[Openvpn-devel] ovpn-dco: hold peer ref in ovpn_tcp_read_sock under RCU

Message ID tencent_4CA59D4D873F946422CA8AAEB20AF7C4CF05@qq.com
State New
Headers
Series [Openvpn-devel] ovpn-dco: hold peer ref in ovpn_tcp_read_sock under RCU |

Commit Message

liuchong Nov. 18, 2026, 7 a.m. UTC
  ovpn_tcp_read_sock() uses sock->peer without taking a reference.
ovpn_peer_del() can drop the last hash-table reference while the
TCP data_ready callback is still attached, so peer->refcount may
already be zero when ovpn_peer_hold() is later called for
ovpn_recv(). That leads to a WARN_ON and a refcount underflow when
the async path puts the peer.

Take the peer reference under rcu_read_lock() before leaving the
RCU critical section, and drop it on all exit paths. If the peer
is already being destroyed, stop consuming the stream quietly.

Fixes: 209247cf6d47 ("ovpn-dco: implement TCP transport support")
Signed-off-by: liuchong <979504756@qq.com>
---
 drivers/net/ovpn-dco/tcp.c | 24 ++++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)
  

Comments

Ralf Lici Nov. 18, 2026, 8:05 a.m. UTC | #1
Hi,

Thanks for the patch.

On Wed, 18 Nov 2026 15:00:37 +0800, liuchong via Openvpn-devel <openvpn-devel@lists.sourceforge.net> wrote:
> ovpn_tcp_read_sock() uses sock->peer without taking a reference.
> ovpn_peer_del() can drop the last hash-table reference while the
> TCP data_ready callback is still attached, so peer->refcount may
> already be zero when ovpn_peer_hold() is later called for
> ovpn_recv(). That leads to a WARN_ON and a refcount underflow when
> the async path puts the peer.
>
> Take the peer reference under rcu_read_lock() before leaving the
> RCU critical section, and drop it on all exit paths. If the peer
> is already being destroyed, stop consuming the stream quietly.
>
> Fixes: 209247cf6d47 ("ovpn-dco: implement TCP transport support")
> Signed-off-by: liuchong <979504756@qq.com>

I'm not sure this SoB tag is fully compliant with the Developer
Certificate of Origin. If you can, please use your real name, as the DCO
asks for a known identity.

> ---
>  drivers/net/ovpn-dco/tcp.c | 24 ++++++++++++++++++------
>  1 file changed, 18 insertions(+), 6 deletions(-)
>
> diff --git a/drivers/net/ovpn-dco/tcp.c b/drivers/net/ovpn-dco/tcp.c
> index fe095f9..d8f9afd 100644
> --- a/drivers/net/ovpn-dco/tcp.c
> +++ b/drivers/net/ovpn-dco/tcp.c
> @@ -35,15 +35,26 @@ static int ovpn_tcp_read_sock(read_descriptor_t *desc, struct sk_buff *in_skb,
>  
>  	rcu_read_lock();
>  	sock = rcu_dereference_sk_user_data(sk);
> -	rcu_read_unlock();
> -
>  	if (unlikely(!sock || !sock->peer)) {
> +		rcu_read_unlock();
>  		pr_err("ovpn: read_sock triggered for socket with no metadata\n");
>  		desc->error = -EINVAL;
>  		return 0;
>  	}
>  
> +	/* Take a reference while still under RCU: ovpn_peer_del() may drop the
> +	 * last hash-table reference before socket callbacks are restored in
> +	 * ovpn_peer_release(). Without this hold, a concurrent delete leaves
> +	 * peer->refcount at 0 while this callback still runs, causing
> +	 * ovpn_peer_hold() for ovpn_recv() to fail and a later put to underflow.
> +	 */

The new comment is quite long for what it says, and the commit message
already carries the same explanation.
Something like

    /* hold the peer while still under RCU, it may be getting deleted */

would be enough.

Aside from these two minor nits, neither of which needs a respin on its
own account, the code LGTM.

Reviewed-by: Ralf Lici <ralf@mandelbit.com>

Regards,
  

Patch

diff --git a/drivers/net/ovpn-dco/tcp.c b/drivers/net/ovpn-dco/tcp.c
index fe095f9..d8f9afd 100644
--- a/drivers/net/ovpn-dco/tcp.c
+++ b/drivers/net/ovpn-dco/tcp.c
@@ -35,15 +35,26 @@  static int ovpn_tcp_read_sock(read_descriptor_t *desc, struct sk_buff *in_skb,
 
 	rcu_read_lock();
 	sock = rcu_dereference_sk_user_data(sk);
-	rcu_read_unlock();
-
 	if (unlikely(!sock || !sock->peer)) {
+		rcu_read_unlock();
 		pr_err("ovpn: read_sock triggered for socket with no metadata\n");
 		desc->error = -EINVAL;
 		return 0;
 	}
 
+	/* Take a reference while still under RCU: ovpn_peer_del() may drop the
+	 * last hash-table reference before socket callbacks are restored in
+	 * ovpn_peer_release(). Without this hold, a concurrent delete leaves
+	 * peer->refcount at 0 while this callback still runs, causing
+	 * ovpn_peer_hold() for ovpn_recv() to fail and a later put to underflow.
+	 */
 	peer = sock->peer;
+	if (unlikely(!ovpn_peer_hold(peer))) {
+		rcu_read_unlock();
+		/* Peer is already being destroyed; stop consuming the stream. */
+		return 0;
+	}
+	rcu_read_unlock();
 
 	while (in_len > 0) {
 		/* no skb allocated means that we have to read (or finish reading) the 2 bytes
@@ -103,11 +114,10 @@  static int ovpn_tcp_read_sock(read_descriptor_t *desc, struct sk_buff *in_skb,
 			 * Queue skb for sending to userspace via recvmsg on the socket
 			 */
 			if (likely(ovpn_opcode_from_skb(peer->tcp.skb, 0) == OVPN_DATA_V2)) {
-				/* hold reference to peer as required by ovpn_recv().
+				/* Extra reference required by ovpn_recv() / decrypt_work.
 				 *
-				 * NOTE: in this context we should already be holding a
-				 * reference to this peer, therefore ovpn_peer_hold() is
-				 * not expected to fail
+				 * NOTE: this cannot fail as this function already holds a
+				 * reference to the peer.
 				 */
 				WARN_ON(!ovpn_peer_hold(peer));
 				status = ovpn_recv(peer->ovpn, peer, peer->tcp.skb);
@@ -140,10 +150,12 @@  next_read:
 		copied += chunk;
 	}
 
+	ovpn_peer_put(peer);
 	return copied;
 err:
 	netdev_err(peer->ovpn->dev, "cannot process incoming TCP data: %d\n", desc->error);
 	ovpn_peer_del(peer, OVPN_DEL_PEER_REASON_TRANSPORT_ERROR);
+	ovpn_peer_put(peer);
 	return 0;
 }