[PATCH] wireguard: Fix leaking sockets in wg_socket_init error paths

[PATCH] wireguard: Fix leaking sockets in wg_socket_init error paths

[Daniel Gröber]

---
 drivers/net/wireguard/socket.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/drivers/net/wireguard/socket.c b/drivers/net/wireguard/socket.c
index 0414d7a6ce74..c35163f503e7 100644
--- a/drivers/net/wireguard/socket.c
+++ b/drivers/net/wireguard/socket.c
@@ -387,7 +387,7 @@ int wg_socket_init(struct wg_device *wg, u16 port)
 	ret = udp_sock_create(net, &port4, &new4);
 	if (ret < 0) {
 		pr_err("%s: Could not create IPv4 socket\n", wg->dev->name);
-		goto out;
+		goto err;
 	}
 	set_sock_opts(new4);
 	setup_udp_tunnel_sock(net, new4, &cfg);
@@ -402,7 +402,7 @@ int wg_socket_init(struct wg_device *wg, u16 port)
 				goto retry;
 			pr_err("%s: Could not create IPv6 socket\n",
 			       wg->dev->name);
-			goto out;
+			goto err;
 		}
 		set_sock_opts(new6);
 		setup_udp_tunnel_sock(net, new6, &cfg);
@@ -414,6 +414,11 @@ int wg_socket_init(struct wg_device *wg, u16 port)
 out:
 	put_net(net);
 	return ret;
+
+err:
+	sock_free(new4 ? new4->sk : NULL);
+	sock_free(new6 ? new6->sk : NULL);
+	goto out;
 }
 
 void wg_socket_reinit(struct wg_device *wg, struct sock *new4,
-- 
2.39.2

[PATCH] wireguard: Fix leaking sockets in wg_socket_init error paths

[Daniel Gröber]

This doesn't seem to be reachable normally, but while working on a patch
for the address binding code I ended up triggering this leak and had to
reboot to get rid of the leaking wg sockets.
---
 drivers/net/wireguard/socket.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/drivers/net/wireguard/socket.c b/drivers/net/wireguard/socket.c
index 0414d7a6ce74..c35163f503e7 100644
--- a/drivers/net/wireguard/socket.c
+++ b/drivers/net/wireguard/socket.c
@@ -387,7 +387,7 @@ int wg_socket_init(struct wg_device *wg, u16 port)
 	ret = udp_sock_create(net, &port4, &new4);
 	if (ret < 0) {
 		pr_err("%s: Could not create IPv4 socket\n", wg->dev->name);
-		goto out;
+		goto err;
 	}
 	set_sock_opts(new4);
 	setup_udp_tunnel_sock(net, new4, &cfg);
@@ -402,7 +402,7 @@ int wg_socket_init(struct wg_device *wg, u16 port)
 				goto retry;
 			pr_err("%s: Could not create IPv6 socket\n",
 			       wg->dev->name);
-			goto out;
+			goto err;
 		}
 		set_sock_opts(new6);
 		setup_udp_tunnel_sock(net, new6, &cfg);
@@ -414,6 +414,11 @@ int wg_socket_init(struct wg_device *wg, u16 port)
 out:
 	put_net(net);
 	return ret;
+
+err:
+	sock_free(new4 ? new4->sk : NULL);
+	sock_free(new6 ? new6->sk : NULL);
+	goto out;
 }
 
 void wg_socket_reinit(struct wg_device *wg, struct sock *new4,
-- 
2.39.2

Re: [PATCH] wireguard: Fix leaking sockets in wg_socket_init error paths

[Jason A. Donenfeld]

Hi,

The signed-off-by is missing and the subject does not match the format
of any other wireguard commits.

On Mon, Oct 23, 2023 at 03:06:09PM +0200, Daniel Gröber wrote:
> This doesn't seem to be reachable normally, but while working on a patch

"Normally" as in what? At all? Or?

> for the address binding code I ended up triggering this leak and had to
> reboot to get rid of the leaking wg sockets.

This commit message doesn't describe any rationale for this patch. Can
you describe the bug?

> ---
>  drivers/net/wireguard/socket.c | 9 +++++++--
>  1 file changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/net/wireguard/socket.c b/drivers/net/wireguard/socket.c
> index 0414d7a6ce74..c35163f503e7 100644
> --- a/drivers/net/wireguard/socket.c
> +++ b/drivers/net/wireguard/socket.c
> @@ -387,7 +387,7 @@ int wg_socket_init(struct wg_device *wg, u16 port)
>  	ret = udp_sock_create(net, &port4, &new4);
>  	if (ret < 0) {
>  		pr_err("%s: Could not create IPv4 socket\n", wg->dev->name);
> -		goto out;
> +		goto err;

`new4` is either NULL or has already been freed here in the `goto retry`
case. `new6` is NULL here.

>  	}
>  	set_sock_opts(new4);
>  	setup_udp_tunnel_sock(net, new4, &cfg);
> @@ -402,7 +402,7 @@ int wg_socket_init(struct wg_device *wg, u16 port)
>  				goto retry;
>  			pr_err("%s: Could not create IPv6 socket\n",
>  			       wg->dev->name);
> -			goto out;
> +			goto err;

`new4` has just been freed by `udp_tunnel_sock_release` just above the
context. `new6` is NULL.

>  		}
>  		set_sock_opts(new6);
>  		setup_udp_tunnel_sock(net, new6, &cfg);
> @@ -414,6 +414,11 @@ int wg_socket_init(struct wg_device *wg, u16 port)
>  out:
>  	put_net(net);
>  	return ret;
> +
> +err:
> +	sock_free(new4 ? new4->sk : NULL);
> +	sock_free(new6 ? new6->sk : NULL);
> +	goto out;
>  }
>  
>  void wg_socket_reinit(struct wg_device *wg, struct sock *new4,

I don't see the bug. If there is one, maybe try again with a real patch
that describes it better. If there isn't one, what is the point?

Jason

Re: [PATCH] wireguard: Fix leaking sockets in wg_socket_init error paths

[Daniel Gröber]

Hi Jason,

On Mon, Oct 23, 2023 at 04:04:13PM +0200, Jason A. Donenfeld wrote:
> The signed-off-by is missing and the subject does not match the format
> of any other wireguard commits.

Ah, I don't usually send kernel patches. Forgot to do format.signOff=true.

> On Mon, Oct 23, 2023 at 03:06:09PM +0200, Daniel Gröber wrote:
> > This doesn't seem to be reachable normally, but while working on a patch
> 
> "Normally" as in what? At all? Or?

I committed this while working on my address/ifindex binding patch[1]
(which I will also resend shortly), at the time I thought this fix makes
sense in isolation but apparently not.

[1]: https://lists.zx2c4.com/pipermail/wireguard/2023-August/008148.html,

> > for the address binding code I ended up triggering this leak and had to
> > reboot to get rid of the leaking wg sockets.
> 
> This commit message doesn't describe any rationale for this patch. Can
> you describe the bug?

It's been a while since I wrote this patch. Unfortunately you didn't
respond to my initial mail in Aug, so some context has already been lost to
time.

I may have been under the mistaken impression that udp_sock_create can
return <0 while leaving *sockp!=NULL, but as I recall it I did re-test with
this patch and it fixed the bug, that I wish I remembered how to trigger
now. Unsatisfying.

--Daniel