Get rid of the post-recovery cleanup step of GIN page splits.

Replace it with an approach similar to what GiST uses: when a page is split,
the left sibling is marked with a flag indicating that the parent hasn't been
updated yet. When the parent is updated, the flag is cleared. If an insertion
steps on a page with the flag set, it will finish split before proceeding
with the insertion.

The post-recovery cleanup mechanism was never totally reliable, as insertion
to the parent could fail e.g because of running out of memory or disk space,
leaving the tree in an inconsistent state.

This also divides the responsibility of WAL-logging more clearly between
the generic ginbtree.c code, and the parts specific to entry and posting
trees. There is now a common WAL record format for insertions and deletions,
which is written by ginbtree.c, followed by tree-specific payload, which is
returned by the placetopage- and split- callbacks.

Author: hlinnaka

src/backend/access/gin/ginbtree.c

@@ -227,6 +227,7 @@ GinDataPageAddItemPointer(Page page, ItemPointer data, OffsetNumber offset)
 	OffsetNumber maxoff = GinPageGetOpaque(page)->maxoff;
 	char	   *ptr;
 
+	Assert(ItemPointerIsValid(data));
 	Assert(GinPageIsLeaf(page));
 
 	if (offset == InvalidOffsetNumber)
@@ -255,6 +256,7 @@ GinDataPageAddPostingItem(Page page, PostingItem *data, OffsetNumber offset)
 	OffsetNumber maxoff = GinPageGetOpaque(page)->maxoff;
 	char	   *ptr;
 
+	Assert(PostingItemGetBlockNumber(data) != InvalidBlockNumber);
 	Assert(!GinPageIsLeaf(page));
 
 	if (offset == InvalidOffsetNumber)
@@ -338,11 +340,8 @@ dataPlaceToPage(GinBtree btree, Buffer buf, OffsetNumber off,
 				XLogRecData **prdata)
 {
 	Page		page = BufferGetPage(buf);
-	int			cnt = 0;
-
 	/* these must be static so they can be returned to caller */
-	static XLogRecData rdata[3];
-	static ginxlogInsert data;
+	static XLogRecData rdata[2];
 
 	/* quick exit if it doesn't fit */
 	if (!dataIsEnoughSpace(btree, buf, off, insertdata))
@@ -359,45 +358,10 @@ dataPlaceToPage(GinBtree btree, Buffer buf, OffsetNumber off,
 		PostingItemSetBlockNumber(pitem, updateblkno);
 	}
 
-	data.updateBlkno = updateblkno;
-	data.node = btree->index->rd_node;
-	data.blkno = BufferGetBlockNumber(buf);
-	data.offset = off;
-	data.nitem = 1;
-	data.isDelete = FALSE;
-	data.isData = TRUE;
-	data.isLeaf = GinPageIsLeaf(page) ? TRUE : FALSE;
-
-	/*
-	 * Prevent full page write if child's split occurs. That is needed to
-	 * remove incomplete splits while replaying WAL
-	 *
-	 * data.updateBlkno contains new block number (of newly created right
-	 * page) for recently splited page.
-	 */
-	if (data.updateBlkno == InvalidBlockNumber)
-	{
-		rdata[0].buffer = buf;
-		rdata[0].buffer_std = FALSE;
-		rdata[0].data = NULL;
-		rdata[0].len = 0;
-		rdata[0].next = &rdata[1];
-		cnt++;
-	}
-
-	rdata[cnt].buffer = InvalidBuffer;
-	rdata[cnt].data = (char *) &data;
-	rdata[cnt].len = sizeof(ginxlogInsert);
-	rdata[cnt].next = &rdata[cnt + 1];
-	cnt++;
-
-	rdata[cnt].buffer = InvalidBuffer;
-	/* data and len filled in below */
-	rdata[cnt].next = NULL;
-
 	if (GinPageIsLeaf(page))
 	{
 		GinBtreeDataLeafInsertData *items = insertdata;
+		static ginxlogInsertDataLeaf data;
 		uint32		savedPos = items->curitem;
 
 		if (GinPageRightMost(page) && off > GinPageGetOpaque(page)->maxoff)
@@ -415,19 +379,29 @@ dataPlaceToPage(GinBtree btree, Buffer buf, OffsetNumber off,
 		{
 			GinDataPageAddItemPointer(page, items->items + items->curitem, off);
 			items->curitem++;
+			data.nitem = 1;
 		}
 
