Fix SPGiST vacuum algorithm to handle concurrent tuple motion properly.

A leaf tuple that we need to delete could get moved as a consequence of an
insertion happening concurrently with the VACUUM scan.  If it moves from a
page past the current scan point to a page before, we'll miss it, which is
not acceptable.  Hence, when we see a leaf-page REDIRECT that could have
been made since our scan started, chase down the redirection pointer much
as if we were doing a normal index search, and be sure to vacuum every page
it leads to.  This fixes the issue because, if the tuple was on page N at
the instant we start our scan, we will surely find it as a consequence of
chasing the redirect from page N, no matter how much it moves around in
between.  Problem noted by Takashi Yamamoto.

src/backend/access/spgist/README

@@ -314,6 +314,27 @@ the reverse map of the nextOffset links (ie, when we see tuple x links to
 tuple y, we set predecessor[y] = x).  Then head tuples are the ones with
 no predecessor.
 
+Because insertions can occur while VACUUM runs, a pure sequential scan
+could miss deleting some target leaf tuples, because they could get moved
+from a not-yet-visited leaf page to an already-visited leaf page as a
+consequence of a PickSplit or MoveLeafs operation.  Failing to delete any
+target TID is not acceptable, so we have to extend the algorithm to cope
+with such cases.  We recognize that such a move might have occurred when
+we see a leaf-page REDIRECT tuple whose XID indicates it might have been
+created after the VACUUM scan started.  We add the redirection target TID
+to a "pending list" of places we need to recheck.  Between pages of the
+main sequential scan, we empty the pending list by visiting each listed
+TID.  If it points to an inner tuple (from a PickSplit), add each downlink
+TID to the pending list.  If it points to a leaf page, vacuum that page.
+(We could just vacuum the single pointed-to chain, but vacuuming the
+whole page simplifies the code and reduces the odds of VACUUM having to
+modify the same page multiple times.)  To ensure that pending-list
+processing can never get into an endless loop, even in the face of
+concurrent index changes, we don't remove list entries immediately but
+only after we've completed all pending-list processing; instead we just
+mark items as done after processing them.  Adding a TID that's already in
+the list is a no-op, whether or not that item is marked done yet.
+
 spgbulkdelete also updates the index's free space map.
 
 Currently, spgvacuumcleanup has nothing to do if spgbulkdelete was

src/backend/access/spgist/spgvacuum.c

@@ -25,9 +25,18 @@
 #include "storage/indexfsm.h"
 #include "storage/lmgr.h"
 #include "storage/procarray.h"
+#include "utils/snapmgr.h"
 
 
-/* local state for vacuum operations */
+/* Entry in pending-list of TIDs we need to revisit */
+typedef struct spgVacPendingItem
+{
+	ItemPointerData tid;				/* redirection target to visit */
+	bool		done;					/* have we dealt with this? */
+	struct spgVacPendingItem *next;		/* list link */
+} spgVacPendingItem;
+
+/* Local state for vacuum operations */
 typedef struct spgBulkDeleteState
 {
 	/* Parameters passed in to spgvacuumscan */
@@ -35,22 +44,87 @@ typedef struct spgBulkDeleteState
 	IndexBulkDeleteResult *stats;
 	IndexBulkDeleteCallback callback;
 	void	   *callback_state;
+
 	/* Additional working state */
-	SpGistState spgstate;
+	SpGistState spgstate;			/* for SPGiST operations that need one */
-	TransactionId OldestXmin;
+	spgVacPendingItem *pendingList;	/* TIDs we need to (re)visit */
-	BlockNumber lastFilledBlock;
+	TransactionId myXmin;			/* for detecting newly-added redirects */
+	TransactionId OldestXmin;		/* for deciding a redirect is obsolete */
+	BlockNumber lastFilledBlock;	/* last non-deletable block */
 } spgBulkDeleteState;
 
