MDEV-13267 At startup with crash recovery: mtr_t::commit_checkpoint(lsn_t,...

MDEV-13267 At startup with crash recovery: mtr_t::commit_checkpoint(lsn_t, bool): Assertion `!recv_no_log_write' failed

This is a bogus debug assertion failure that should be possible
starting with MariaDB 10.2.2 (which merged WL#7142 via MySQL 5.7.9).

While generating page-change redo log records is strictly out of the
question during tat certain parts of crash recovery, the
fil_names_clear() is only emitting informational MLOG_FILE_NAME
and MLOG_CHECKPOINT records to guarantee that if the server is killed
during or soon after the crash recovery, subsequent crash recovery
will be possible.

The metadata buffer that fil_names_clear() is flushing to the redo log
is being filled by recv_init_crash_recovery_spaces(), right before
starting to apply redo log, by invoking fil_names_dirty() on every
discovered tablespace for which there are changes to apply.

When it comes to Mariabackup (xtrabackup --prepare), it is strictly out
of the question to generate any redo log whatsoever, because that could
break the restore of incremental backups by causing LSN deviation.
So, the fil_names_dirty() call must be skipped when restoring backups.

recv_recovery_from_checkpoint_start(): Do not invoke fil_names_clear()
when restoring a backup.

mtr_t::commit_checkpoint(): Remove the failing assertion. The only
caller is fil_names_clear(), and it must be called by
recv_recovery_from_checkpoint_start() for normal server startup to be
crash-safe. The debug assertion in mtr_t::commit() will still
catch rogue redo log writes.

storage/innobase/log/log0recv.cc

@@ -3126,6 +3126,9 @@ recv_recovery_from_checkpoint_start(lsn_t flush_lsn)
 	byte*		buf;
 	dberr_t		err = DB_SUCCESS;
 
+	ut_ad(srv_operation == SRV_OPERATION_NORMAL
+	      || srv_operation == SRV_OPERATION_RESTORE);
+
 	/* Initialize red-black tree for fast insertions into the
 	flush_list during recovery process. */
 	buf_flush_init_flush_rbt();
@@ -3342,9 +3345,11 @@ recv_recovery_from_checkpoint_start(lsn_t flush_lsn)
 
 	log_sys->last_checkpoint_lsn = checkpoint_lsn;
 
-	if (!srv_read_only_mode) {
+	if (!srv_read_only_mode && srv_operation == SRV_OPERATION_NORMAL) {
 		/* Write a MLOG_CHECKPOINT marker as the first thing,
-		before generating any other redo log. */
+		before generating any other redo log. This ensures
+		that subsequent crash recovery will be possible even
+		if the server were killed soon after this. */
 		fil_names_clear(log_sys->last_checkpoint_lsn, true);
 	}
 

storage/innobase/mtr/mtr0mtr.cc

@@ -599,9 +599,6 @@ mtr_t::commit_checkpoint(
 	ut_d(m_impl.m_state = MTR_STATE_COMMITTING);
 	ut_ad(write_mlog_checkpoint || m_impl.m_n_log_recs > 1);
 
-	/* This is a dirty read, for debugging. */
-	ut_ad(!recv_no_log_write);
-
 	switch (m_impl.m_n_log_recs) {
 	case 0:
 		break;