Re: [syzbot] KASAN: use-after-free Write in sco_sock_timeout

[Desmond Cheong Zhi Xi]

On 29/8/21 4:29 pm, Hillf Danton wrote:
> On Fri, 27 Aug 2021 15:58:34 +0800 Desmond Cheong Zhi Xi wrote:
> > On 27/8/21 9:19 am, Hillf Danton wrote:
> > > On Thu, 26 Aug 2021 09:29:24 -0700
> > > > syzbot found the following issue on:
> > > >
> > > > HEAD commit:    e3f30ab28ac8 Merge branch 'pktgen-samples-next'
> > > > git tree:       net-next
> > > > console output: https://syzkaller.appspot.com/x/log.txt?x=13249c96300000
> > > > kernel config:  https://syzkaller.appspot.com/x/.config?x=ef482942966bf763
> > > > dashboard link: https://syzkaller.appspot.com/bug?extid=2bef95d3ab4daa10155b
> > > > compiler:       gcc (Debian 10.2.1-6) 10.2.1 20210110, GNU ld (GNU Binutils for Debian) 2.35.1
> > > > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=16a29ea9300000
> > > >
> > > > The issue was bisected to:
> > > >
> > > > commit e1dee2c1de2b4dd00eb44004a4bda6326ed07b59
> > > > Author: Desmond Cheong Zhi Xi <desmondcheongzx@xxxxxxxxx>
> > > > Date:   Tue Aug 10 04:14:10 2021 +0000
> > > >
> > > >       Bluetooth: fix repeated calls to sco_sock_kill
>
> To fix the uaf, grab another hold to sock to make the timeout work safe.
>
> #syz test: git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git e3f30ab28ac8
>
> --- a/net/bluetooth/sco.c
> +++ b/net/bluetooth/sco.c
> @@ -190,15 +190,14 @@ static void sco_conn_del(struct hci_conn
>   	sco_conn_unlock(conn);
>   
>   	if (sk) {
> -		sock_hold(sk);
>   		lock_sock(sk);
>   		sco_sock_clear_timer(sk);
>   		sco_chan_del(sk, err);
>   		release_sock(sk);
> -		sock_put(sk);
>   
>   		/* Ensure no more work items will run before freeing conn. */
>   		cancel_delayed_work_sync(&conn->timeout_work);
> +		sock_put(sk);

Hi Hillf,

Saw that this passed the reproducer. But on closer inspection, I think 
what's happening is that sco_conn_del is never run.

So the extra sock_hold prevents a UAF, but that's because now the 
reference count never goes to 0. In my opinion, something closer to your 
previous proposal (+ also addressing other calls to __sco_sock_close) 
where we call cancel_delayed_work_sync after the channel is deleted 
would address the root cause better.

Just my two cents.

>   	}
>   
>   	hcon->sco_data = NULL;
> @@ -212,6 +211,8 @@ static void __sco_chan_add(struct sco_co
>   
>   	sco_pi(sk)->conn = conn;
>   	conn->sk = sk;
> +	/* make timeout_work safe; will be put in  sco_conn_del() */
> +	sock_hold(sk);
>   
>   	INIT_DELAYED_WORK(&conn->timeout_work, sco_sock_timeout);
>   
> --