 
+/*
+ * Add TID to pendingList, but only if not already present.
+ *
+ * Note that new items are always appended at the end of the list; this
+ * ensures that scans of the list don't miss items added during the scan.
+ */
+static void
+spgAddPendingTID(spgBulkDeleteState *bds, ItemPointer tid)
+{
+	spgVacPendingItem *pitem;
+	spgVacPendingItem **listLink;
+
+	/* search the list for pre-existing entry */
+	listLink = &bds->pendingList;
+	while (*listLink != NULL)
+	{
+		pitem = *listLink;
+		if (ItemPointerEquals(tid, &pitem->tid))
+			return;				/* already in list, do nothing */
+		listLink = &pitem->next;
+	}
+	/* not there, so append new entry */
+	pitem = (spgVacPendingItem *) palloc(sizeof(spgVacPendingItem));
+	pitem->tid = *tid;
+	pitem->done = false;
+	pitem->next = NULL;
+	*listLink = pitem;
+}
+
+/*
+ * Clear pendingList
+ */
+static void
+spgClearPendingList(spgBulkDeleteState *bds)
+{
+	spgVacPendingItem *pitem;
+	spgVacPendingItem *nitem;
+
+	for (pitem = bds->pendingList; pitem != NULL; pitem = nitem)
+	{
+		nitem = pitem->next;
+		/* All items in list should have been dealt with */
+		Assert(pitem->done);
+		pfree(pitem);
+	}
+	bds->pendingList = NULL;
+}
+
 /*
  * Vacuum a regular (non-root) leaf page
  *
  * We must delete tuples that are targeted for deletion by the VACUUM,
  * but not move any tuples that are referenced by outside links; we assume
  * those are the ones that are heads of chains.
+ *
+ * If we find a REDIRECT that was made by a concurrently-running transaction,
+ * we must add its target TID to pendingList.  (We don't try to visit the
+ * target immediately, first because we don't want VACUUM locking more than
+ * one buffer at a time, and second because the duplicate-filtering logic
+ * in spgAddPendingTID is useful to ensure we can't get caught in an infinite
+ * loop in the face of continuous concurrent insertions.)
+ *
+ * If forPending is true, we are examining the page as a consequence of
+ * chasing a redirect link, not as part of the normal sequential scan.
+ * We still vacuum the page normally, but we don't increment the stats
+ * about live tuples; else we'd double-count those tuples, since the page
+ * has been or will be visited in the sequential scan as well.
  */
 static void
-vacuumLeafPage(spgBulkDeleteState *bds, Relation index, Buffer buffer)
+vacuumLeafPage(spgBulkDeleteState *bds, Relation index, Buffer buffer,
+			   bool forPending)
 {
 	Page		page = BufferGetPage(buffer);
 	spgxlogVacuumLeaf xlrec;
@@ -90,7 +164,8 @@ vacuumLeafPage(spgBulkDeleteState *bds, Relation index, Buffer buffer)
 			}
 			else
 			{
-				bds->stats->num_index_tuples += 1;
+				if (!forPending)
+					bds->stats->num_index_tuples += 1;
 			}
 
 			/* Form predecessor map, too */
@@ -106,6 +181,25 @@ vacuumLeafPage(spgBulkDeleteState *bds, Relation index, Buffer buffer)
 				predecessor[lt->nextOffset] = i;
 			}
 		}
+		else if (lt->tupstate == SPGIST_REDIRECT)
+		{
+			SpGistDeadTuple dt = (SpGistDeadTuple) lt;
+
+			Assert(dt->nextOffset == InvalidOffsetNumber);
+			Assert(ItemPointerIsValid(&dt->pointer));
+
+			/*
+			 * Add target TID to pending list if the redirection could have
+			 * happened since VACUUM started.
+			 *
+			 * Note: we could make a tighter test by seeing if the xid is
+			 * "running" according to the active snapshot; but tqual.c doesn't
+			 * currently export a suitable API, and it's not entirely clear
+			 * that a tighter test is worth the cycles anyway.
+			 */
+			if (TransactionIdFollowsOrEquals(dt->xid, bds->myXmin))
+				spgAddPendingTID(bds, &dt->pointer);
+		}
 		else
 		{
 			Assert(lt->nextOffset == InvalidOffsetNumber);
@@ -545,7 +639,7 @@ spgvacuumpage(spgBulkDeleteState *bds, BlockNumber blkno)
 		}
 		else
 		{
-			vacuumLeafPage(bds, index, buffer);
+			vacuumLeafPage(bds, index, buffer, false);
 			vacuumRedirectAndPlaceholder(index, buffer, bds->OldestXmin);
 		}
 	}
@@ -556,8 +650,8 @@ spgvacuumpage(spgBulkDeleteState *bds, BlockNumber blkno)
 	}
 
 	/*
-	 * The root page must never be deleted, nor marked as available in FSM,
+	 * The root pages must never be deleted, nor marked as available in FSM,
-	 * because we don't want it ever returned by a search for a place to
+	 * because we don't want them ever returned by a search for a place to
 	 * put a new tuple.  Otherwise, check for empty/deletable page, and
 	 * make sure FSM knows about it.
 	 */
@@ -585,6 +679,118 @@ spgvacuumpage(spgBulkDeleteState *bds, BlockNumber blkno)
 	UnlockReleaseBuffer(buffer);
 }
 
