diff --git a/contrib/amcheck/t/003_cic_2pc.pl b/contrib/amcheck/t/003_cic_2pc.pl
new file mode 100644
index 0000000000..803b99a738
--- /dev/null
+++ b/contrib/amcheck/t/003_cic_2pc.pl
@@ -0,0 +1,188 @@
+
+# Copyright (c) 2021, PostgreSQL Global Development Group
+
+# Test CREATE INDEX CONCURRENTLY with concurrent prepared-xact modifications
+use strict;
+use warnings;
+
+use Config;
+use PostgresNode;
+use TestLib;
+
+use Test::More tests => 6;
+
+my ($node, $result);
+
+#
+# Test set-up
+#
+$node = PostgresNode->new('CIC_2PC_test');
+$node->init;
+$node->append_conf('postgresql.conf', 'max_prepared_transactions = 10');
+$node->append_conf('postgresql.conf', 'lock_timeout = 180000');
+$node->start;
+$node->safe_psql('postgres', q(CREATE EXTENSION amcheck));
+$node->safe_psql('postgres', q(CREATE TABLE tbl(i int)));
+
+
+#
+# Run 3 overlapping 2PC transactions with CIC
+#
+# We have two concurrent background psql processes: $main_h for INSERTs and
+# $cic_h for CIC.  Also, we use non-background psql for some COMMIT PREPARED
+# statements.
+#
+
+my $main_in    = '';
+my $main_out   = '';
+my $main_timer = IPC::Run::timeout(180);
+
+my $main_h =
+  $node->background_psql('postgres', \$main_in, \$main_out,
+	$main_timer, on_error_stop => 1);
+$main_in .= q(
+BEGIN;
+INSERT INTO tbl VALUES(0);
+\echo syncpoint1
+);
+pump $main_h until $main_out =~ /syncpoint1/ || $main_timer->is_expired;
+
+my $cic_in    = '';
+my $cic_out   = '';
+my $cic_timer = IPC::Run::timeout(180);
+my $cic_h =
+  $node->background_psql('postgres', \$cic_in, \$cic_out,
+	$cic_timer, on_error_stop => 1);
+$cic_in .= q(
+\echo start
+CREATE INDEX CONCURRENTLY idx ON tbl(i);
+);
+pump $cic_h until $cic_out =~ /start/ || $cic_timer->is_expired;
+
+$main_in .= q(
+PREPARE TRANSACTION 'a';
+);
+
+$main_in .= q(
+BEGIN;
+INSERT INTO tbl VALUES(0);
+\echo syncpoint2
+);
+pump $main_h until $main_out =~ /syncpoint2/ || $main_timer->is_expired;
+
+$node->safe_psql('postgres', q(COMMIT PREPARED 'a';));
+
+$main_in .= q(
+PREPARE TRANSACTION 'b';
+BEGIN;
+INSERT INTO tbl VALUES(0);
+\echo syncpoint3
+);
+pump $main_h until $main_out =~ /syncpoint3/ || $main_timer->is_expired;
+
+$node->safe_psql('postgres', q(COMMIT PREPARED 'b';));
+
+$main_in .= q(
+PREPARE TRANSACTION 'c';
+COMMIT PREPARED 'c';
+);
+$main_h->pump_nb;
+
+$main_h->finish;
+$cic_h->finish;
+
+$result = $node->psql('postgres', q(SELECT bt_index_check('idx',true)));
+is($result, '0', 'bt_index_check after overlapping 2PC');
+
+
+#
+# Server restart shall not change whether prepared xact blocks CIC
+#
+
+$node->safe_psql(
+	'postgres', q(
+BEGIN;
+INSERT INTO tbl VALUES(0);
+PREPARE TRANSACTION 'spans_restart';
+BEGIN;
+CREATE TABLE unused ();
+PREPARE TRANSACTION 'persists_forever';
+));
+$node->restart;
+
+my $reindex_in    = '';
+my $reindex_out   = '';
+my $reindex_timer = IPC::Run::timeout(180);
+my $reindex_h =
+  $node->background_psql('postgres', \$reindex_in, \$reindex_out,
+	$reindex_timer, on_error_stop => 1);
+$reindex_in .= q(
+\echo start