-		rdata[cnt].data = (char *) &items->items[savedPos];
-		rdata[cnt].len = sizeof(ItemPointerData) * data.nitem;
+		rdata[0].buffer = InvalidBuffer;
+		rdata[0].data = (char *) &data;
+		rdata[0].len = offsetof(ginxlogInsertDataLeaf, items);
+		rdata[0].next = &rdata[1];
+
+		rdata[1].buffer = InvalidBuffer;
+		rdata[1].data = (char *) &items->items[savedPos];
+		rdata[1].len = sizeof(ItemPointerData) * data.nitem;
+		rdata[1].next = NULL;
 	}
 	else
 	{
 		PostingItem *pitem = insertdata;
 
 		GinDataPageAddPostingItem(page, pitem, off);
 
-		rdata[cnt].data = (char *) pitem;
-		rdata[cnt].len = sizeof(PostingItem);
+		rdata[0].buffer = InvalidBuffer;
+		rdata[0].data = (char *) pitem;
+		rdata[0].len = sizeof(PostingItem);
+		rdata[0].next = NULL;
 	}
 
 	return true;
@@ -456,8 +430,8 @@ dataSplitPage(GinBtree btree, Buffer lbuf, Buffer rbuf, OffsetNumber off,
 	Size		freeSpace;
 
 	/* these must be static so they can be returned to caller */
-	static ginxlogSplit data;
-	static XLogRecData rdata[4];
+	static ginxlogSplitData data;
+	static XLogRecData rdata[2];
 	static char vector[2 * BLCKSZ];
 
 	GinInitPage(rpage, GinPageGetOpaque(lpage)->flags, pageSize);
@@ -488,6 +462,7 @@ dataSplitPage(GinBtree btree, Buffer lbuf, Buffer rbuf, OffsetNumber off,
 
 	if (isleaf && GinPageRightMost(lpage) && off > GinPageGetOpaque(lpage)->maxoff)
 	{
+		/* append new items to the end */
 		GinBtreeDataLeafInsertData *items = insertdata;
 
 		while (items->curitem < items->nitem &&
@@ -566,25 +541,18 @@ dataSplitPage(GinBtree btree, Buffer lbuf, Buffer rbuf, OffsetNumber off,
 	bound = GinDataPageGetRightBound(rpage);
 	*bound = oldbound;
 
-	data.node = btree->index->rd_node;
-	data.rootBlkno = InvalidBlockNumber;
-	data.lblkno = BufferGetBlockNumber(lbuf);
-	data.rblkno = BufferGetBlockNumber(rbuf);
 	data.separator = separator;
 	data.nitem = maxoff;
-	data.isData = TRUE;
-	data.isLeaf = GinPageIsLeaf(lpage) ? TRUE : FALSE;
-	data.isRootSplit = FALSE;
 	data.rightbound = oldbound;
 
 	rdata[0].buffer = InvalidBuffer;
 	rdata[0].data = (char *) &data;
-	rdata[0].len = sizeof(ginxlogSplit);
+	rdata[0].len = sizeof(ginxlogSplitData);
 	rdata[0].next = &rdata[1];
 
 	rdata[1].buffer = InvalidBuffer;
 	rdata[1].data = vector;
-	rdata[1].len = MAXALIGN(maxoff * sizeofitem);
+	rdata[1].len = maxoff * sizeofitem;
 	rdata[1].next = NULL;
 
 	return lpage;
@@ -610,21 +578,18 @@ dataPrepareDownlink(GinBtree btree, Buffer lbuf)
  * Also called from ginxlog, should not use btree
  */
 void
-ginDataFillRoot(GinBtree btree, Buffer root, Buffer lbuf, Buffer rbuf)
+ginDataFillRoot(GinBtree btree, Page root, BlockNumber lblkno, Page lpage, BlockNumber rblkno, Page rpage)
 {
-	Page		page = BufferGetPage(root),
-				lpage = BufferGetPage(lbuf),
-				rpage = BufferGetPage(rbuf);
 	PostingItem li,
 				ri;
 
