mds: flush mds log before finishing STATE_STARTING

If we donn't flush mds log before requesting STATE_ACTIVE, and
mds happens to stop later before the log reaches journal.
The take-over mds will have no SubtreeMap to replay, and fail
later at non-empty subtree check.

Fixes: https://tracker.ceph.com/issues/65094
Signed-off-by: ethanwu <ethanwu@synology.com>
(cherry picked from commit ee5472edd6fe80ed043c3eb73322348367f9bb22)

diff --git a/src/mds/MDSRank.cc b/src/mds/MDSRank.cc
index 9fd50b477ded21ddfcfb1b088c9205fd7913048c..c2f5b47f25421440803a23762d976bb8bf4f753b 100644 (file)
--- a/src/mds/MDSRank.cc
+++ b/src/mds/MDSRank.cc
@@ -1724,7 +1724,10 @@ void MDSRank::boot_start(BootStep step, int r)
       } else {
         dout(2) << "Booting: " << step << ": positioning at end of old mds log" << dendl;
         mdlog->append();
-        starting_done();
+        auto sle = mdcache->create_subtree_map();
+        mdlog->submit_entry(sle);
+        mdlog->flush();
+        mdlog->wait_for_safe(new C_MDS_VoidFn(this, &MDSRank::starting_done));
       }
       break;
     case MDS_BOOT_REPLAY_DONE:
@@ -1771,9 +1774,6 @@ void MDSRank::starting_done()
   ceph_assert(is_starting());
   request_state(MDSMap::STATE_ACTIVE);
 
-  auto sle = mdcache->create_subtree_map();
-  mdlog->submit_entry(sle);
-
   // sync snaptable cache
   snapclient->sync(new C_MDSInternalNoop);
 }