Re: IORING_REGISTER_CREDS[_UPDATE]() and credfd_create()?

[Jens Axboe <axboe@kernel.dk>]

On 1/28/20 1:50 PM, Pavel Begunkov wrote:
> On 28/01/2020 23:19, Jens Axboe wrote:
>> On 1/28/20 1:16 PM, Pavel Begunkov wrote:
>>> On 28/01/2020 22:42, Jens Axboe wrote:
>>>> On 1/28/20 11:04 AM, Jens Axboe wrote:
>>>>> On 1/28/20 10:19 AM, Jens Axboe wrote:
>>>>>> On 1/28/20 9:19 AM, Jens Axboe wrote:
>>>>>>> On 1/28/20 9:17 AM, Stefan Metzmacher wrote:
>>>>>> OK, so here are two patches for testing:
>>>>>>
>>>>>> https://git.kernel.dk/cgit/linux-block/log/?h=for-5.6/io_uring-vfs-creds
>>>>>>
>>>>>> #1 adds support for registering the personality of the invoking task,
>>>>>> and #2 adds support for IORING_OP_USE_CREDS. Right now it's limited to
>>>>>> just having one link, it doesn't support a chain of them.
>>>>>>
>>>>>> I'll try and write a test case for this just to see if it actually works,
>>>>>> so far it's totally untested. 
>>>>>>
>>>>>> Adding Pavel to the CC.
>>>>>
>>>>> Minor tweak to ensuring we do the right thing for async offload as well,
>>>>> and it tests fine for me. Test case is:
>>>>>
>>>>> - Run as root
>>>>> - Register personality for root
>>>>> - create root only file
>>>>> - check we can IORING_OP_OPENAT the file
>>>>> - switch to user id test
>>>>> - check we cannot IORING_OP_OPENAT the file
>>>>> - check that we can open the file with IORING_OP_USE_CREDS linked
>>>>
>>>> I didn't like it becoming a bit too complicated, both in terms of
>>>> implementation and use. And the fact that we'd have to jump through
>>>> hoops to make this work for a full chain.
>>>>
>>>> So I punted and just added sqe->personality and IOSQE_PERSONALITY.
>>>> This makes it way easier to use. Same branch:
>>>>
>>>> https://git.kernel.dk/cgit/linux-block/log/?h=for-5.6/io_uring-vfs-creds
>>>>
>>>> I'd feel much better with this variant for 5.6.
>>>>
>>>
>>> To be honest, sounds pretty dangerous. Especially since somebody started talking
>>> about stealing fds from a process, it could lead to a nasty loophole somehow.
>>> E.g. root registers its credentials, passes io_uring it to non-privileged
>>> children, and then some process steals the uring fd (though, it would need
>>> priviledged mode for code-injection or else). Could we Cc here someone really
>>> keen on security?
>>
>> Link? If you can steal fds, then surely you've already lost any sense of
> 
> https://lwn.net/Articles/808997/
> But I didn't looked up it yet.

This isn't new by any stretch, it's always been possible to pass file
descriptors through SCM_RIGHTS. This just gives you a new way to do it.
That's not stealing or leaking, it's deliberately passing it to someone
else.

>> security in the first place? Besides, if root registered the ring, the root
>> credentials are already IN the ring. I don't see how this adds any extra
>> holes.
> 
> Isn't it what you fixed in ("don't use static creds/mm assignments") ?

Sure, but SQPOLL still uses it.

> I'm not sure what capability (and whether any) it would need, but
> better to think such cases through. Just saying, I would prefer to ask
> a person with extensive security experience, unlike me.

I don't disagree, but I really don't think this is any different than
what we already allow.

-- 
Jens Axboe