Make _bt_killitems drop pins it acquired itself.

Teach nbtree's _bt_killitems to leave the so->currPos page that it sets
LP_DEAD items on in whatever state it was in when _bt_killitems was
called.  In particular, make sure that so->dropPin scans don't acquire a
pin whose reference is saved in so->currPos.buf.

Allowing _bt_killitems to change so->currPos.buf like this is wrong.
The immediate consequence of allowing it is that code in _bt_steppage
(that copies so->currPos into so->markPos) will behave as if the scan is
a !so->dropPin scan.  so->markPos will therefore retain the buffer pin
indefinitely, even though _bt_killitems only needs to acquire a pin
(along with a lock) for long enough to mark known-dead items LP_DEAD.

This issue came to light following a report of a failure of an assertion
from recent commit e6eed40e.  The test case in question involves the use
of mark and restore.  An initial call to _bt_killitems takes place that
leaves so->currPos.buf in a state that is inconsistent with the scan
being so->dropPin.  A subsequent call to _bt_killitems for the same
position (following so->currPos being saved in so->markPos, and then
restored as so->currPos) resulted in the failure of an assertion that
tests that so->currPos.buf is InvalidBuffer when the scan is so->dropPin
(non-assert builds got a "resource was not closed" WARNING instead).

The same problem exists on earlier releases, though the issue is far
more subtle there.  Recent commit e6eed40e introduced the so->dropPin
field as a partial replacement for testing so->currPos.buf directly.
Earlier releases won't get an assertion failure (or buffer pin leak),
but they will allow the second _bt_killitems call from the test case to
behave as if a buffer pin was consistently held since the original call
to _bt_readpage.  This is wrong; there will have been an initial window
during which no pin was held on the so->currPos page, and yet the second
_bt_killitems call will neglect to check if so->currPos.lsn continues to
match the page's now-current LSN.

As a result of all this, it's just about possible that _bt_killitems
will set the wrong items LP_DEAD (on release branches).  This could only
happen with merge joins (the sole user of nbtree mark/restore support),
when a concurrently inserted index tuple used a recently-recycled TID
(and only when the new tuple was inserted onto the same page as a
distinct concurrently-removed tuple with the same TID).  This is exactly
the scenario that _bt_killitems' check of the page's now-current LSN
against the LSN stashed in currPos was supposed to prevent.

A follow-up commit will make nbtree completely stop conditioning whether
or not a position's pin needs to be dropped on whether the 'buf' field
is set.  All call sites that might need to drop a still-held pin will be
taught to rely on the scan-level so->dropPin field recently introduced
by commit e6eed40e.  That will make bugs of the same general nature as
this one impossible (or make them much easier to detect, at least).

Author: Peter Geoghegan <pg@bowt.ie>
Reported-By: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/545be1e5-3786-439a-9257-a90d30f8b849@gmail.com
Backpatch-through: 13

src/backend/access/nbtree/nbtree.c

@@ -417,6 +417,8 @@ btrescan(IndexScanDesc scan, ScanKey scankey, int nscankeys,
 	 * way, so we might as well avoid wasting cycles on acquiring page LSNs.
 	 *
 	 * See nbtree/README section on making concurrent TID recycling safe.
+	 *
+	 * Note: so->dropPin should never change across rescans.
 	 */
 	so->dropPin = (!scan->xs_want_itup &&
 				   IsMVCCSnapshot(scan->xs_snapshot) &&

