemptive microvacuuming is performed here when that could allow
+ *		caller to insert on to the page in buf.
+ *
+ *		Returns true if caller should proceed with insert on buf's page.
+ *		Otherwise, caller should move on to the page to the right (caller
+ *		must always be able to still move right following call here).
+ */
+static bool
+_bt_useduplicatepage(Relation rel, Relation heapRel, Buffer buf,
+					 bool *restorebinsrch, Size itemsz)
+{
+	Page		page = BufferGetPage(buf);
+	BTPageOpaque lpageop;
+
+	lpageop = (BTPageOpaque) PageGetSpecialPointer(page);
+	Assert(P_ISLEAF(lpageop) && !P_RIGHTMOST(lpageop));
+
+	/* Easy case -- there is space free on this page already */
+	if (PageGetFreeSpace(page) >= itemsz)
+		return true;
+
+	if (P_HAS_GARBAGE(lpageop))
+	{
+		_bt_vacuum_one_page(rel, buf, heapRel);
+
+		*restorebinsrch = false;
+		if (PageGetFreeSpace(page) >= itemsz)
+			return true;		/* OK, now we have enough space */
+	}
+
+	/*----------
+	 * It's now clear that _bt_findinsertloc() caller will need to split
+	 * the page if it is to insert new item on to it.  The choice to move
+	 * right to the next page remains open to it, but we should not search
+	 * for free space exhaustively when there are many pages to look through.
+	 *
+	 *	_bt_findinsertloc() keeps scanning right until it:
+	 *		(a) reaches the last page where the tuple can legally go
+	 *	Or until we:
+	 *		(b) find a page with enough free space, or
+	 *		(c) get tired of searching.
+	 * (c) is not flippant; it is important because if there are many
+	 * pages' worth of equal keys, it's better to split one of the early
+	 * pages than to scan all the way to the end of the run of equal keys
+	 * on every insert.  We implement "get tired" as a random choice,
+	 * since stopping after scanning a fixed number of pages wouldn't work
+	 * well (we'd never reach the right-hand side of previously split
+	 * pages).  The probability of moving right is set at 0.99, which may
+	 * seem too high to change the behavior much, but it does an excellent
+	 * job of preventing O(N^2) behavior with many equal keys.
+	 *----------
+	 */
+	return random() <= (MAX_RANDOM_VALUE / 100);
 }
 
 /*----------
@@ -1189,8 +1275,9 @@ _bt_split(Relation rel, Buffer buf, Buffer cbuf, OffsetNumber firstright,
 	 * If the page we're splitting is not the rightmost page at its level in
 	 * the tree, then the first entry on the page is the high key for the
 	 * page.  We need to copy that to the right half.  Otherwise (meaning the
-	 * rightmost page case), all the items on the right half will be user
-	 * data.
+	 * rightmost page case), all the items on the right half will be user data
+	 * (there is no existing high key that needs to be relocated to the new
+	 * right page).
 	 */
 	rightoff = P_HIKEY;
 
@@ -2310,24 +2397,21 @@ _bt_pgaddtup(Page page,
  * Rule is simple: NOT_NULL not equal NULL, NULL not equal NULL too.
  */
 static bool
-_bt_isequal(TupleDesc itupdesc, Page page, OffsetNumber offnum,
-			int keysz, ScanKey scankey)
+_bt_isequal(TupleDesc itupdesc, BTScanInsert itup_key, Page page,
+			OffsetNumber offnum)
 {
 	IndexTuple	itup;
+	ScanKey		scankey;
 	int			i;
 
-	/* Better be comparing to a leaf item */
+	/* Better be comparing to a non-pivot item */
 	Assert(P_ISLEAF((BTPageOpaque) PageGetSpecialPointer(page)));
+	Assert(offnum >= P_FIRSTDATAKEY((BTPageOpaque) PageGetSpecialPointer(page)));
 
+	scankey = itup_key->scankeys;
 	itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offnum));
 
-	/*
-	 * It's okay that we might perform a comparison against a truncated page
-	 * high key when caller needs to determine if _bt_check_unique scan must
-	 * continue on to the next page.  Caller never asks us to compare non-key
-	 * attributes within an INCLUDE index.
-	 */
-	for (i = 1; i <= keysz; i++)
+	for (i = 1; i <= itup_key->keysz; i++)
 	{
 		AttrNumber	attno;
 		Datum		datum;
diff --git a/src/backend/access/nbtree/nbtpage.c b/src/backend/access/nbtree/nbtpage.c
index 9c785bca95..56041c3d38 100644
--- a/src/backend/access/nbtree/nbtpage.c
+++ b/src/backend/access/nbtree/nbtpage.c
@@ -1371,7 +1371,7 @@ _bt_pagedel(Relation rel, Buffer buf)
 			 */
 			if (!stack)
 			{
-				ScanKey		itup_scankey;
+				BTScanInsert itup_key;
 				ItemId		itemid;
 				IndexTuple	targetkey;
 				Buffer		lbuf;
@@ -1421,12 +1421,10 @@ _bt_pagedel(Relation rel, Buffer buf)
 				}
 