+DROP INDEX CONCURRENTLY idx;
+CREATE INDEX CONCURRENTLY idx ON tbl(i);
+);
+pump $reindex_h until $reindex_out =~ /start/ || $reindex_timer->is_expired;
+
+$node->safe_psql('postgres', "COMMIT PREPARED 'spans_restart'");
+$reindex_h->finish;
+$result = $node->psql('postgres', q(SELECT bt_index_check('idx',true)));
+is($result, '0', 'bt_index_check after 2PC and restart');
+
+
+#
+# Stress CIC+2PC with pgbench
+#
+
+# Fix broken index first
+$node->safe_psql('postgres', q(REINDEX TABLE tbl;));
+
+# Run background pgbench with CIC. We cannot mix-in this script into single
+# pgbench: CIC will deadlock with itself occasionally.
+my $pgbench_out   = '';
+my $pgbench_timer = IPC::Run::timeout(180);
+my $pgbench_h     = $node->background_pgbench(
+	'--no-vacuum --client=1 --transactions=100',
+	{
+		'002_pgbench_concurrent_cic' => q(
+			DROP INDEX CONCURRENTLY idx;
+			CREATE INDEX CONCURRENTLY idx ON tbl(i);
+			SELECT bt_index_check('idx',true);
+		   )
+	},
+	\$pgbench_out,
+	$pgbench_timer);
+
+# Run pgbench.
+$node->pgbench(
+	'--no-vacuum --client=5 --transactions=100',
+	0,
+	[qr{actually processed}],
+	[qr{^$}],
+	'concurrent INSERTs w/ 2PC',
+	{
+		'002_pgbench_concurrent_2pc' => q(
+			BEGIN;
+			INSERT INTO tbl VALUES(0);
+			PREPARE TRANSACTION 'c:client_id';
+			COMMIT PREPARED 'c:client_id';
+		  ),
+		'002_pgbench_concurrent_2pc_savepoint' => q(
+			BEGIN;
+			SAVEPOINT s1;
+			INSERT INTO tbl VALUES(0);
+			PREPARE TRANSACTION 'c:client_id';
+			COMMIT PREPARED 'c:client_id';
+		  )
+	});
+
+$pgbench_h->pump_nb;
+$pgbench_h->finish();
+$result =
+    ($Config{osname} eq "MSWin32")
+  ? ($pgbench_h->full_results)[0]
+  : $pgbench_h->result(0);
+is($result, 0, "pgbench with CIC works");
+
+# done
+$node->stop;
+done_testing();
diff --git a/src/backend/access/transam/twophase.c b/src/backend/access/transam/twophase.c
index 2156de187c..f6e7fa71d8 100644
--- a/src/backend/access/transam/twophase.c
+++ b/src/backend/access/transam/twophase.c
@@ -459,14 +459,24 @@ MarkAsPreparingGuts(GlobalTransaction gxact, TransactionId xid, const char *gid,
 	proc->pgprocno = gxact->pgprocno;
 	SHMQueueElemInit(&(proc->links));
 	proc->waitStatus = PROC_WAIT_STATUS_OK;
-	/* We set up the gxact's VXID as InvalidBackendId/XID */
-	proc->lxid = (LocalTransactionId) xid;
+	if (LocalTransactionIdIsValid(MyProc->lxid))
+	{
+		/* clone VXID, for TwoPhaseGetXidByVirtualXID() to find */
+		proc->lxid = MyProc->lxid;
+		proc->backendId = MyBackendId;
+	}
+	else
+	{
+		Assert(AmStartupProcess() || !IsPostmasterEnvironment);
+		/* GetLockConflicts() uses this to specify a wait on the XID */
+		proc->lxid = xid;
+		proc->backendId = InvalidBackendId;
+	}
 	proc->xid = xid;
 	Assert(proc->xmin == InvalidTransactionId);
 	proc->delayChkpt = false;
 	proc->statusFlags = 0;
 	proc->pid = 0;
-	proc->backendId = InvalidBackendId;
 	proc->databaseId = databaseid;
 	proc->roleId = owner;
 	proc->tempNamespaceId = InvalidOid;
@@ -846,6 +856,53 @@ TwoPhaseGetGXact(TransactionId xid, bool lock_held)
 	return result;
 }
 
