Re: [PATCH net v6] net: rose: fix null-ptr-deref caused by rose_kill_by_neigh

[Paolo Abeni <pabeni@redhat.com>]

On Mon, 2022-07-11 at 09:31 +0800, Duoming Zhou wrote:
> When the link layer connection is broken, the rose->neighbour is
> set to null. But rose->neighbour could be used by rose_connection()
> and rose_release() later, because there is no synchronization among
> them. As a result, the null-ptr-deref bugs will happen.
> 
> One of the null-ptr-deref bugs is shown below:
> 
>     (thread 1)                  |        (thread 2)
>                                 |  rose_connect
> rose_kill_by_neigh              |    lock_sock(sk)
>   spin_lock_bh(&rose_list_lock) |    if (!rose->neighbour)
>   rose->neighbour = NULL;//(1)  |
>                                 |    rose->neighbour->use++;//(2)
> 
> The rose->neighbour is set to null in position (1) and dereferenced
> in position (2).
> 
> The KASAN report triggered by POC is shown below:
> 
> KASAN: null-ptr-deref in range [0x0000000000000028-0x000000000000002f]
> ...
> RIP: 0010:rose_connect+0x6c2/0xf30
> RSP: 0018:ffff88800ab47d60 EFLAGS: 00000206
> RAX: 0000000000000005 RBX: 000000000000002a RCX: 0000000000000000
> RDX: ffff88800ab38000 RSI: ffff88800ab47e48 RDI: ffff88800ab38309
> RBP: dffffc0000000000 R08: 0000000000000000 R09: ffffed1001567062
> R10: dfffe91001567063 R11: 1ffff11001567061 R12: 1ffff11000d17cd0
> R13: ffff8880068be680 R14: 0000000000000002 R15: 1ffff11000d17cd0
> ...
> Call Trace:
>   <TASK>
>   ? __local_bh_enable_ip+0x54/0x80
>   ? selinux_netlbl_socket_connect+0x26/0x30
>   ? rose_bind+0x5b0/0x5b0
>   __sys_connect+0x216/0x280
>   __x64_sys_connect+0x71/0x80
>   do_syscall_64+0x43/0x90
>   entry_SYSCALL_64_after_hwframe+0x46/0xb0
> 
> This patch adds lock_sock() in rose_kill_by_neigh() in order to
> synchronize with rose_connect() and rose_release(). Then, changing
> type of 'neighbour->use' from unsigned short to atomic_t in order to
> mitigate race conditions caused by holding different socket lock while
> updating 'neighbour->use'.
> 
> Meanwhile, this patch adds sock_hold() protected by rose_list_lock
> that could synchronize with rose_remove_socket() in order to mitigate
> UAF bug caused by lock_sock() we add.
> 
> What's more, there is no need using rose_neigh_list_lock to protect
> rose_kill_by_neigh(). Because we have already used rose_neigh_list_lock
> to protect the state change of rose_neigh in rose_link_failed(), which
> is well synchronized.
> 
> Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2")
> Signed-off-by: Duoming Zhou <duoming@zju.edu.cn>
> ---
> Changes in v6:
>   - Change sk_for_each() to sk_for_each_safe().
>   - Change type of 'neighbour->use' from unsigned short to atomic_t.
> 
>  include/net/rose.h    |  2 +-
>  net/rose/af_rose.c    | 19 +++++++++++++------
>  net/rose/rose_in.c    | 12 ++++++------
>  net/rose/rose_route.c | 24 ++++++++++++------------
>  net/rose/rose_timer.c |  2 +-
>  5 files changed, 33 insertions(+), 26 deletions(-)
> 
> diff --git a/include/net/rose.h b/include/net/rose.h
> index 0f0a4ce0fee..d5ddebc556d 100644
> --- a/include/net/rose.h
> +++ b/include/net/rose.h
> @@ -95,7 +95,7 @@ struct rose_neigh {
>  	ax25_cb			*ax25;
>  	struct net_device		*dev;
>  	unsigned short		count;
> -	unsigned short		use;
> +	atomic_t		use;
>  	unsigned int		number;
>  	char			restarted;
>  	char			dce_mode;
> diff --git a/net/rose/af_rose.c b/net/rose/af_rose.c
> index bf2d986a6bc..54e7b76c4f3 100644
> --- a/net/rose/af_rose.c
> +++ b/net/rose/af_rose.c
> @@ -163,16 +163,23 @@ static void rose_remove_socket(struct sock *sk)
>  void rose_kill_by_neigh(struct rose_neigh *neigh)
>  {
>  	struct sock *s;
> +	struct hlist_node *tmp;
>  
>  	spin_lock_bh(&rose_list_lock);
> -	sk_for_each(s, &rose_list) {
> +	sk_for_each_safe(s, tmp, &rose_list) {
>  		struct rose_sock *rose = rose_sk(s);
>  
> +		sock_hold(s);
> +		spin_unlock_bh(&rose_list_lock);
> +		lock_sock(s);
>  		if (rose->neighbour == neigh) {
>  			rose_disconnect(s, ENETUNREACH, ROSE_OUT_OF_ORDER, 0);
> -			rose->neighbour->use--;
> +			atomic_dec(&rose->neighbour->use);
>  			rose->neighbour = NULL;
>  		}
> +		release_sock(s);
> +		sock_put(s);

I'm sorry, this does not work. At this point both 's' and 'tmp' sockets
can be freed and reused. Both iterators are not valid anymore when you
acquire the 'rose_list_lock' later.

I really think you should resort to something similar to the following
(completelly untested, just to give an idea). In any case it would be
better to split this change in 2 separate patches: the first patch
replaces 'int use;' with an antomic_t and the 2nd one addresses the
race you describe above.

---
diff --git a/net/rose/af_rose.c b/net/rose/af_rose.c
index bf2d986a6bc3..27b1027aaedf 100644
--- a/net/rose/af_rose.c
+++ b/net/rose/af_rose.c
@@ -156,25 +156,45 @@ static void rose_remove_socket(struct sock *sk)
 	spin_unlock_bh(&rose_list_lock);
 }
 
+static DEFINE_MUTEX(kill_lock);
+
 /*
  *	Kill all bound sockets on a broken link layer connection to a
  *	particular neighbour.
  */
 void rose_kill_by_neigh(struct rose_neigh *neigh)
 {
-	struct sock *s;
+	HLIST_HEAD(rose_list_copy);
+	struct sock *s, *tmp;
+
+	mutex_lock(&kill_lock);
 
 	spin_lock_bh(&rose_list_lock);
 	sk_for_each(s, &rose_list) {
+		sock_hold(s);
+		/* sk_bind_node is apparently unused by rose. Alternatively
+		 * you can add another hlist_node to rose_sock and use it here
+		 */
+		sk_add_bind_node(s, &rose_list_copy);
+	}
+	spin_unlock_bh(&rose_list_lock);
+
+	hlist_for_each_entry_safe(s, tmp, &rose_list_copy, sk_bind_node) {
 		struct rose_sock *rose = rose_sk(s);
 
+		__sk_del_bind_node(s);
+		lock_sock(s);
 		if (rose->neighbour == neigh) {
 			rose_disconnect(s, ENETUNREACH, ROSE_OUT_OF_ORDER, 0);
-			rose->neighbour->use--;
+			atomic_dec(&rose->neighbour->use);
 			rose->neighbour = NULL;
 		}
+		release_sock(s);
+
+		sock_put(s);
 	}
-	spin_unlock_bh(&rose_list_lock);
+
+	mutex_unlock(&kill_lock);
 }
 
 /*
---
/P