 				/* we need an insertion scan key for the search, so build one */
-				itup_scankey = _bt_mkscankey(rel, targetkey);
-				/* find the leftmost leaf page containing this key */
-				stack = _bt_search(rel,
-								   IndexRelationGetNumberOfKeyAttributes(rel),
-								   itup_scankey, false, &lbuf, BT_READ, NULL);
-				/* don't need a pin on the page */
+				itup_key = _bt_mkscankey(rel, targetkey);
+				/* get stack to leaf page by searching index */
+				stack = _bt_search(rel, itup_key, &lbuf, BT_READ, NULL);
+				/* don't need a lock or second pin on the page */
 				_bt_relbuf(rel, lbuf);
 
 				/*
diff --git a/src/backend/access/nbtree/nbtsearch.c b/src/backend/access/nbtree/nbtsearch.c
index 92832237a8..7940297305 100644
--- a/src/backend/access/nbtree/nbtsearch.c
+++ b/src/backend/access/nbtree/nbtsearch.c
@@ -71,13 +71,9 @@ _bt_drop_lock_and_maybe_pin(IndexScanDesc scan, BTScanPos sp)
  *	_bt_search() -- Search the tree for a particular scankey,
  *		or more precisely for the first leaf page it could be on.
  *
- * The passed scankey must be an insertion-type scankey (see nbtree/README),
+ * The passed scankey is an insertion-type scankey (see nbtree/README),
  * but it can omit the rightmost column(s) of the index.
  *
- * When nextkey is false (the usual case), we are looking for the first
- * item >= scankey.  When nextkey is true, we are looking for the first
- * item strictly greater than scankey.
- *
  * Return value is a stack of parent-page pointers.  *bufP is set to the
  * address of the leaf-page buffer, which is read-locked and pinned.
  * No locks are held on the parent pages, however!
@@ -93,8 +89,8 @@ _bt_drop_lock_and_maybe_pin(IndexScanDesc scan, BTScanPos sp)
  * during the search will be finished.
  */
 BTStack
-_bt_search(Relation rel, int keysz, ScanKey scankey, bool nextkey,
-		   Buffer *bufP, int access, Snapshot snapshot)
+_bt_search(Relation rel, BTScanInsert key, Buffer *bufP, int access,
+		   Snapshot snapshot)
 {
 	BTStack		stack_in = NULL;
 	int			page_access = BT_READ;
@@ -130,8 +126,7 @@ _bt_search(Relation rel, int keysz, ScanKey scankey, bool nextkey,
 		 * if the leaf page is split and we insert to the parent page).  But
 		 * this is a good opportunity to finish splits of internal pages too.
 		 */
-		*bufP = _bt_moveright(rel, *bufP, keysz, scankey, nextkey,
-							  (access == BT_WRITE), stack_in,
+		*bufP = _bt_moveright(rel, key, *bufP, (access == BT_WRITE), stack_in,
 							  page_access, snapshot);
 
 		/* if this is a leaf page, we're done */
@@ -144,7 +139,7 @@ _bt_search(Relation rel, int keysz, ScanKey scankey, bool nextkey,
 		 * Find the appropriate item on the internal page, and get the child
 		 * page that it points to.
 		 */
-		offnum = _bt_binsrch(rel, *bufP, keysz, scankey, nextkey);
+		offnum = _bt_binsrch(rel, key, *bufP);
 		itemid = PageGetItemId(page, offnum);
 		itup = (IndexTuple) PageGetItem(page, itemid);
 		blkno = BTreeInnerTupleGetDownLink(itup);
@@ -198,8 +193,8 @@ _bt_search(Relation rel, int keysz, ScanKey scankey, bool nextkey,
 		 * need to move right in the tree.  See Lehman and Yao for an
 		 * excruciatingly precise description.
 		 */
-		*bufP = _bt_moveright(rel, *bufP, keysz, scankey, nextkey,
-							  true, stack_in, BT_WRITE, snapshot);
+		*bufP = _bt_moveright(rel, key, *bufP, true, stack_in, BT_WRITE,
+							  snapshot);
 	}
 