+/*
+ * TwoPhaseGetXidByVirtualXID
+ *		Lookup VXID among xacts prepared since last startup.
+ *
+ * (This won't find recovered xacts.)  If more than one matches, return any
+ * and set "have_more" to true.  To witness multiple matches, a single
+ * BackendId must consume 2^32 LXIDs, with no intervening database restart.
+ */
+TransactionId
+TwoPhaseGetXidByVirtualXID(VirtualTransactionId vxid,
+						   bool *have_more)
+{
+	int			i;
+	TransactionId result = InvalidTransactionId;
+
+	Assert(VirtualTransactionIdIsValid(vxid));
+	LWLockAcquire(TwoPhaseStateLock, LW_SHARED);
+
+	for (i = 0; i < TwoPhaseState->numPrepXacts; i++)
+	{
+		GlobalTransaction gxact = TwoPhaseState->prepXacts[i];
+		PGPROC	   *proc;
+		VirtualTransactionId proc_vxid;
+
+		if (!gxact->valid)
+			continue;
+		proc = &ProcGlobal->allProcs[gxact->pgprocno];
+		GET_VXID_FROM_PGPROC(proc_vxid, *proc);
+		if (VirtualTransactionIdEquals(vxid, proc_vxid))
+		{
+			/* Startup process sets proc->backendId to InvalidBackendId. */
+			Assert(!gxact->inredo);
+
+			if (result != InvalidTransactionId)
+			{
+				*have_more = true;
+				break;
+			}
+			result = gxact->xid;
+		}
+	}
+
+	LWLockRelease(TwoPhaseStateLock);
+
+	return result;
+}
+
 /*
  * TwoPhaseGetDummyBackendId
  *		Get the dummy backend ID for prepared transaction specified by XID
diff --git a/src/backend/access/transam/xact.c b/src/backend/access/transam/xact.c
index d96881c0de..ca6f6d57d3 100644
--- a/src/backend/access/transam/xact.c
+++ b/src/backend/access/transam/xact.c
@@ -2506,6 +2506,13 @@ PrepareTransaction(void)
 	/* Reset XactLastRecEnd until the next transaction writes something */
 	XactLastRecEnd = 0;
 