 	li.key = *GinDataPageGetRightBound(lpage);
-	PostingItemSetBlockNumber(&li, BufferGetBlockNumber(lbuf));
-	GinDataPageAddPostingItem(page, &li, InvalidOffsetNumber);
+	PostingItemSetBlockNumber(&li, lblkno);
+	GinDataPageAddPostingItem(root, &li, InvalidOffsetNumber);
 
 	ri.key = *GinDataPageGetRightBound(rpage);
-	PostingItemSetBlockNumber(&ri, BufferGetBlockNumber(rbuf));
-	GinDataPageAddPostingItem(page, &ri, InvalidOffsetNumber);
+	PostingItemSetBlockNumber(&ri, rblkno);
+	GinDataPageAddPostingItem(root, &ri, InvalidOffsetNumber);
 }
 
 /*

src/backend/access/gin/gindatapage.c

@@ -504,15 +504,14 @@ entryPlaceToPage(GinBtree btree, Buffer buf, OffsetNumber off,
 
 	/* these must be static so they can be returned to caller */
 	static XLogRecData rdata[3];
-	static ginxlogInsert data;
+	static ginxlogInsertEntry data;
 
 	/* quick exit if it doesn't fit */
 	if (!entryIsEnoughSpace(btree, buf, off, insertData))
 		return false;
 
 	*prdata = rdata;
 	entryPreparePage(btree, page, off, insertData, updateblkno);
-	data.updateBlkno = updateblkno;
 
 	placed = PageAddItem(page,
 						 (Item) insertData->entry,
@@ -522,34 +521,11 @@ entryPlaceToPage(GinBtree btree, Buffer buf, OffsetNumber off,
 		elog(ERROR, "failed to add item to index page in \"%s\"",
 			 RelationGetRelationName(btree->index));
 
-	data.node = btree->index->rd_node;
-	data.blkno = BufferGetBlockNumber(buf);
-	data.offset = off;
-	data.nitem = 1;
 	data.isDelete = insertData->isDelete;
-	data.isData = false;
-	data.isLeaf = GinPageIsLeaf(page) ? TRUE : FALSE;
-
-	/*
-	 * Prevent full page write if child's split occurs. That is needed to
-	 * remove incomplete splits while replaying WAL
-	 *
-	 * data.updateBlkno contains new block number (of newly created right
-	 * page) for recently splited page.
-	 */
-	if (data.updateBlkno == InvalidBlockNumber)
-	{
-		rdata[0].buffer = buf;
-		rdata[0].buffer_std = TRUE;
-		rdata[0].data = NULL;
-		rdata[0].len = 0;
-		rdata[0].next = &rdata[1];
-		cnt++;
-	}
 
 	rdata[cnt].buffer = InvalidBuffer;
 	rdata[cnt].data = (char *) &data;
-	rdata[cnt].len = sizeof(ginxlogInsert);
+	rdata[cnt].len = offsetof(ginxlogInsertEntry, tuple);
 	rdata[cnt].next = &rdata[cnt + 1];
 	cnt++;
 
@@ -577,6 +553,7 @@ entrySplitPage(GinBtree btree, Buffer lbuf, Buffer rbuf, OffsetNumber off,
 				maxoff,
 				separator = InvalidOffsetNumber;
 	Size		totalsize = 0;
+	Size		tupstoresize;
 	Size		lsize = 0,
 				size;
 	char	   *ptr;
@@ -588,18 +565,18 @@ entrySplitPage(GinBtree btree, Buffer lbuf, Buffer rbuf, OffsetNumber off,
 
 	/* these must be static so they can be returned to caller */
 	static XLogRecData rdata[2];
-	static ginxlogSplit data;
+	static ginxlogSplitEntry data;
 	static char tupstore[2 * BLCKSZ];
 
 	*prdata = rdata;
-	data.leftChildBlkno = (GinPageIsLeaf(lpage)) ?
-		InvalidOffsetNumber : GinGetDownlink(insertData->entry);
-	data.updateBlkno = updateblkno;
 	entryPreparePage(btree, lpage, off, insertData, updateblkno);
 