 	return stack_in;
@@ -215,16 +210,17 @@ _bt_search(Relation rel, int keysz, ScanKey scankey, bool nextkey,
  * or strictly to the right of it.
  *
  * This routine decides whether or not we need to move right in the
- * tree by examining the high key entry on the page.  If that entry
- * is strictly less than the scankey, or <= the scankey in the nextkey=true
- * case, then we followed the wrong link and we need to move right.
+ * tree by examining the high key entry on the page.  If that entry is
+ * strictly less than the scankey, or <= the scankey in the
+ * key.nextkey=true case, then we followed the wrong link and we need
+ * to move right.
  *
- * The passed scankey must be an insertion-type scankey (see nbtree/README),
- * but it can omit the rightmost column(s) of the index.
+ * The passed insertion-type scankey can omit the rightmost column(s) of the
+ * index. (see nbtree/README)
  *
- * When nextkey is false (the usual case), we are looking for the first
- * item >= scankey.  When nextkey is true, we are looking for the first
- * item strictly greater than scankey.
+ * When key.nextkey is false (the usual case), we are looking for the first
+ * item >= key.  When key.nextkey is true, we are looking for the first item
+ * strictly greater than key.
  *
  * If forupdate is true, we will attempt to finish any incomplete splits
  * that we encounter.  This is required when locking a target page for an
@@ -241,10 +237,8 @@ _bt_search(Relation rel, int keysz, ScanKey scankey, bool nextkey,
  */
 Buffer
 _bt_moveright(Relation rel,
+			  BTScanInsert key,
 			  Buffer buf,
-			  int keysz,
-			  ScanKey scankey,
-			  bool nextkey,
 			  bool forupdate,
 			  BTStack stack,
 			  int access,
@@ -269,7 +263,7 @@ _bt_moveright(Relation rel,
 	 * We also have to move right if we followed a link that brought us to a
 	 * dead page.
 	 */
-	cmpval = nextkey ? 0 : 1;
+	cmpval = key->nextkey ? 0 : 1;
 
 	for (;;)
 	{
@@ -304,7 +298,7 @@ _bt_moveright(Relation rel,
 			continue;
 		}
 
-		if (P_IGNORE(opaque) || _bt_compare(rel, keysz, scankey, page, P_HIKEY) >= cmpval)
+		if (P_IGNORE(opaque) || _bt_compare(rel, key, page, P_HIKEY) >= cmpval)
 		{
 			/* step right one page */
 			buf = _bt_relandgetbuf(rel, buf, opaque->btpo_next, access);
@@ -324,13 +318,6 @@ _bt_moveright(Relation rel,
 /*
  *	_bt_binsrch() -- Do a binary search for a key on a particular page.
  *
- * The passed scankey must be an insertion-type scankey (see nbtree/README),
- * but it can omit the rightmost column(s) of the index.
- *
- * When nextkey is false (the usual case), we are looking for the first
- * item >= scankey.  When nextkey is true, we are looking for the first
- * item strictly greater than scankey.
- *
  * On a leaf page, _bt_binsrch() returns the OffsetNumber of the first
  * key >= given scankey, or > scankey if nextkey is true.  (NOTE: in
  * particular, this means it is possible to return a value 1 greater than the
@@ -346,27 +333,45 @@ _bt_moveright(Relation rel,
  *
  * This procedure is not responsible for walking right, it just examines
  * the given page.  _bt_binsrch() has no lock or refcount side effects
- * on the buffer.
+ * on the buffer.  When key.savebinsrch is set, modifies mutable fields
+ * of insertion scan key, so that a subsequent call where caller sets
+ * key.savebinsrch can reuse the low and strict high bound of original
+ * binary search.  Callers that use these fields directly must be
+ * prepared for the case where stricthigh isn't on the same page (it
+ * exceeds maxoff for the page), and the case where there are no items
+ * on the page (high < low).
  */
 OffsetNumber
 _bt_binsrch(Relation rel,
-			Buffer buf,
-			int keysz,
-			ScanKey scankey,
-			bool nextkey)
+			BTScanInsert key,
+			Buffer buf)
 {
 	Page		page;
 	BTPageOpaque opaque;
 	OffsetNumber low,
-				high;
+				high,
+				stricthigh;
 	int32		result,
 				cmpval;
+	bool		isleaf;
 
 	page = BufferGetPage(buf);
 	opaque = (BTPageOpaque) PageGetSpecialPointer(page);
 
-	low = P_FIRSTDATAKEY(opaque);
-	high = PageGetMaxOffsetNumber(page);
+	if (!key->restorebinsrch)
+	{
+		low = P_FIRSTDATAKEY(opaque);
+		high = PageGetMaxOffsetNumber(page);
+		isleaf = P_ISLEAF(opaque);
+	}
+	else
+	{
+		/* Restore result of previous binary search against same page */
+		Assert(P_ISLEAF(opaque));
+		low = key->low;
+		high = key->stricthigh;
+		isleaf = true;
+	}
 