+	/*
+	 * Transfer our locks to a dummy PGPROC.  This has to be done before
+	 * ProcArrayClearTransaction().  Otherwise, a GetLockConflicts() would
+	 * conclude "xact already committed or aborted" for our locks.
+	 */
+	PostPrepare_Locks(xid);
+
 	/*
 	 * Let others know about no transaction in progress by me.  This has to be
 	 * done *after* the prepared transaction has been marked valid, else
@@ -2545,7 +2552,6 @@ PrepareTransaction(void)
 
 	PostPrepare_MultiXact(xid);
 
-	PostPrepare_Locks(xid);
 	PostPrepare_PredicateLocks(xid);
 
 	ResourceOwnerRelease(TopTransactionResourceOwner,
diff --git a/src/backend/storage/lmgr/lmgr.c b/src/backend/storage/lmgr/lmgr.c
index cdf2266d6d..2db0424ad9 100644
--- a/src/backend/storage/lmgr/lmgr.c
+++ b/src/backend/storage/lmgr/lmgr.c
@@ -871,9 +871,10 @@ XactLockTableWaitErrorCb(void *arg)
  * To do this, obtain the current list of lockers, and wait on their VXIDs
  * until they are finished.
  *
- * Note we don't try to acquire the locks on the given locktags, only the VXIDs
- * of its lock holders; if somebody grabs a conflicting lock on the objects
- * after we obtained our initial list of lockers, we will not wait for them.
+ * Note we don't try to acquire the locks on the given locktags, only the
+ * VXIDs and XIDs of their lock holders; if somebody grabs a conflicting lock
+ * on the objects after we obtained our initial list of lockers, we will not
+ * wait for them.
  */
 void
 WaitForLockersMultiple(List *locktags, LOCKMODE lockmode, bool progress)
diff --git a/src/backend/storage/lmgr/lock.c b/src/backend/storage/lmgr/lock.c
index 364654e106..c25af7fe09 100644
--- a/src/backend/storage/lmgr/lock.c
+++ b/src/backend/storage/lmgr/lock.c
@@ -2900,8 +2900,12 @@ FastPathGetRelationLockEntry(LOCALLOCK *locallock)
  * The result array is palloc'd and is terminated with an invalid VXID.
  * *countp, if not null, is updated to the number of items set.
  *
- * Of course, the result could be out of date by the time it's returned,
- * so use of this function has to be thought about carefully.
+ * Of course, the result could be out of date by the time it's returned, so
+ * use of this function has to be thought about carefully.  Similarly, a
+ * PGPROC with no "lxid" will be considered non-conflicting regardless of any
+ * lock it holds.  Existing callers don't care about a locker after that
+ * locker's pg_xact updates complete.  CommitTransaction() clears "lxid" after
+ * pg_xact updates and before releasing locks.
  *
  * Note we never include the current xact's vxid in the result array,
  * since an xact never blocks itself.
@@ -4529,37 +4533,80 @@ VirtualXactLockTableCleanup(void)
 	}
 }
 
+/*
+ *		XactLockForVirtualXact
+ *
+ * If TransactionIdIsValid(xid), this is essentially XactLockTableWait(xid,
+ * NULL, NULL, XLTW_None) or ConditionalXactLockTableWait(xid).  Unlike those
+ * functions, it assumes "xid" is never a subtransaction and that "xid" is
+ * prepared, committed, or aborted.
+ *
+ * If !TransactionIdIsValid(xid), this locks every prepared XID having been
+ * known as "vxid" before its PREPARE TRANSACTION.
+ */
+static bool
+XactLockForVirtualXact(VirtualTransactionId vxid,
+					   TransactionId xid, bool wait)
+{
+	bool		more = false;
+
+	/* There is no point to wait for 2PCs if you have no 2PCs. */
+	if (max_prepared_xacts == 0)
+		return true;
+
+	do
+	{
+		LockAcquireResult lar;
+		LOCKTAG		tag;
+
+		/* Clear state from previous iterations. */
+		if (more)
+		{
+			xid = InvalidTransactionId;
+			more = false;
+		}
+
+		/* If we have no xid, try to find one. */
+		if (!TransactionIdIsValid(xid))
+			xid = TwoPhaseGetXidByVirtualXID(vxid, &more);
+		if (!TransactionIdIsValid(xid))
+		{
+			Assert(!more);
+			return true;
+		}
+
+		/* Check or wait for XID completion. */
+		SET_LOCKTAG_TRANSACTION(tag, xid);
+		lar = LockAcquire(&tag, ShareLock, false, !wait);
+		if (lar == LOCKACQUIRE_NOT_AVAIL)
+			return false;
+		LockRelease(&tag, ShareLock, false);
+	} while (more);
+
+	return true;
+}
+
 /*
  *		VirtualXactLock
  *
- * If wait = true, wait until the given VXID has been released, and then
- * return true.
+ * If wait = true, wait as long as the given VXID or any XID acquired by the
+ * same transaction is still running.  Then, return true.
  *
- * If wait = false, just check whether the VXID is still running, and return
- * true or false.
+ * If wait = false, just check whether that VXID or one of those XIDs is still
+ * running, and return true or false.
  */
 bool
 VirtualXactLock(VirtualTransactionId vxid, bool wait)
 {
 	LOCKTAG		tag;
 	PGPROC	   *proc;
+	TransactionId xid = InvalidTransactionId;
 
