From fb5d5dc9e1e8fc395752906490e4d18d645e5994 Mon Sep 17 00:00:00 2001
From: Mikhail Nikalayeu <mihailnikalayeu@gmail.com>
Date: Tue, 4 Aug 2026 00:46:42 +0200
Subject: [PATCH v1] Do not hold rd_smgr across the fork loop when enabling
 data checksums

ProcessSingleRelationByOid() called RelationGetSmgr(rel), discarded the
result, and then read rel->rd_smgr directly on each iteration of the loop
over forks. Only RelationGetSmgr() is authorized to read that field because
a relcache invalidation resets rd_smgr to NULL.

When one lands between two forks of the same relation, the next iteration
passes NULL to smgrexists() and the worker dies.
---
 src/backend/postmaster/datachecksum_state.c   |  3 +-
 src/test/modules/test_checksums/meson.build   |  1 +
 .../test_checksums/t/010_discard_caches.pl    | 49 +++++++++++++++++++
 3 files changed, 51 insertions(+), 2 deletions(-)
 create mode 100644 src/test/modules/test_checksums/t/010_discard_caches.pl

diff --git a/src/backend/postmaster/datachecksum_state.c b/src/backend/postmaster/datachecksum_state.c
index 569399fb2b1..be308a282d5 100644
--- a/src/backend/postmaster/datachecksum_state.c
+++ b/src/backend/postmaster/datachecksum_state.c
@@ -841,11 +841,10 @@ ProcessSingleRelationByOid(Oid relationId, BufferAccessStrategy strategy)
 		pgstat_report_activity(STATE_IDLE, NULL);
 		return true;
 	}
-	RelationGetSmgr(rel);
 
 	for (ForkNumber fnum = 0; fnum <= MAX_FORKNUM; fnum++)
 	{
-		if (smgrexists(rel->rd_smgr, fnum))
+		if (smgrexists(RelationGetSmgr(rel), fnum))
 		{
 			if (!ProcessSingleRelationFork(rel, fnum, strategy))
 			{
diff --git a/src/test/modules/test_checksums/meson.build b/src/test/modules/test_checksums/meson.build
index 9b1421a9b91..c19e84210a2 100644
--- a/src/test/modules/test_checksums/meson.build
+++ b/src/test/modules/test_checksums/meson.build
@@ -33,6 +33,7 @@ tests += {
       't/007_pgbench_standby.pl',
       't/008_pitr.pl',
       't/009_fpi.pl',
+      't/010_discard_caches.pl',
     ],
   },
 }
diff --git a/src/test/modules/test_checksums/t/010_discard_caches.pl b/src/test/modules/test_checksums/t/010_discard_caches.pl
new file mode 100644
index 00000000000..131fbc7ff7f
--- /dev/null
+++ b/src/test/modules/test_checksums/t/010_discard_caches.pl
@@ -0,0 +1,49 @@
+
+# Copyright (c) 2026, PostgreSQL Global Development Group
+
+# Enabling data checksums while the relcache is being invalidated.
+
+use strict;
+use warnings FATAL => 'all';
+
+use PostgreSQL::Test::Cluster;
+use PostgreSQL::Test::Utils;
+use Test::More;
+
+use FindBin;
+use lib $FindBin::RealBin;
+
+use DataChecksums::Utils;
+
+# debug_discard_caches only does anything where DISCARD_CACHES_ENABLED is
+# defined, which follows USE_ASSERT_CHECKING; elsewhere the GUC rejects
+# any value above zero and the server would refuse to start.
+if (!check_pg_config('#define USE_ASSERT_CHECKING 1'))
+{
+	plan skip_all => 'this build does not have debug_discard_caches';
+}
+
+my $node = PostgreSQL::Test::Cluster->new('discard_caches');
+$node->init(no_data_checksums => 1);
+$node->append_conf('postgresql.conf', 'debug_discard_caches = 1');
+$node->start;
+
+test_checksum_state($node, 'off');
+
+# A little user data, so the worker has relations of its own to walk
+# besides the catalogs.
+$node->safe_psql('postgres',
+	'CREATE TABLE t AS SELECT generate_series(1, 1000) AS a');
+
+# The whole test: the worker must survive walking the cluster while every
+# catalog access flushes the caches.
+enable_data_checksums($node, wait => 'on');
+
+test_checksum_state($node, 'on');
+
+is($node->safe_psql('postgres', 'SELECT count(*) FROM t'),
+	'1000', 'the cluster survived enabling checksums');
+
+$node->stop;
+
+done_testing();
-- 
2.43.0