 	/*
 	 * If there are no keys on the page, return the first available slot. Note
@@ -375,8 +380,19 @@ _bt_binsrch(Relation rel,
 	 * This can never happen on an internal page, however, since they are
 	 * never empty (an internal page must have children).
 	 */
-	if (high < low)
+	if (unlikely(high < low))
+	{
+		if (key->savebinsrch)
+		{
+			Assert(isleaf);
+			/* Caller can't use stricthigh */
+			key->low = low;
+			key->stricthigh = high;
+		}
+		key->savebinsrch = false;
+		key->restorebinsrch = false;
 		return low;
+	}
 
 	/*
 	 * Binary search to find the first key on the page >= scan key, or first
@@ -390,9 +406,12 @@ _bt_binsrch(Relation rel,
 	 *
 	 * We can fall out when high == low.
 	 */
-	high++;						/* establish the loop invariant for high */
+	if (!key->restorebinsrch)
+		high++;					/* establish the loop invariant for high */
+	key->restorebinsrch = false;
+	stricthigh = high;			/* high initially strictly higher */
 
-	cmpval = nextkey ? 0 : 1;	/* select comparison value */
+	cmpval = key->nextkey ? 0 : 1;	/* select comparison value */
 
 	while (high > low)
 	{
@@ -400,12 +419,21 @@ _bt_binsrch(Relation rel,
 
 		/* We have low <= mid < high, so mid points at a real slot */
 
-		result = _bt_compare(rel, keysz, scankey, page, mid);
+		result = _bt_compare(rel, key, page, mid);
 
 		if (result >= cmpval)
 			low = mid + 1;
 		else
+		{
 			high = mid;
+
+			/*
+			 * high can only be reused by more restrictive binary search when
+			 * it's known to be strictly greater than the original scankey
+			 */
+			if (result != 0)
+				stricthigh = high;
+		}
 	}
 
 	/*
@@ -415,7 +443,14 @@ _bt_binsrch(Relation rel,
 	 * On a leaf page, we always return the first key >= scan key (resp. >
 	 * scan key), which could be the last slot + 1.
 	 */
-	if (P_ISLEAF(opaque))
+	if (key->savebinsrch)
+	{
+		Assert(isleaf);
+		key->low = low;
+		key->stricthigh = stricthigh;
+		key->savebinsrch = false;
+	}
+	if (isleaf)
 		return low;
 
 	/*
@@ -428,13 +463,8 @@ _bt_binsrch(Relation rel,
 }
 
 /*----------
- *	_bt_compare() -- Compare scankey to a particular tuple on the page.
+ *	_bt_compare() -- Compare insertion-type scankey to tuple on a page.
  *
- * The passed scankey must be an insertion-type scankey (see nbtree/README),
- * but it can omit the rightmost column(s) of the index.
- *
- *	keysz: number of key conditions to be checked (might be less than the
- *		number of index columns!)
  *	page/offnum: location of btree item to be compared to.
  *
  *		This routine returns:
@@ -455,17 +485,17 @@ _bt_binsrch(Relation rel,
  */
 int32
 _bt_compare(Relation rel,
-			int keysz,
-			ScanKey scankey,
+			BTScanInsert key,
 			Page page,
 			OffsetNumber offnum)
 {
 	TupleDesc	itupdesc = RelationGetDescr(rel);
 	BTPageOpaque opaque = (BTPageOpaque) PageGetSpecialPointer(page);
 	IndexTuple	itup;
-	int			i;
+	ScanKey		scankey;
 
 	Assert(_bt_check_natts(rel, page, offnum));
+	Assert(key->keysz <= IndexRelationGetNumberOfKeyAttributes(rel));
 
 	/*
 	 * Force result ">" if target item is first data item on an internal page
@@ -488,7 +518,8 @@ _bt_compare(Relation rel,
 	 * _bt_first).
 	 */
 
-	for (i = 1; i <= keysz; i++)
+	scankey = key->scankeys;
+	for (int i = 1; i <= key->keysz; i++)
 	{
 		Datum		datum;
 		bool		isNull;
@@ -574,8 +605,8 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 	StrategyNumber strat;
 	bool		nextkey;
 	bool		goback;
+	BTScanInsertData inskey;
 	ScanKey		startKeys[INDEX_MAX_KEYS];
-	ScanKeyData scankeys[INDEX_MAX_KEYS];
 	ScanKeyData notnullkeys[INDEX_MAX_KEYS];
 	int			keysCount = 0;
 	int			i;
@@ -821,8 +852,9 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 	/*
 	 * We want to start the scan somewhere within the index.  Set up an
 	 * insertion scankey we can use to search for the boundary point we
-	 * identified above.  The insertion scankey is built in the local
-	 * scankeys[] array, using the keys identified by startKeys[].
+	 * identified above.  The insertion scankey is built using the keys
+	 * identified by startKeys[].  (Remaining insertion scankey fields are
+	 * initialized after initial-positioning strategy is finalized.)
 	 */
 	Assert(keysCount <= INDEX_MAX_KEYS);
 	for (i = 0; i < keysCount; i++)
@@ -850,7 +882,7 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 				_bt_parallel_done(scan);
 				return false;
 			}
-			memcpy(scankeys + i, subkey, sizeof(ScanKeyData));
+			memcpy(inskey.scankeys + i, subkey, sizeof(ScanKeyData));
 
