Skip foreign tablespaces when running pg_checksums/pg_verify_checksums

michaelpq · michaelpq · commit 428a2609ef64 · 2020-02-27T15:31:27.000+09:00
Attempting to use pg_checksums (pg_verify_checksums in 11) on a data folder which includes tablespace paths used across multiple major versions would cause pg_checksums to scan all directories present in pg_tblspc, and not only marked with TABLESPACE_VERSION_DIRECTORY. This could lead to failures when for example running sanity checks on an upgraded instance with --check. Even worse, it was possible to rewrite on-disk pages with --enable for a cluster potentially online. This commit makes pg_checksums skip any directories not named TABLESPACE_VERSION_DIRECTORY, similarly to what is done for base backups. Reported-by: Michael Banck Author: Michael Banck, Bernd Helmle Discussion: https://postgr.es/m/62031974fd8e941dd8351fbc8c7eff60d59c5338.camel@credativ.de backpatch-through: 11
diff --git a/src/bin/pg_checksums/pg_checksums.c b/src/bin/pg_checksums/pg_checksums.c
@@ -385,7 +385,52 @@ scan_directory(const char *basedir, const char *subdir, bool sizeonly)
 #else
 		else if (S_ISDIR(st.st_mode) || pgwin32_is_junction(fn))
 #endif
-			dirsize += scan_directory(path, de->d_name, sizeonly);
+		{
+			/*
+			 * If going through the entries of pg_tblspc, we assume to operate
+			 * on tablespace locations where only TABLESPACE_VERSION_DIRECTORY
+			 * is valid, resolving the linked locations and dive into them
+			 * directly.
+			 */
+			if (strncmp("pg_tblspc", subdir, strlen("pg_tblspc")) == 0)
+			{
+				char		tblspc_path[MAXPGPATH];
+				struct stat tblspc_st;
+
+				/*
+				 * Resolve tablespace location path and check whether
+				 * TABLESPACE_VERSION_DIRECTORY exists.  Not finding a valid
+				 * location is unexpected, since there should be no orphaned
+				 * links and no links pointing to something else than a
+				 * directory.
+				 */
+				snprintf(tblspc_path, sizeof(tblspc_path), "%s/%s/%s",
+						 path, de->d_name, TABLESPACE_VERSION_DIRECTORY);
+
+				if (lstat(tblspc_path, &tblspc_st) < 0)
+				{
+					pg_log_error("could not stat file \"%s\": %m",
+								 tblspc_path);
+					exit(1);
+				}
+
+				/*
+				 * Move backwards once as the scan needs to happen for the
+				 * contents of TABLESPACE_VERSION_DIRECTORY.
+				 */
+				snprintf(tblspc_path, sizeof(tblspc_path), "%s/%s",
+						 path, de->d_name);
+
+				/* Looks like a valid tablespace location */
+				dirsize += scan_directory(tblspc_path,
+										  TABLESPACE_VERSION_DIRECTORY,
+										  sizeonly);
+			}
+			else
+			{
+				dirsize += scan_directory(path, de->d_name, sizeonly);
+			}
+		}
 	}
 	closedir(dir);
 	return dirsize;
diff --git a/src/bin/pg_checksums/t/002_actions.pl b/src/bin/pg_checksums/t/002_actions.pl
@@ -5,7 +5,7 @@
 use warnings;
 use PostgresNode;
 use TestLib;
-use Test::More tests => 62;
+use Test::More tests => 63;
 
 
 # Utility routine to create and check a table with corrupted checksums
@@ -217,6 +217,13 @@ sub fail_corrupt
 # Stop instance for the follow-up checks.
 $node->stop;
 
+# Create a fake tablespace location that should not be scanned
+# when verifying checksums.
+mkdir "$tablespace_dir/PG_99_999999991/";
+append_to_file "$tablespace_dir/PG_99_999999991/foo", "123";
+command_ok([ 'pg_checksums', '--check', '-D', $pgdata ],
+	"succeeds with foreign tablespace");
+
 # Authorized relation files filled with corrupted data cause the
 # checksum checks to fail.  Make sure to use file names different
 # than the previous ones.
