nfsd4: extend state lock over seqid replay logic

There are currently a couple races in the seqid replay code: a
retransmission could come while we're still encoding the original reply,
or a new seqid-mutating call could come as we're encoding a replay.

So, extend the state lock over the encoding (both encoding of a replayed
reply and caching of the original encoded reply).

I really hate doing this, and previously added the stateowner
reference-counting code to avoid it (which was insufficient)--but I
don't see a less complicated alternative at the moment.

Signed-off-by: J. Bruce Fields <bfields@redhat.com>
diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index bc1a9db..6cf729a 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -3501,7 +3501,8 @@
 
 	nfsd4_create_clid_dir(sop->so_client);
 out:
-	nfs4_unlock_state();
+	if (!cstate->replay_owner)
+		nfs4_unlock_state();
 	return status;
 }
 
@@ -3568,7 +3569,8 @@
 	memcpy(&od->od_stateid, &stp->st_stateid, sizeof(stateid_t));
 	status = nfs_ok;
 out:
-	nfs4_unlock_state();
+	if (!cstate->replay_owner)
+		nfs4_unlock_state();
 	return status;
 }
 
@@ -3609,7 +3611,8 @@
 	if (list_empty(&so->so_stateids))
 		move_to_close_lru(so);
 out:
-	nfs4_unlock_state();
+	if (!cstate->replay_owner)
+		nfs4_unlock_state();
 	return status;
 }
 
@@ -4071,7 +4074,8 @@
 out:
 	if (status && lock->lk_is_new && lock_sop)
 		release_lockowner(lock_sop);
-	nfs4_unlock_state();
+	if (!cstate->replay_owner)
+		nfs4_unlock_state();
 	return status;
 }