 			/*
 			 * If the row comparison is the last positioning key we accepted,
@@ -882,7 +914,8 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 					if (subkey->sk_flags & SK_ISNULL)
 						break;	/* can't use null keys */
 					Assert(keysCount < INDEX_MAX_KEYS);
-					memcpy(scankeys + keysCount, subkey, sizeof(ScanKeyData));
+					memcpy(inskey.scankeys + keysCount, subkey,
+						   sizeof(ScanKeyData));
 					keysCount++;
 					if (subkey->sk_flags & SK_ROW_END)
 					{
@@ -928,7 +961,7 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 				FmgrInfo   *procinfo;
 
 				procinfo = index_getprocinfo(rel, cur->sk_attno, BTORDER_PROC);
-				ScanKeyEntryInitializeWithInfo(scankeys + i,
+				ScanKeyEntryInitializeWithInfo(inskey.scankeys + i,
 											   cur->sk_flags,
 											   cur->sk_attno,
 											   InvalidStrategy,
@@ -949,7 +982,7 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 					elog(ERROR, "missing support function %d(%u,%u) for attribute %d of index \"%s\"",
 						 BTORDER_PROC, rel->rd_opcintype[i], cur->sk_subtype,
 						 cur->sk_attno, RelationGetRelationName(rel));
-				ScanKeyEntryInitialize(scankeys + i,
+				ScanKeyEntryInitialize(inskey.scankeys + i,
 									   cur->sk_flags,
 									   cur->sk_attno,
 									   InvalidStrategy,
@@ -1052,12 +1085,17 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 			return false;
 	}
 
+	/* Initialize remaining insertion scan key fields */
+	inskey.savebinsrch = inskey.restorebinsrch = false;
+	inskey.low = inskey.stricthigh = InvalidOffsetNumber;
+	inskey.nextkey = nextkey;
+	inskey.keysz = keysCount;
+
 	/*
 	 * Use the manufactured insertion scan key to descend the tree and
 	 * position ourselves on the target leaf page.
 	 */
-	stack = _bt_search(rel, keysCount, scankeys, nextkey, &buf, BT_READ,
-					   scan->xs_snapshot);
+	stack = _bt_search(rel, &inskey, &buf, BT_READ, scan->xs_snapshot);
 
 	/* don't need to keep the stack around... */
 	_bt_freestack(stack);
@@ -1086,7 +1124,7 @@ _bt_first(IndexScanDesc scan, ScanDirection dir)
 	_bt_initialize_more_data(so, dir);
 
 	/* position to the precise item on the page */
-	offnum = _bt_binsrch(rel, buf, keysCount, scankeys, nextkey);
+	offnum = _bt_binsrch(rel, &inskey, buf);
 
 	/*
 	 * If nextkey = false, we are positioned at the first item >= scan key, or
diff --git a/src/backend/access/nbtree/nbtsort.c b/src/backend/access/nbtree/nbtsort.c
index dc398e1186..759859c302 100644
--- a/src/backend/access/nbtree/nbtsort.c
+++ b/src/backend/access/nbtree/nbtsort.c
@@ -254,6 +254,7 @@ typedef struct BTWriteState
 {
 	Relation	heap;
 	Relation	index;
+	BTScanInsert inskey;		/* generic insertion scankey */
 	bool		btws_use_wal;	/* dump pages to WAL? */
 	BlockNumber btws_pages_alloced; /* # pages allocated */
 	BlockNumber btws_pages_written; /* # pages written out */
@@ -531,6 +532,7 @@ _bt_leafbuild(BTSpool *btspool, BTSpool *btspool2)
 
 	wstate.heap = btspool->heap;
 	wstate.index = btspool->index;
+	wstate.inskey = _bt_mkscankey(wstate.index, NULL);
 
 	/*
 	 * We need to log index creation in WAL iff WAL archiving/streaming is
@@ -1076,7 +1078,6 @@ _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
 	TupleDesc	tupdes = RelationGetDescr(wstate->index);
 	int			i,
 				keysz = IndexRelationGetNumberOfKeyAttributes(wstate->index);
-	ScanKey		indexScanKey = NULL;
 	SortSupport sortKeys;
 
 	if (merge)
@@ -1089,7 +1090,6 @@ _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
 		/* the preparation of merge */
 		itup = tuplesort_getindextuple(btspool->sortstate, true);
 		itup2 = tuplesort_getindextuple(btspool2->sortstate, true);
-		indexScanKey = _bt_mkscankey_nodata(wstate->index);
 
