client: flush the mdlog in _fsync before waiting on unstable reqs

Recently, we changed ganesha to call to ceph_ll_sync_inode after
issuing a setattr call, to ensure that the attribute changes makes it
to the backing store before returning.

While testing with this, I noticed that we'd occasionally see setattr
calls stall for a long time. Analysis showed that the client's _fsync
request was stalled waiting on an mdlog flush.

This patch adds a call to flush_mdlog_sync prior to waiting on unsafe
requests in _fsync and seems to resolve the issue.

Tracker: http://tracker.ceph.com/issues/23714
Signed-off-by: Jeff Layton <jlayton@redhat.com>

diff --git a/src/client/Client.cc b/src/client/Client.cc
index c529535b11500153ad921b43670e0048b381ce63..9f82607609b90757728c58ec6300d00b3ede8b42 100644 (file)
--- a/src/client/Client.cc
+++ b/src/client/Client.cc
@@ -9397,6 +9397,8 @@ int Client::_fsync(Inode *in, bool syncdataonly)
   } else ldout(cct, 10) << "no metadata needs to commit" << dendl;
 
   if (!syncdataonly && !in->unsafe_ops.empty()) {
+    flush_mdlog_sync();
+
     MetaRequest *req = in->unsafe_ops.back();
     ldout(cct, 15) << "waiting on unsafe requests, last tid " << req->get_tid() <<  dendl;