os/FileStore: do not loop in sync_entry on shutdown

Normally we repeat a sync if the journal is full and asks for it or if we
have pending sync waiters.  This isn't for correctness: it's to move things
along quickly.

If we are shutting down *and* the journal is full, however, we currently
can get stuck in a loop where we repeatedly sync.  However, the journal
never gets unfull because we aren't expiring anything.

We tried to fix this in 682b9daacb1960eb60b5f0e9b40d4a3a3e9afd5b but that
didn't work because we don't tell the journal to stop before we shut down
the filestore sync thread.

I'm pretty sure there is another issue in the journal code as

 full_state == FULL_FULL
 submit_manager.op_seq = 14761
 apply_manager.committing_seq = 14761
 apply_manager.committed_seq = 14761

which seems odd.  But I have no longs.  Meanwhile, this fixes the hang.

Fixes: #9545
Signed-off-by: Sage Weil <sage@redhat.com>

diff --git a/src/os/FileStore.cc b/src/os/FileStore.cc
index 99a2502aeae165b50686023308aa32f864f84fc6..1bb4de96cdd211e0f869df31d08edb1470e03bbd 100644 (file)
--- a/src/os/FileStore.cc
+++ b/src/os/FileStore.cc
@@ -3533,7 +3533,7 @@ void FileStore::sync_entry()
       dout(10) << "sync_entry more waiters, committing again" << dendl;
       goto again;
     }
-    if (journal && journal->should_commit_now()) {
+    if (!stop && journal && journal->should_commit_now()) {
       dout(10) << "sync_entry journal says we should commit again (probably is/was full)" << dendl;
       goto again;
     }