Skip to contentSkip to conversation

Question about nfs4_destroy_session()

DISCUSSION Paul E. McKenney 2 participants 3 messages Feb 12, 2014

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

Hello, Trond,

In nfs4_destroy_session(), there is an rcu_dereference() that looks to leak the returned pointer out of an RCU read-side critical section. If the pointed-to object might have just now been created, this is a bug because xprt_destroy_backchannel() dereferences this pointer.

So, does xprt_destroy_backchannel() exclude creation-side code? (If so, no bug -- but a comment might be good.)

Thanx, Paul

void nfs4_destroy_session(struct nfs4_session *session)

{
	struct rpc_xprt *xprt;
	struct rpc_cred *cred;

	cred = nfs4_get_clid_cred(session->clp);
	nfs4_proc_destroy_session(session, cred);
	if (cred)
		put_rpccred(cred);

	rcu_read_lock();
	xprt = rcu_dereference(session->clp->cl_rpcclient->cl_xprt);
	rcu_read_unlock();
	dprintk("%s Destroy backchannel for xprt %p\n",
		__func__, xprt);
	xprt_destroy_backchannel(xprt, NFS41_BC_MIN_CALLBACKS);
	nfs4_destroy_session_slot_tables(session);
	kfree(session);
}
#1 lore raw
Quoted 23 lines · Paul E. McKenney

Hello, Trond,

In nfs4_destroy_session(), there is an rcu_dereference() that looks to leak the returned pointer out of an RCU read-side critical section. If the pointed-to object might have just now been created, this is a bug because xprt_destroy_backchannel() dereferences this pointer.

So, does xprt_destroy_backchannel() exclude creation-side code? (If so, no bug -- but a comment might be good.)

Thanx, Paul

void nfs4_destroy_session(struct nfs4_session *session)

{
	struct rpc_xprt *xprt;
	struct rpc_cred *cred;

	cred = nfs4_get_clid_cred(session->clp);
	nfs4_proc_destroy_session(session, cred);
	if (cred)
		put_rpccred(cred);

	rcu_read_lock();
Quoted 9 lines
	xprt = rcu_dereference(session->clp->cl_rpcclient->cl_xprt);
	rcu_read_unlock();
	dprintk("%s Destroy backchannel for xprt %p\n",
		__func__, xprt);
	xprt_destroy_backchannel(xprt, NFS41_BC_MIN_CALLBACKS);
	nfs4_destroy_session_slot_tables(session);
	kfree(session);
}

Hi Paul,

nfs4_destroy_session() is only called when we’re tearing down the struct nfs_client that owns the cl_rppcclient, and the associated cl_xprt, so the code above should be safe, despite being ugly.

Is there a better annotation for use in the above kind of situation?

Cheers,
Trond

_________________________________
Trond Myklebust
Linux NFS client maintainer, PrimaryData
[email protected]

#2 lore raw
Quoted 36 lines · Trond Myklebust
Quoted 23 lines · Paul E. McKenney

Hello, Trond,

In nfs4_destroy_session(), there is an rcu_dereference() that looks to leak the returned pointer out of an RCU read-side critical section. If the pointed-to object might have just now been created, this is a bug because xprt_destroy_backchannel() dereferences this pointer.

So, does xprt_destroy_backchannel() exclude creation-side code? (If so, no bug -- but a comment might be good.)

Thanx, Paul

void nfs4_destroy_session(struct nfs4_session *session)

{
	struct rpc_xprt *xprt;
	struct rpc_cred *cred;

	cred = nfs4_get_clid_cred(session->clp);
	nfs4_proc_destroy_session(session, cred);
	if (cred)
		put_rpccred(cred);

	rcu_read_lock();
Quoted 9 lines
	xprt = rcu_dereference(session->clp->cl_rpcclient->cl_xprt);
	rcu_read_unlock();
	dprintk("%s Destroy backchannel for xprt %p\n",
		__func__, xprt);
	xprt_destroy_backchannel(xprt, NFS41_BC_MIN_CALLBACKS);
	nfs4_destroy_session_slot_tables(session);
	kfree(session);
}
Quoted 5 lines

Hi Paul,

nfs4_destroy_session() is only called when we’re tearing down the struct nfs_client that owns the cl_rppcclient, and the associated cl_xprt, so the code above should be safe, despite being ugly.

Is there a better annotation for use in the above kind of situation?

One approach would be to add a comment on the rcu_dereference() stating that creation-side code is excluded, e.g., via locking or by the data structures no longer being accessible. Another approach would be to move the rcu_read_unlock() to follow the xprt_destroy_backchannel(), assuming none of the code that would be pulled into the RCU read-side critical section can block.

The second approach would prevent false positives from the RCU pointer leak detectors that are being worked on.

Thanx, Paul