 	Assert(VirtualTransactionIdIsValid(vxid));
 
-	if (VirtualTransactionIdIsPreparedXact(vxid))
-	{
-		LockAcquireResult lar;
-
-		/*
-		 * Prepared transactions don't hold vxid locks.  The
-		 * LocalTransactionId is always a normal, locked XID.
-		 */
-		SET_LOCKTAG_TRANSACTION(tag, vxid.localTransactionId);
-		lar = LockAcquire(&tag, ShareLock, false, !wait);
-		if (lar != LOCKACQUIRE_NOT_AVAIL)
-			LockRelease(&tag, ShareLock, false);
-		return lar != LOCKACQUIRE_NOT_AVAIL;
-	}
+	if (VirtualTransactionIdIsRecoveredPreparedXact(vxid))
+		/* no vxid lock; localTransactionId is a normal, locked XID */
+		return XactLockForVirtualXact(vxid, vxid.localTransactionId, wait);
 
 	SET_LOCKTAG_VIRTUALTRANSACTION(tag, vxid);
 
@@ -4573,7 +4620,7 @@ VirtualXactLock(VirtualTransactionId vxid, bool wait)
 	 */
 	proc = BackendIdGetProc(vxid.backendId);
 	if (proc == NULL)
-		return true;
+		return XactLockForVirtualXact(vxid, InvalidTransactionId, wait);
 
 	/*
 	 * We must acquire this lock before checking the backendId and lxid
@@ -4582,12 +4629,12 @@ VirtualXactLock(VirtualTransactionId vxid, bool wait)
 	 */
 	LWLockAcquire(&proc->fpInfoLock, LW_EXCLUSIVE);
 
-	/* If the transaction has ended, our work here is done. */
 	if (proc->backendId != vxid.backendId
 		|| proc->fpLocalTransactionId != vxid.localTransactionId)
 	{
+		/* VXID ended */
 		LWLockRelease(&proc->fpInfoLock);
-		return true;
+		return XactLockForVirtualXact(vxid, InvalidTransactionId, wait);
 	}
 
 	/*
@@ -4634,6 +4681,16 @@ VirtualXactLock(VirtualTransactionId vxid, bool wait)
 		proc->fpVXIDLock = false;
 	}
 
+	/*
+	 * If the proc has an XID now, we'll avoid a TwoPhaseGetXidByVirtualXID()
+	 * search.  The proc might have assigned this XID but not yet locked it,
+	 * in which case the proc will lock this XID before releasing the VXID.
+	 * The fpInfoLock critical section excludes VirtualXactLockTableCleanup(),
+	 * so we won't save an XID of a different VXID.  It doesn't matter whether
+	 * we save this before or after setting up the primary lock table entry.
+	 */
+	xid = proc->xid;
+
 	/* Done with proc->fpLockBits */
 	LWLockRelease(&proc->fpInfoLock);
 
@@ -4641,7 +4698,7 @@ VirtualXactLock(VirtualTransactionId vxid, bool wait)
 	(void) LockAcquire(&tag, ShareLock, false, false);
 
 	LockRelease(&tag, ShareLock, false);
-	return true;
+	return XactLockForVirtualXact(vxid, xid, wait);
 }
 