+/*
+ * Process the pending-TID list between pages of the main scan
+ */
+static void
+spgprocesspending(spgBulkDeleteState *bds)
+{
+	Relation	index = bds->info->index;
+	spgVacPendingItem *pitem;
+	spgVacPendingItem *nitem;
+	BlockNumber	blkno;
+	Buffer		buffer;
+	Page		page;
+
+	for (pitem = bds->pendingList; pitem != NULL; pitem = pitem->next)
+	{
+		if (pitem->done)
+			continue;			/* ignore already-done items */
+
+		/* call vacuum_delay_point while not holding any buffer lock */
+		vacuum_delay_point();
+
+		/* examine the referenced page */
+		blkno = ItemPointerGetBlockNumber(&pitem->tid);
+		buffer = ReadBufferExtended(index, MAIN_FORKNUM, blkno,
+									RBM_NORMAL, bds->info->strategy);
+		LockBuffer(buffer, BUFFER_LOCK_EXCLUSIVE);
+		page = (Page) BufferGetPage(buffer);
+
+		if (PageIsNew(page) || SpGistPageIsDeleted(page))
+		{
+			/* Probably shouldn't happen, but ignore it */
+		}
+		else if (SpGistPageIsLeaf(page))
+		{
+			if (SpGistBlockIsRoot(blkno))
+			{
+				/* this should definitely not happen */
+				elog(ERROR, "redirection leads to root page of index \"%s\"",
+					 RelationGetRelationName(index));
+			}
+
+			/* deal with any deletable tuples */
+			vacuumLeafPage(bds, index, buffer, true);
+			/* might as well do this while we are here */
+			vacuumRedirectAndPlaceholder(index, buffer, bds->OldestXmin);
+
+			SpGistSetLastUsedPage(index, buffer);
+
+			/*
+			 * We can mark as done not only this item, but any later ones
+			 * pointing at the same page, since we vacuumed the whole page.
+			 */
+			pitem->done = true;
+			for (nitem = pitem->next; nitem != NULL; nitem = nitem->next)
+			{
+				if (ItemPointerGetBlockNumber(&nitem->tid) == blkno)
+					nitem->done = true;
+			}
+		}
+		else
+		{
+			/*
+			 * On an inner page, visit the referenced inner tuple and add
+			 * all its downlinks to the pending list.  We might have pending
+			 * items for more than one inner tuple on the same page (in fact
+			 * this is pretty likely given the way space allocation works),
+			 * so get them all while we are here.
+			 */
+			for (nitem = pitem; nitem != NULL; nitem = nitem->next)
+			{
+				if (nitem->done)
+					continue;
+				if (ItemPointerGetBlockNumber(&nitem->tid) == blkno)
+				{
+					OffsetNumber offset;
+					SpGistInnerTuple innerTuple;
+
+					offset = ItemPointerGetOffsetNumber(&nitem->tid);
+					innerTuple = (SpGistInnerTuple) PageGetItem(page,
+												PageGetItemId(page, offset));
+					if (innerTuple->tupstate == SPGIST_LIVE)
+					{
+						SpGistNodeTuple node;
+						int			i;
+
+						SGITITERATE(innerTuple, i, node)
+						{
+							if (ItemPointerIsValid(&node->t_tid))
+								spgAddPendingTID(bds, &node->t_tid);
+						}
+					}
+					else if (innerTuple->tupstate == SPGIST_REDIRECT)
+					{
+						/* transfer attention to redirect point */
+						spgAddPendingTID(bds,
+										 &((SpGistDeadTuple) innerTuple)->pointer);
+					}
+					else
+						elog(ERROR, "unexpected SPGiST tuple state: %d",
+							 innerTuple->tupstate);
+
+					nitem->done = true;
+				}
+			}
+		}
+
+		UnlockReleaseBuffer(buffer);
+	}
+
+	spgClearPendingList(bds);
+}
+
 /*
  * Perform a bulkdelete scan
  */
@@ -598,6 +804,8 @@ spgvacuumscan(spgBulkDeleteState *bds)
 
 	/* Finish setting up spgBulkDeleteState */
 	initSpGistState(&bds->spgstate, index);
+	bds->pendingList = NULL;
+	bds->myXmin = GetActiveSnapshot()->xmin;
 	bds->OldestXmin = GetOldestXmin(true, false);
 	bds->lastFilledBlock = SPGIST_LAST_FIXED_BLKNO;
 
@@ -637,6 +845,9 @@ spgvacuumscan(spgBulkDeleteState *bds)
 		for (; blkno < num_pages; blkno++)
 		{
 			spgvacuumpage(bds, blkno);
+			/* empty the pending-list after each page */
+			if (bds->pendingList != NULL)
+				spgprocesspending(bds);
 		}
 	}
 
@@ -747,7 +958,7 @@ spgvacuumcleanup(PG_FUNCTION_ARGS)
 	IndexFreeSpaceMapVacuum(index);
 
 	/*
-	 * It's quite possible for us to be fooled by concurrent page splits into
+	 * It's quite possible for us to be fooled by concurrent tuple moves into
 	 * double-counting some index tuples, so disbelieve any total that exceeds
 	 * the underlying heap's count ... if we know that accurately.  Otherwise
 	 * this might just make matters worse.