GiST: Invalidate killed items consistently.

GiST neglected to invalidate its killedItems[] array on a rescan.  As a
result, it was just about possible for the wrong tuples from the wrong
index page to be LP_DEAD-marked on a rescan.  The scan mistakenly
believed that the previous rescan's killedItems[] were for this rescan's
curBlkno, causing index corruption.

To fix, bring GiST in line with nbtree and hash: call gistkillitems from
both gistrescan and gistendscan (the existing gistgettuple caller still
handles the common case where we need to LP_DEAD-mark before moving on
to the next page).  That way the scan's pending killedItems[] are passed
to gistkillitems while they still describe items from curBlkno.  When
gistkillitems runs, it'll invalidate the array in passing (and won't
needlessly miss out on an opportunity to LP_DEAD-mark eligible index
tuples).  Back branches just get minimal hardening: we invalidate
killedItems[] at the places where the master branch gets new calls to
gistkillitems (and we invalidate curBlkno and curPageLSN on a rescan).

The test that proved corruption on master didn't result in corruption on
any stable branch, though only because, without commit 9c9ddf109, we'd
clobber curPageLSN without also updating curBlkno -- which accidentally
prevented it.  Relying on gistkillitems to not LP_DEAD-mark by passing
it a curBlkno whose curPageLSN was taken from an entirely different page
seems like a very bad idea, which is why this issue is being treated as
a bug affecting all stable branches.

Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Andrey Borodin <x4mmm@yandex-team.ru>
Discussion: https://postgr.es/m/CAH2-WzmwEThnQf17Ju+t0N9_KJLsEQSXzYrFnaS2=s4KnGGrqw@mail.gmail.com
Backpatch-through: 14

2 files changed, 12 insertions, 0 deletions

diff --git a/src/backend/access/gist/gistget.c b/src/backend/access/gist/gistget.c
index adbf622c83c..b87c43b98d0 100644
--- a/src/backend/access/gist/gistget.c
+++ b/src/backend/access/gist/gistget.c
@@ -425,7 +425,10 @@ gistScanPage(IndexScanDesc scan, GISTSearchItem *pageItem,
 		 * killed tuple as not passing the qual.
 		 */
 		if (scan->ignore_killed_tuples && ItemIdIsDead(iid))
+		{
+			Assert(GistPageIsLeaf(page));
 			continue;
+		}
 
 		it = (IndexTuple) PageGetItem(page, iid);
 
@@ -722,6 +725,7 @@ gistgettuple(IndexScanDesc scan, ScanDirection dir)
 				CHECK_FOR_INTERRUPTS();
 
 				/* save current item BlockNumber for next gistkillitems() call */
+				Assert(so->numKilled == 0);
 				so->curBlkno = item->blkno;
 
 				/*
diff --git a/src/backend/access/gist/gistscan.c b/src/backend/access/gist/gistscan.c
index 4252ff44b25..d8a44c43b4a 100644
--- a/src/backend/access/gist/gistscan.c
+++ b/src/backend/access/gist/gistscan.c
@@ -133,7 +133,12 @@ gistrescan(IndexScanDesc scan, ScanKey key, int nkeys,
 	int			i;
 	MemoryContext oldCxt;
 
+	/* invalidate any killed items still pending */
+	so->numKilled = 0;
+
 	/* rescan an existing indexscan --- reset state */
+	so->curBlkno = InvalidBlockNumber;
+	so->curPageLSN = InvalidXLogRecPtr;
 
 	/*
 	 * The first time through, we create the search queue in the scanCxt.
@@ -350,6 +355,9 @@ gistendscan(IndexScanDesc scan)
 {
 	GISTScanOpaque so = (GISTScanOpaque) scan->opaque;
 
+	/* invalidate any killed items still pending */
+	so->numKilled = 0;
+
 	/*
 	 * freeGISTstate is enough to clean up everything made by gistbeginscan,
 	 * as well as the queueCxt if there is a separate context for it.