PG-1842 PG-1843 Optimize deletion of leftover relation keys

Instead of first deleting any leftover key and then writing the new key
we do a single pass through the file where we replace any old key that
we find. To make this happen on redo too we need to stop generating a
separate WAL record for the key deletion for encrypted tables and only
generate that record for unencrypted tables where we still need a key
deletion record.

We except this optimization to primarily be visible on WAL replay where
only a single backend is used to replay everything, but it also speeds
up table creation in general on workloads with many tables.

Author: jeltz

contrib/pg_tde/src/access/pg_tde_tdemap.c

@@ -97,7 +97,7 @@ static void pg_tde_write_one_map_entry(int fd, const TDEMapEntry *map_entry, off
 static int	pg_tde_file_header_write(const char *tde_filename, int fd, const TDESignedPrincipalKeyInfo *signed_key_info, off_t *bytes_written);
 static void pg_tde_initialize_map_entry(TDEMapEntry *map_entry, const TDEPrincipalKey *principal_key, const RelFileLocator *rlocator, const InternalKey *rel_key_data);
 static int	pg_tde_open_file_write(const char *tde_filename, const TDESignedPrincipalKeyInfo *signed_key_info, bool truncate, off_t *curr_pos);
-static void pg_tde_write_key_map_entry(const RelFileLocator *rlocator, const InternalKey *rel_key_data, const TDEPrincipalKey *principal_key);
+static void pg_tde_replace_key_map_entry(const RelFileLocator *rlocator, const InternalKey *rel_key_data, const TDEPrincipalKey *principal_key);
 
 void
 pg_tde_save_smgr_key(RelFileLocator rel, const InternalKey *rel_key_data)
@@ -114,7 +114,7 @@ pg_tde_save_smgr_key(RelFileLocator rel, const InternalKey *rel_key_data)
 				errhint("Use pg_tde_set_key_using_database_key_provider() or pg_tde_set_key_using_global_key_provider() to configure one."));
 	}
 
-	pg_tde_write_key_map_entry(&rel, rel_key_data, principal_key);
+	pg_tde_replace_key_map_entry(&rel, rel_key_data, principal_key);
 	LWLockRelease(lock_pk);
 }
 
@@ -452,11 +452,12 @@ pg_tde_write_one_map_entry(int fd, const TDEMapEntry *map_entry, off_t *offset,
  * concurrent in place updates leading to data conflicts.
  */
 void
-pg_tde_write_key_map_entry(const RelFileLocator *rlocator, const InternalKey *rel_key_data, const TDEPrincipalKey *principal_key)
+pg_tde_replace_key_map_entry(const RelFileLocator *rlocator, const InternalKey *rel_key_data, const TDEPrincipalKey *principal_key)
 {
 	char		db_map_path[MAXPGPATH];
 	int			map_fd;
 	off_t		curr_pos = 0;
+	off_t		write_pos = 0;
 	TDEMapEntry write_map_entry;
 	TDESignedPrincipalKeyInfo signed_key_Info;
 
@@ -481,22 +482,26 @@ pg_tde_write_key_map_entry(const RelFileLocator *rlocator, const InternalKey *re
 
 		if (!pg_tde_read_one_map_entry(map_fd, &read_map_entry, &curr_pos))
 		{
-			curr_pos = prev_pos;
+			if (write_pos == 0)
+				write_pos = prev_pos;
 			break;
 		}
 
-		if (read_map_entry.type == MAP_ENTRY_TYPE_EMPTY)
+		if (read_map_entry.spcOid == rlocator->spcOid && read_map_entry.relNumber == rlocator->relNumber)
 		{
-			curr_pos = prev_pos;
+			write_pos = prev_pos;
 			break;
 		}
+
+		if (write_pos == 0 && read_map_entry.type == MAP_ENTRY_TYPE_EMPTY)
+			write_pos = prev_pos;
 	}
 
 	/* Initialize map entry and encrypt key */
 	pg_tde_initialize_map_entry(&write_map_entry, principal_key, rlocator, rel_key_data);
 
 	/* Write the given entry at curr_pos; i.e. the free entry. */
-	pg_tde_write_one_map_entry(map_fd, &write_map_entry, &curr_pos, db_map_path);
+	pg_tde_write_one_map_entry(map_fd, &write_map_entry, &write_pos, db_map_path);
 
 	CloseTransientFile(map_fd);
 }

contrib/pg_tde/src/smgr/pg_tde_smgr.c

@@ -113,9 +113,6 @@ tde_smgr_create_key_redo(const RelFileLocator *rlocator)
 {
 	InternalKey key;
 
-	if (pg_tde_has_smgr_key(*rlocator))
-		return;
-
 	pg_tde_generate_internal_key(&key);
 
 	pg_tde_save_smgr_key(*rlocator, &key);
@@ -383,18 +380,18 @@ tde_mdcreate(RelFileLocator relold, SMgrRelation reln, ForkNumber forknum, bool
 	if (forknum != MAIN_FORKNUM)
 		return;
 
-	/*
-	 * If we have a key for this relation already, we need to remove it. This
-	 * can happen if OID is re-used after a crash left a key for a
-	 * non-existing relation in the key file.
-	 *
-	 * If we're in redo, a separate WAL record will make sure the key is
-	 * removed.
-	 */
-	tde_smgr_delete_leftover_key(&reln->smgr_rlocator);
-
 	if (!tde_smgr_should_encrypt(&reln->smgr_rlocator, &relold))
 	{
+		/*
+		 * If we have a key for this relation already, we need to remove it.
+		 * This can happen if OID is re-used after a crash left a key for a
+		 * non-existing relation in the key file.
+		 *
+		 * Old keys for encrypted tables are replace when creating the new
+		 * key.
+		 */
+		tde_smgr_delete_leftover_key(&reln->smgr_rlocator);
+
 		tdereln->encryption_status = RELATION_NOT_ENCRYPTED;
 		return;
 	}