Re: [PATCH 4.4 stable net] net: tcp: Fix use-after-free in tcp_write_xmit

[maowenan <maowenan@huawei.com>]

On 2019/7/25 14:19, Eric Dumazet wrote:
> 
> 
> On 7/25/19 6:29 AM, maowenan wrote:
>>
> 
>>>>>> Syzkaller reproducer():
>>>>>> r0 = socket$packet(0x11, 0x3, 0x300)
>>>>>> r1 = socket$inet_tcp(0x2, 0x1, 0x0)
>>>>>> bind$inet(r1, &(0x7f0000000300)={0x2, 0x4e21, @multicast1}, 0x10)
>>>>>> connect$inet(r1, &(0x7f0000000140)={0x2, 0x1000004e21, @loopback}, 0x10)
>>>>>> recvmmsg(r1, &(0x7f0000001e40)=[{{0x0, 0x0, &(0x7f0000000100)=[{&(0x7f00000005c0)=""/88, 0x58}], 0x1}}], 0x1, 0x40000000, 0x0)
>>>>>> sendto$inet(r1, &(0x7f0000000000)="e2f7ad5b661c761edf", 0x9, 0x8080, 0x0, 0x0)
>>>>>> r2 = fcntl$dupfd(r1, 0x0, r0)
>>>>>> connect$unix(r2, &(0x7f00000001c0)=@file={0x0, './file0\x00'}, 0x6e)
>>>>>>
>>>
>>> It does call tcp_disconnect(), by one of the connect() call.
>>
>> yes, __inet_stream_connect will call tcp_disconnect when sa_family == AF_UNSPEC, in c repro if it
>> passes sa_family with AF_INET it won't call disconnect, and then sk_send_head won't be NULL when tcp_connect.
>>
> 
> 
> Look again at the Syzkaller reproducer()
> 
> It definitely uses tcp_disconnect()
> 
> Do not be fooled by connect$unix(), this is a connect() call really, with AF_UNSPEC

Right, in syzkaller reproducer, it calls connect() with AF_UNSPEC, actually I can reproduce the issue only with C repro(https://syzkaller.appspot.com/text?tag=ReproC&x=14db474f800000).
syscall procedure in C:
__NR_socket
__NR_bind
__NR_sendto  (flag=0x20000000,MSG_FASTOPEN, it will call __inet_stream_connect with sa_family = AF_INET, sk->sk_send_head = NULL)
__NR_write
__NR_connect (call __inet_stream_connect with sa_family = AF_UNSPEC, it will call tcp_disconnect and set sk->sk_send_head = NULL)
__NR_connect (call __inet_stream_connect with sa_family = AF_INET, if sk->sk_send_head != NULL UAF happen)

I debug why tcp_disconnect has already set sk->sk_send_head = NULL, but it is NOT NULL after next __NR_connect.
I find that some packets send out before second __NR_connect(with AF_INET), so the sk_send_head is modified by: tcp_sendmsg->skb_entail->tcp_add_write_queue_tail
static inline void tcp_add_write_queue_tail(struct sock *sk, struct sk_buff *skb)
{
	__tcp_add_write_queue_tail(sk, skb);

	/* Queue it, remembering where we must start sending. */
	if (sk->sk_send_head == NULL) {
		sk->sk_send_head = skb;  //here, sk->sk_send_head is changed.

		if (tcp_sk(sk)->highest_sack == NULL)
			tcp_sk(sk)->highest_sack = skb;
	}
}




> 
>