 /*
diff --git a/src/backend/utils/cache/inval.c b/src/backend/utils/cache/inval.c
index dd8586ab4d..777fab4915 100644
--- a/src/backend/utils/cache/inval.c
+++ b/src/backend/utils/cache/inval.c
@@ -65,6 +65,20 @@
  *	(XXX is it worth testing likewise for duplicate catcache flush entries?
  *	Probably not.)
  *
+ *	Many subsystems own higher-level caches that depend on relcache and/or
+ *	catcache, and they register callbacks here to invalidate their caches.
+ *	While building a higher-level cache entry, a backend may receive a
+ *	callback for the being-built entry or one of its dependencies.  This
+ *	implies the new higher-level entry would be born stale, and it might
+ *	remain stale for the life of the backend.  Many caches do not prevent
+ *	that.  They rely on DDL for can't-miss catalog changes taking
+ *	AccessExclusiveLock on suitable objects.  (For a change made with less
+ *	locking, backends might never read the change.)  The relation cache,
+ *	however, needs to reflect changes from CREATE INDEX CONCURRENTLY no later
+ *	than the beginning of the next transaction.  Hence, when a relevant
+ *	invalidation callback arrives during a build, relcache.c reattempts that
+ *	build.  Caches with similar needs could do likewise.
+ *
  *	If a relcache flush is issued for a system relation that we preload
  *	from the relcache init file, we must also delete the init file so that
  *	it will be rebuilt during the next backend restart.  The actual work of
diff --git a/src/include/access/twophase.h b/src/include/access/twophase.h
index e27e1a8fe8..2d758422a7 100644
--- a/src/include/access/twophase.h
+++ b/src/include/access/twophase.h
@@ -34,6 +34,8 @@ extern void TwoPhaseShmemInit(void);
 extern void AtAbort_Twophase(void);
 extern void PostPrepare_Twophase(void);
 
+extern TransactionId TwoPhaseGetXidByVirtualXID(VirtualTransactionId vxid,
+												bool *have_more);
 extern PGPROC *TwoPhaseGetDummyProc(TransactionId xid, bool lock_held);
 extern BackendId TwoPhaseGetDummyBackendId(TransactionId xid, bool lock_held);
 
diff --git a/src/include/storage/lock.h b/src/include/storage/lock.h
index 9b2a421c32..a5286fab89 100644
--- a/src/include/storage/lock.h
+++ b/src/include/storage/lock.h
@@ -47,10 +47,11 @@ extern bool Debug_deadlocks;
 
 /*
  * Top-level transactions are identified by VirtualTransactionIDs comprising
- * PGPROC fields backendId and lxid.  For prepared transactions, the
- * LocalTransactionId is an ordinary XID.  These are guaranteed unique over
- * the short term, but will be reused after a database restart or XID
- * wraparound; hence they should never be stored on disk.
+ * PGPROC fields backendId and lxid.  For recovered prepared transactions, the
+ * LocalTransactionId is an ordinary XID; LOCKTAG_VIRTUALTRANSACTION never
+ * refers to that kind.  These are guaranteed unique over the short term, but
+ * will be reused after a database restart or XID wraparound; hence they
+ * should never be stored on disk.
  *
  * Note that struct VirtualTransactionId can not be assumed to be atomically
  * assignable as a whole.  However, type LocalTransactionId is assumed to
@@ -70,7 +71,7 @@ typedef struct
 #define LocalTransactionIdIsValid(lxid) ((lxid) != InvalidLocalTransactionId)
 #define VirtualTransactionIdIsValid(vxid) \
 	(LocalTransactionIdIsValid((vxid).localTransactionId))
-#define VirtualTransactionIdIsPreparedXact(vxid) \
+#define VirtualTransactionIdIsRecoveredPreparedXact(vxid) \
 	((vxid).backendId == InvalidBackendId)
 #define VirtualTransactionIdEquals(vxid1, vxid2) \
 	((vxid1).backendId == (vxid2).backendId && \