Re: [PATCH 9/9] netfilter: implement xt_cgroup cgroup2 path match

[Daniel Borkmann <daniel-FeC+5ew28dpmcu3hnIyYJQ@public.gmane.org>]

On 11/23/2015 02:43 PM, Daniel Borkmann wrote:
> On 11/21/2015 07:54 PM, Florian Westphal wrote:
>> Tejun Heo <tj-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org> wrote:
>>> On Sat, Nov 21, 2015 at 05:56:06PM +0100, Florian Westphal wrote:
>>>>> +struct xt_cgroup_info_v1 {
>>>>> +    __u8        has_path;
>>>>> +    __u8        has_classid;
>>>>> +    __u8        invert_path;
>>>>> +    __u8        invert_classid;
>>>>> +    char        path[PATH_MAX];
>>>>> +    __u32        classid;
>>>>> +
>>>>> +    /* kernel internal data */
>>>>> +    void        *priv __attribute__((aligned(8)));
>>>>> +};
>>>>
>>>> Ahem.  Am I reading this right? This struct is > 4k in size?
>>>> If so -- Ugh.  Does sizeof(path) really have to be PATH_MAX?
>>>
>>> Hmmm... yeap but would this be an acutual problem?
>>
>> Since rule blob can be allocated via vmalloc i guess "no", its not
>> really a problem unless someone needs realy insane amount of such rules.
>>
>> I don't have any better suggestion, so I guess its necessary evil.
>>
>> The only other question I have is wheter PATH_MAX might be a possible
>> ABI breaker in future.  It would have to be guaranteed that this is the
>> same size forever, else you'd get strange errors on rule insertion if
>> the sizes of the kernel and userspace version differs.
>
> Haven't looked deeply into kernfs, but if it's possible to get the object
> from the struct file eventually, you could let iptables frontend open that
> path and just pass the fd down. Would be sizeof(int) vs PATH_MAX then, i.e.
> when you have a large number of rules to load.

( ... but with the downside that things like save/restore wouldn't work. )