 		/* Prepare SortSupport data for each column */
 		sortKeys = (SortSupport) palloc0(keysz * sizeof(SortSupportData));
@@ -1097,7 +1097,7 @@ _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
 		for (i = 0; i < keysz; i++)
 		{
 			SortSupport sortKey = sortKeys + i;
-			ScanKey		scanKey = indexScanKey + i;
+			ScanKey		scanKey = wstate->inskey->scankeys + i;
 			int16		strategy;
 
 			sortKey->ssup_cxt = CurrentMemoryContext;
@@ -1116,8 +1116,6 @@ _bt_load(BTWriteState *wstate, BTSpool *btspool, BTSpool *btspool2)
 			PrepareSortSupportFromIndexRel(wstate->index, strategy, sortKey);
 		}
 
-		_bt_freeskey(indexScanKey);
-
 		for (;;)
 		{
 			load1 = true;		/* load BTSpool next ? */
diff --git a/src/backend/access/nbtree/nbtutils.c b/src/backend/access/nbtree/nbtutils.c
index 2c05fb5e45..e010bcdcfa 100644
--- a/src/backend/access/nbtree/nbtutils.c
+++ b/src/backend/access/nbtree/nbtutils.c
@@ -56,34 +56,39 @@ static bool _bt_check_rowcompare(ScanKey skey,
  *		Build an insertion scan key that contains comparison data from itup
  *		as well as comparator routines appropriate to the key datatypes.
  *
- *		The result is intended for use with _bt_compare().
+ *		Result is intended for use with _bt_compare().  Callers that don't
+ *		need to fill out the insertion scankey arguments (e.g. they use an own
+ *		ad-hoc comparison routine) can pass a NULL index tuple.
  */
-ScanKey
+BTScanInsert
 _bt_mkscankey(Relation rel, IndexTuple itup)
 {
+	BTScanInsert key;
 	ScanKey		skey;
 	TupleDesc	itupdesc;
-	int			indnatts PG_USED_FOR_ASSERTS_ONLY;
 	int			indnkeyatts;
 	int16	   *indoption;
+	int			tupnatts;
 	int			i;
 
 	itupdesc = RelationGetDescr(rel);
-	indnatts = IndexRelationGetNumberOfAttributes(rel);
 	indnkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
 	indoption = rel->rd_indoption;
+	tupnatts = itup ? BTreeTupleGetNAtts(itup, rel) : 0;
 
-	Assert(indnkeyatts > 0);
-	Assert(indnkeyatts <= indnatts);
-	Assert(BTreeTupleGetNAtts(itup, rel) == indnatts ||
-		   BTreeTupleGetNAtts(itup, rel) == indnkeyatts);
+	Assert(tupnatts <= IndexRelationGetNumberOfAttributes(rel));
 
 	/*
 	 * We'll execute search using scan key constructed on key columns. Non-key
 	 * (INCLUDE index) columns are always omitted from scan keys.
 	 */
-	skey = (ScanKey) palloc(indnkeyatts * sizeof(ScanKeyData));
-
+	key = palloc(offsetof(BTScanInsertData, scankeys) +
+				 sizeof(ScanKeyData) * indnkeyatts);
+	key->savebinsrch = key->restorebinsrch = false;
+	key->low = key->stricthigh = InvalidOffsetNumber;
+	key->nextkey = false;
+	key->keysz = Min(indnkeyatts, tupnatts);
+	skey = key->scankeys;
 	for (i = 0; i < indnkeyatts; i++)
 	{
 		FmgrInfo   *procinfo;
@@ -96,7 +101,19 @@ _bt_mkscankey(Relation rel, IndexTuple itup)
 		 * comparison can be needed.
 		 */
 		procinfo = index_getprocinfo(rel, i + 1, BTORDER_PROC);
-		arg = index_getattr(itup, i + 1, itupdesc, &null);
+
+		/*
+		 * Key arguments built when caller provides no tuple are defensively
+		 * represented as NULL values, though they should still not
+		 * participate in comparisons.
+		 */
+		if (i < tupnatts)
+			arg = index_getattr(itup, i + 1, itupdesc, &null);
+		else
+		{
+			arg = (Datum) 0;
+			null = true;
+		}
 		flags = (null ? SK_ISNULL : 0) | (indoption[i] << SK_BT_INDOPTION_SHIFT);
 		ScanKeyEntryInitializeWithInfo(&skey[i],
 									   flags,
@@ -108,64 +125,7 @@ _bt_mkscankey(Relation rel, IndexTuple itup)
 									   arg);
 	}
 
