nfsd4: fix a deadlock on state owner replay mutex
diff mbox series

Message ID
State New
Headers show
  • nfsd4: fix a deadlock on state owner replay mutex
Related show

Commit Message

黄乐 June 27, 2019, 6:16 p.m. UTC
from: Huang Le <>

In move_to_close_lru(), which only be called on path of nfsd4 CLOSE op,
the code could wait for its stid ref count drop to 2 while holding its
state owner replay mutex.  However, the other stid ref holder (normally
a parallel CLOSE op) that move_to_close_lru() is waiting for might be
accquiring the same replay mutex.

This patch fix the issue by clearing the replay owner before waiting, and
assign it back after then.

Signed-off-by: Huang Le <>

I guess we should cc this patch to stable tree, since a malicious client
could craft parallel CLOSE ops to put all nfsd tasks in D state shortly.

diff mbox series

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index 618e660..5f6a48f 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -3829,12 +3829,12 @@  static void nfs4_free_openowner(struct nfs4_stateowner *so)
  * them before returning however.
 static void
-move_to_close_lru(struct nfs4_ol_stateid *s, struct net *net)
+move_to_close_lru(struct nfsd4_compound_state *cstate, struct nfs4_ol_stateid *s,
+		struct net *net)
 	struct nfs4_ol_stateid *last;
 	struct nfs4_openowner *oo = openowner(s->st_stateowner);
-	struct nfsd_net *nn = net_generic(s->st_stid.sc_client->net,
-						nfsd_net_id);
+	struct nfsd_net *nn = net_generic(net, nfsd_net_id);
 	dprintk("NFSD: move_to_close_lru nfs4_openowner %p\n", oo);
@@ -3846,8 +3846,19 @@  static void nfs4_free_openowner(struct nfs4_stateowner *so)
 	 * Wait for the refcount to drop to 2. Since it has been unhashed,
 	 * there should be no danger of the refcount going back up again at
 	 * this point.
+	 *
+	 * Before waiting, we clear cstate->replay_owner to release its
+	 * so_replay.rp_mutex, since other reference holder might be accquiring
+	 * the same mutex before they could drop the references.  The replay_owner
+	 * can be assigned back safely after they done their jobs.
-	wait_event(close_wq, refcount_read(&s->st_stid.sc_count) == 2);
+	if (refcount_read(&s->st_stid.sc_count) != 2) {
+		struct nfs4_stateowner *so = cstate->replay_owner;
+		nfsd4_cstate_clear_replay(cstate);
+		wait_event(close_wq, refcount_read(&s->st_stid.sc_count) == 2);
+		nfsd4_cstate_assign_replay(cstate, so);
+	}
 	if (s->st_stid.sc_file) {
@@ -5531,7 +5542,8 @@  static inline void nfs4_stateid_downgrade(struct nfs4_ol_stateid *stp, u32 to_ac
 	return status;
-static void nfsd4_close_open_stateid(struct nfs4_ol_stateid *s)
+static void nfsd4_close_open_stateid(struct nfsd4_compound_state *cstate,
+		struct nfs4_ol_stateid *s)
 	struct nfs4_client *clp = s->st_stid.sc_client;
 	bool unhashed;
@@ -5549,7 +5561,7 @@  static void nfsd4_close_open_stateid(struct nfs4_ol_stateid *s)
 		if (unhashed)
-			move_to_close_lru(s, clp->net);
+			move_to_close_lru(cstate, s, clp->net);
@@ -5587,7 +5599,7 @@  static void nfsd4_close_open_stateid(struct nfs4_ol_stateid *s)
 	nfs4_inc_and_copy_stateid(&close->cl_stateid, &stp->st_stid);
-	nfsd4_close_open_stateid(stp);
+	nfsd4_close_open_stateid(cstate, stp);
 	/* v4.1+ suggests that we send a special stateid in here, since the