Re: [PATCH net-next] sockopt: Change getsockopt() of SO_BINDTODEVICE to return an interface name

[Brian Haley <brian.haley@hp.com>]

On 11/02/2012 05:34 PM, Ben Hutchings wrote:
> On Fri, 2012-11-02 at 11:02 -0400, Brian Haley wrote:
>> On 11/02/2012 05:36 AM, Pavel Emelyanov wrote:
>>>> +static int sock_getbindtodevice(struct sock *sk, char __user *optval,
>>>> +				int __user *optlen, int len)
>>>> +{
>>>> +	int ret = -ENOPROTOOPT;
>>>> +#ifdef CONFIG_NETDEVICES
>>>> +	struct net *net = sock_net(sk);
>>>> +	struct net_device *dev;
>>>> +	char devname[IFNAMSIZ];
>>>> +
>>>> +	if (sk->sk_bound_dev_if == 0) {
>>>> +		len = 0;
>>>> +		goto zero;
>>>> +	}
>>>> +
>>>> +	ret = -EINVAL;
>>>> +	if (len < IFNAMSIZ)
>>>> +		goto out;
>>>> +
>>>> +	rcu_read_lock();
>>>> +	dev = dev_get_by_index_rcu(net, sk->sk_bound_dev_if);
>>>> +	if (dev)
>>>> +		strcpy(devname, dev->name);
>>>
>>> This still races with the device name change, potentially providing
>>> a name which never existed in the system, doesn't it?
>>
>> My only argument here is that SIOCGIFNAME has had this same code forever, and
>> noone has ever complained about that returning a garbled name.  Even
>> dev_get_by_name() only holds an rcu lock when doing a strncmp().
>>
>> We'd need to audit the whole kernel to catch all the places where we potentially
>> look at dev->name while it could change.  Is it really worth it?
>
> A net device name can't be changed while the device is up, or while
> another task holds the RTNL lock.  I think that covers almost all uses.
> I don't know whether it's worth going out to look for exceptions, but we
> might as well fix the cases we know about.

So do you think we can fix these corner cases later and get the API 
right first?

-Brian