-	return skey;
-}
-
-/*
- * _bt_mkscankey_nodata
- *		Build an insertion scan key that contains 3-way comparator routines
- *		appropriate to the key datatypes, but no comparison data.  The
- *		comparison data ultimately used must match the key datatypes.
- *
- *		The result cannot be used with _bt_compare(), unless comparison
- *		data is first stored into the key entries.  Currently this
- *		routine is only called by nbtsort.c and tuplesort.c, which have
- *		their own comparison routines.
- */
-ScanKey
-_bt_mkscankey_nodata(Relation rel)
-{
-	ScanKey		skey;
-	int			indnkeyatts;
-	int16	   *indoption;
-	int			i;
-
-	indnkeyatts = IndexRelationGetNumberOfKeyAttributes(rel);
-	indoption = rel->rd_indoption;
-
-	skey = (ScanKey) palloc(indnkeyatts * sizeof(ScanKeyData));
-
-	for (i = 0; i < indnkeyatts; i++)
-	{
-		FmgrInfo   *procinfo;
-		int			flags;
-
-		/*
-		 * We can use the cached (default) support procs since no cross-type
-		 * comparison can be needed.
-		 */
-		procinfo = index_getprocinfo(rel, i + 1, BTORDER_PROC);
-		flags = SK_ISNULL | (indoption[i] << SK_BT_INDOPTION_SHIFT);
-		ScanKeyEntryInitializeWithInfo(&skey[i],
-									   flags,
-									   (AttrNumber) (i + 1),
-									   InvalidStrategy,
-									   InvalidOid,
-									   rel->rd_indcollation[i],
-									   procinfo,
-									   (Datum) 0);
-	}
-
-	return skey;
-}
-
-/*
- * free a scan key made by either _bt_mkscankey or _bt_mkscankey_nodata.
- */
-void
-_bt_freeskey(ScanKey skey)
-{
-	pfree(skey);
+	return key;
 }
 
 /*
diff --git a/src/backend/utils/sort/tuplesort.c b/src/backend/utils/sort/tuplesort.c
index 7b10fd2974..f97a82ae7b 100644
--- a/src/backend/utils/sort/tuplesort.c
+++ b/src/backend/utils/sort/tuplesort.c
@@ -884,7 +884,7 @@ tuplesort_begin_cluster(TupleDesc tupDesc,
 {
 	Tuplesortstate *state = tuplesort_begin_common(workMem, coordinate,
 												   randomAccess);
-	ScanKey		indexScanKey;
+	BTScanInsert indexScanKey;
 	MemoryContext oldcontext;
 	int			i;
 
@@ -919,7 +919,7 @@ tuplesort_begin_cluster(TupleDesc tupDesc,
 
 	state->tupDesc = tupDesc;	/* assume we need not copy tupDesc */
 
-	indexScanKey = _bt_mkscankey_nodata(indexRel);
+	indexScanKey = _bt_mkscankey(indexRel, NULL);
 
 	if (state->indexInfo->ii_Expressions != NULL)
 	{
@@ -945,7 +945,7 @@ tuplesort_begin_cluster(TupleDesc tupDesc,
 	for (i = 0; i < state->nKeys; i++)
 	{
 		SortSupport sortKey = state->sortKeys + i;
-		ScanKey		scanKey = indexScanKey + i;
+		ScanKey		scanKey = indexScanKey->scankeys + i;
 		int16		strategy;
 
 		sortKey->ssup_cxt = CurrentMemoryContext;
@@ -964,7 +964,7 @@ tuplesort_begin_cluster(TupleDesc tupDesc,
 		PrepareSortSupportFromIndexRel(indexRel, strategy, sortKey);
 	}
 
-	_bt_freeskey(indexScanKey);
+	pfree(indexScanKey);
 
 	MemoryContextSwitchTo(oldcontext);
 
@@ -981,7 +981,7 @@ tuplesort_begin_index_btree(Relation heapRel,
 {
 	Tuplesortstate *state = tuplesort_begin_common(workMem, coordinate,
 												   randomAccess);
-	ScanKey		indexScanKey;
+	BTScanInsert indexScanKey;
 	MemoryContext oldcontext;
 	int			i;
 
@@ -1014,7 +1014,7 @@ tuplesort_begin_index_btree(Relation heapRel,
 	state->indexRel = indexRel;
 	state->enforceUnique = enforceUnique;
 
-	indexScanKey = _bt_mkscankey_nodata(indexRel);
+	indexScanKey = _bt_mkscankey(indexRel, NULL);
 
 	/* Prepare SortSupport data for each column */
 	state->sortKeys = (SortSupport) palloc0(state->nKeys *
@@ -1023,7 +1023,7 @@ tuplesort_begin_index_btree(Relation heapRel,
 	for (i = 0; i < state->nKeys; i++)
 	{
 		SortSupport sortKey = state->sortKeys + i;
-		ScanKey		scanKey = indexScanKey + i;
+		ScanKey		scanKey = indexScanKey->scankeys + i;
 		int16		strategy;
 
 		sortKey->ssup_cxt = CurrentMemoryContext;
@@ -1042,7 +1042,7 @@ tuplesort_begin_index_btree(Relation heapRel,
 		PrepareSortSupportFromIndexRel(indexRel, strategy, sortKey);
 	}
 