+	/*
+	 * First, append all the existing tuples and the new tuple we're inserting
+	 * one after another in a temporary workspace.
+	 */
 	maxoff = PageGetMaxOffsetNumber(lpage);
 	ptr = tupstore;
-
 	for (i = FirstOffsetNumber; i <= maxoff; i++)
 	{
 		if (i == off)
@@ -624,7 +601,12 @@ entrySplitPage(GinBtree btree, Buffer lbuf, Buffer rbuf, OffsetNumber off,
 		ptr += size;
 		totalsize += size + sizeof(ItemIdData);
 	}
+	tupstoresize = ptr - tupstore;
 
+	/*
+	 * Initialize the left and right pages, and copy all the tuples back to
+	 * them.
+	 */
 	GinInitPage(rpage, GinPageGetOpaque(lpage)->flags, pageSize);
 	GinInitPage(lpage, GinPageGetOpaque(rpage)->flags, pageSize);
 
@@ -654,24 +636,17 @@ entrySplitPage(GinBtree btree, Buffer lbuf, Buffer rbuf, OffsetNumber off,
 		ptr += MAXALIGN(IndexTupleSize(itup));
 	}
 
-	data.node = btree->index->rd_node;
-	data.rootBlkno = InvalidBlockNumber;
-	data.lblkno = BufferGetBlockNumber(lbuf);
-	data.rblkno = BufferGetBlockNumber(rbuf);
 	data.separator = separator;
 	data.nitem = maxoff;
-	data.isData = FALSE;
-	data.isLeaf = GinPageIsLeaf(lpage) ? TRUE : FALSE;
-	data.isRootSplit = FALSE;
 
 	rdata[0].buffer = InvalidBuffer;
 	rdata[0].data = (char *) &data;
-	rdata[0].len = sizeof(ginxlogSplit);
+	rdata[0].len = sizeof(ginxlogSplitEntry);
 	rdata[0].next = &rdata[1];
 
 	rdata[1].buffer = InvalidBuffer;
 	rdata[1].data = tupstore;
-	rdata[1].len = MAXALIGN(totalsize);
+	rdata[1].len = tupstoresize;
 	rdata[1].next = NULL;
 
 	return lpage;
@@ -702,24 +677,19 @@ entryPrepareDownlink(GinBtree btree, Buffer lbuf)
  * Also called from ginxlog, should not use btree
  */
 void
-ginEntryFillRoot(GinBtree btree, Buffer root, Buffer lbuf, Buffer rbuf)
+ginEntryFillRoot(GinBtree btree, Page root,
+				 BlockNumber lblkno, Page lpage,
+				 BlockNumber rblkno, Page rpage)
 {
-	Page		page = BufferGetPage(root);
-	Page		lpage = BufferGetPage(lbuf);
-	Page		rpage = BufferGetPage(rbuf);
 	IndexTuple	itup;
 
-	itup = GinFormInteriorTuple(getRightMostTuple(lpage),
-								lpage,
-								BufferGetBlockNumber(lbuf));
-	if (PageAddItem(page, (Item) itup, IndexTupleSize(itup), InvalidOffsetNumber, false, false) == InvalidOffsetNumber)
+	itup = GinFormInteriorTuple(getRightMostTuple(lpage), lpage, lblkno);
+	if (PageAddItem(root, (Item) itup, IndexTupleSize(itup), InvalidOffsetNumber, false, false) == InvalidOffsetNumber)
 		elog(ERROR, "failed to add item to index root page");
 	pfree(itup);
 
-	itup = GinFormInteriorTuple(getRightMostTuple(rpage),
-								rpage,
-								BufferGetBlockNumber(rbuf));
-	if (PageAddItem(page, (Item) itup, IndexTupleSize(itup), InvalidOffsetNumber, false, false) == InvalidOffsetNumber)
+	itup = GinFormInteriorTuple(getRightMostTuple(rpage), rpage, rblkno);
+	if (PageAddItem(root, (Item) itup, IndexTupleSize(itup), InvalidOffsetNumber, false, false) == InvalidOffsetNumber)
 		elog(ERROR, "failed to add item to index root page");
 	pfree(itup);
 }