Re: [RFCv2 1/9] tcp: authopt: Initial support and key management

[Dmitry Safonov]

Hi David,

On 8/11/21 6:15 PM, David Ahern wrote:
> On 8/11/21 8:31 AM, Dmitry Safonov wrote:
>> On 8/11/21 9:29 AM, Leonard Crestez wrote:
>>> On 8/10/21 11:41 PM, Dmitry Safonov wrote:
[..]
>>>> I'm pretty sure it's not a good choice to write partly tcp_authopt.
>>>> And a user has no way to check what's the correct len on this kernel.
>>>> Instead of len = min_t(unsigned int, len, sizeof(info)), it should be
>>>> if (len != sizeof(info))
>>>>      return -EINVAL;
>>>
>>> Purpose is to allow sockopts to grow as md5 has grown.
>>
>> md5 has not grown. See above.
> 
> MD5 uapi has - e.g., 8917a777be3ba and  6b102db50cdde. We want similar
> capabilities for growth with this API.

So, you mean adding a new setsockopt when the struct has to be extended?
Like TCP_AUTHOPT_EXT?

It can work, but sounds like adding a new syscall every time the old one
can't be extended. I can see that with current limitations on TCP-AO RFC
the ABI in these patches will have to be extended.

The second commit started using new cmd.tcpm_flags, where unknown flags
are still at this moment silently ignored by the kernel. So 6b102db50cdd
could have introduced a regression in userspace. By luck and by reason
that md5 isn't probably frequently used it didn't.
Not nice at all example for newer APIs.

>> Another issue with your approach
>>
>> +       /* If userspace optlen is too short fill the rest with zeros */
>> +       if (optlen > sizeof(opt))
>> +               return -EINVAL;
>> +       memset(&opt, 0, sizeof(opt));
>> +       if (copy_from_sockptr(&opt, optval, optlen))
>> +               return -EFAULT;
>>
>> is that userspace compiled with updated/grew structure will fail on
>> older kernel. So, no extension without breaking something is possible.
>> Which also reminds me that currently you don't validate (opt.flags) for
>> unknown by kernel flags.
>>
>> Extending syscalls is impossible without breaking userspace if ABI is
>> not designed with extensibility in mind. That was quite a big problem,
>> and still is. Please, read this carefully:
>> https://lwn.net/Articles/830666/
>>
>> That is why I'm suggesting you all those changes that will be harder to
>> fix when/if your patches get accepted.
>> As an example how it should work see in copy_clone_args_from_user().
>>
> 
> Look at how TCP_ZEROCOPY_RECEIVE has grown over releases as an example
> of how to properly handle this.

Exactly.

: switch (len) {
:		case offsetofend(...)
:		case offsetofend(...)

And than also:
:		if (unlikely(len > sizeof(zc))) {
:			err = check_zeroed_user(optval + sizeof(zc),
:						len - sizeof(zc));

Does it sound similar to what I've written in my ABI review?
And what the LWN article has in it.
Please, look again at the patch I replied to.

Thanks,
         Dmitry