-	_bt_freeskey(indexScanKey);
+	pfree(indexScanKey);
 
 	MemoryContextSwitchTo(oldcontext);
 
diff --git a/src/include/access/nbtree.h b/src/include/access/nbtree.h
index 60622ea790..950a61958d 100644
--- a/src/include/access/nbtree.h
+++ b/src/include/access/nbtree.h
@@ -319,6 +319,47 @@ typedef struct BTStackData
 
 typedef BTStackData *BTStack;
 
+/*
+ * BTScanInsert is the btree-private state needed to find an initial position
+ * for an indexscan, or to insert new tuples -- an "insertion scankey" (not to
+ * be confused with a search scankey).  It's used to descend a B-Tree using
+ * _bt_search.  For details on its mutable state, see _bt_binsrch and
+ * _bt_findinsertloc.
+ *
+ * When nextkey is false (the usual case), _bt_search and _bt_binsrch will
+ * locate the first item >= scankey.  When nextkey is true, they will locate
+ * the first item > scan key.
+ *
+ * keysz is the number of insertion scankeys present.
+ *
+ * scankeys is an array of scan key entries for attributes that are compared.
+ * During insertion, there must be a scan key for every attribute, but when
+ * starting a regular index scan some can be omitted.  The array is used as a
+ * flexible array member, though it's sized in a way that makes it possible to
+ * use stack allocations.  See nbtree/README for full details.
+ */
+
+typedef struct BTScanInsertData
+{
+	/*
+	 * Mutable state used by _bt_binsrch to inexpensively repeat a binary
+	 * search on the leaf level.  Only used for insertions where
+	 * _bt_check_unique is called.
+	 */
+	bool		savebinsrch;
+	bool		restorebinsrch;
+	OffsetNumber low;
+	OffsetNumber stricthigh;
+
+	/* State used to locate a position at the leaf level */
+	bool		nextkey;
+	int			keysz;			/* Size of scankeys */
+	ScanKeyData scankeys[INDEX_MAX_KEYS];	/* Must appear last */
+} BTScanInsertData;
+
+typedef BTScanInsertData *BTScanInsert;
+
+
 /*
  * BTScanOpaqueData is the btree-private state needed for an indexscan.
  * This consists of preprocessed scan keys (see _bt_preprocess_keys() for
@@ -558,16 +599,12 @@ extern int	_bt_pagedel(Relation rel, Buffer buf);
 /*
  * prototypes for functions in nbtsearch.c
  */
-extern BTStack _bt_search(Relation rel,
-		   int keysz, ScanKey scankey, bool nextkey,
-		   Buffer *bufP, int access, Snapshot snapshot);
-extern Buffer _bt_moveright(Relation rel, Buffer buf, int keysz,
-			  ScanKey scankey, bool nextkey, bool forupdate, BTStack stack,
-			  int access, Snapshot snapshot);
-extern OffsetNumber _bt_binsrch(Relation rel, Buffer buf, int keysz,
-			ScanKey scankey, bool nextkey);
-extern int32 _bt_compare(Relation rel, int keysz, ScanKey scankey,
-			Page page, OffsetNumber offnum);
+extern BTStack _bt_search(Relation rel, BTScanInsert key, Buffer *bufP,
+		   int access, Snapshot snapshot);
+extern Buffer _bt_moveright(Relation rel, BTScanInsert key, Buffer buf,
+			  bool forupdate, BTStack stack, int access, Snapshot snapshot);
+extern OffsetNumber _bt_binsrch(Relation rel, BTScanInsert key, Buffer buf);
+extern int32 _bt_compare(Relation rel, BTScanInsert key, Page page, OffsetNumber offnum);
 extern bool _bt_first(IndexScanDesc scan, ScanDirection dir);
 extern bool _bt_next(IndexScanDesc scan, ScanDirection dir);
 extern Buffer _bt_get_endpoint(Relation rel, uint32 level, bool rightmost,
@@ -576,9 +613,7 @@ extern Buffer _bt_get_endpoint(Relation rel, uint32 level, bool rightmost,
 /*
  * prototypes for functions in nbtutils.c
  */
-extern ScanKey _bt_mkscankey(Relation rel, IndexTuple itup);
-extern ScanKey _bt_mkscankey_nodata(Relation rel);
-extern void _bt_freeskey(ScanKey skey);
+extern BTScanInsert _bt_mkscankey(Relation rel, IndexTuple itup);
 extern void _bt_freestack(BTStack stack);
 extern void _bt_preprocess_array_keys(IndexScanDesc scan);
 extern void _bt_start_array_keys(IndexScanDesc scan, ScanDirection dir);
-- 
2.17.1