src/backend/access/nbtree/nbtutils.c

@@ -3323,24 +3323,26 @@ _bt_checkkeys_look_ahead(IndexScanDesc scan, BTReadPageState *pstate,
  * current page and killed tuples thereon (generally, this should only be
  * called if so->numKilled > 0).
  *
- * The caller does not have a lock on the page and may or may not have the
- * page pinned in a buffer.  Note that read-lock is sufficient for setting
- * LP_DEAD status (which is only a hint).
+ * Caller should not have a lock on the so->currPos page, but must hold a
+ * buffer pin when !so->dropPin.  When we return, it still won't be locked.
+ * It'll continue to hold whatever pins were held before calling here.
  *
- * We match items by heap TID before assuming they are the right ones to
- * delete.  We cope with cases where items have moved right due to insertions.
- * If an item has moved off the current page due to a split, we'll fail to
- * find it and do nothing (this is not an error case --- we assume the item
- * will eventually get marked in a future indexscan).
- *
- * Note that if we hold a pin on the target page continuously from initially
- * reading the items until applying this function, VACUUM cannot have deleted
- * any items on the page, so the page's TIDs can't have been recycled by now.
- * There's no risk that we'll confuse a new index tuple that happens to use a
- * recycled TID with a now-removed tuple with the same TID (that used to be on
- * this same page).  We can't rely on that during scans that drop pins eagerly
+ * We match items by heap TID before assuming they are the right ones to set
+ * LP_DEAD.  If the scan is one that holds a buffer pin on the target page
+ * continuously from initially reading the items until applying this function
+ * (if it is a !so->dropPin scan), VACUUM cannot have deleted any items on the
+ * page, so the page's TIDs can't have been recycled by now.  There's no risk
+ * that we'll confuse a new index tuple that happens to use a recycled TID
+ * with a now-removed tuple with the same TID (that used to be on this same
+ * page).  We can't rely on that during scans that drop buffer pins eagerly
  * (so->dropPin scans), though, so we must condition setting LP_DEAD bits on
  * the page LSN having not changed since back when _bt_readpage saw the page.
+ * We totally give up on setting LP_DEAD bits when the page LSN changed.
+ *
+ * We give up much less often during !so->dropPin scans, but it still happens.
+ * We cope with cases where items have moved right due to insertions.  If an
+ * item has moved off the current page due to a split, we'll fail to find it
+ * and just give up on it.
  */
 void
 _bt_killitems(IndexScanDesc scan)
@@ -3353,6 +3355,7 @@ _bt_killitems(IndexScanDesc scan)
 	OffsetNumber maxoff;
 	int			numKilled = so->numKilled;
 	bool		killedsomething = false;
+	Buffer		buf;
 
 	Assert(numKilled > 0);
 	Assert(BTScanPosIsValid(so->currPos));
@@ -3369,11 +3372,11 @@ _bt_killitems(IndexScanDesc scan)
 		 * concurrent VACUUMs from recycling any of the TIDs on the page.
 		 */
 		Assert(BTScanPosIsPinned(so->currPos));
-		_bt_lockbuf(rel, so->currPos.buf, BT_READ);
+		buf = so->currPos.buf;
+		_bt_lockbuf(rel, buf, BT_READ);
 	}
 	else
 	{
-		Buffer		buf;
 		XLogRecPtr	latestlsn;
 
 		Assert(!BTScanPosIsPinned(so->currPos));
@@ -3391,10 +3394,9 @@ _bt_killitems(IndexScanDesc scan)
 		}
 
 		/* Unmodified, hinting is safe */
-		so->currPos.buf = buf;
 	}
 
-	page = BufferGetPage(so->currPos.buf);
+	page = BufferGetPage(buf);
 	opaque = BTPageGetOpaque(page);
 	minoff = P_FIRSTDATAKEY(opaque);
 	maxoff = PageGetMaxOffsetNumber(page);
@@ -3511,10 +3513,13 @@ _bt_killitems(IndexScanDesc scan)
 	if (killedsomething)
 	{
 		opaque->btpo_flags |= BTP_HAS_GARBAGE;
-		MarkBufferDirtyHint(so->currPos.buf, true);
+		MarkBufferDirtyHint(buf, true);
 	}
 
-	_bt_unlockbuf(rel, so->currPos.buf);
+	if (!so->dropPin)
+		_bt_unlockbuf(rel, buf);
+	else
+		_bt_relbuf(rel, buf);
 }