[PATCH 3/7] vfs: Add a mount-notification facility

[Casey Schaufler]

On 5/29/2019 12:47 PM, Jann Horn wrote:
> On Wed, May 29, 2019 at 9:28 PM Casey Schaufler <casey at schaufler-ca.com> wrote:
>> On 5/29/2019 11:11 AM, Jann Horn wrote:
>>> On Wed, May 29, 2019 at 7:46 PM Casey Schaufler <casey at schaufler-ca.com> wrote:
>>>> On 5/29/2019 10:13 AM, Andy Lutomirski wrote:
>>>>>> On May 29, 2019, at 8:53 AM, Casey Schaufler <casey at schaufler-ca.com> wrote:
>>>>>>> On 5/29/2019 4:00 AM, David Howells wrote:
>>>>>>> Jann Horn <jannh at google.com> wrote:
>>>>>>>
>>>>>>>>> +void post_mount_notification(struct mount *changed,
>>>>>>>>> +                            struct mount_notification *notify)
>>>>>>>>> +{
>>>>>>>>> +       const struct cred *cred = current_cred();
>>>>>>>> This current_cred() looks bogus to me. Can't mount topology changes
>>>>>>>> come from all sorts of places? For example, umount_mnt() from
>>>>>>>> umount_tree() from dissolve_on_fput() from __fput(), which could
>>>>>>>> happen pretty much anywhere depending on where the last reference gets
>>>>>>>> dropped?
>>>>>>> IIRC, that's what Casey argued is the right thing to do from a security PoV.
>>>>>>> Casey?
>>>>>> You need to identify the credential of the subject that triggered
>>>>>> the event. If it isn't current_cred(), the cred needs to be passed
>>>>>> in to post_mount_notification(), or derived by some other means.
>>>>> Taking a step back, why do we care who triggered the event?  It seems to me that we should care whether the event happened and whether the *receiver* is permitted to know that.
>>>> There are two filesystems, "dot" and "dash". I am not allowed
>>>> to communicate with Fred on the system, and all precautions have
>>>> been taken to ensure I cannot. Fred asks for notifications on
>>>> all mount activity. I perform actions that result in notifications
>>>> on "dot" and "dash". Fred receives notifications and interprets
>>>> them using Morse code. This is not OK. If Wilma, who *is* allowed
>>>> to communicate with Fred, does the same actions, he should be
>>>> allowed to get the messages via Morse.
>>> In other words, a classic covert channel. You can't really prevent two
>>> cooperating processes from communicating through a covert channel on a
>>> modern computer.
>> That doesn't give you permission to design them in.
>> Plus, the LSMs that implement mandatory access controls
>> are going to want to intervene. No unclassified user
>> should see notifications caused by Top Secret users.
> But that's probably because they're worried about *side* channels, not
> covert channels?

The security evaluators from the 1990's considered any channel
with greater than 1 bit/second bandwidth a show-stopper. That was
true for covert and side channels. Further, if you knew that a
mechanism had a channel, as this one does, and you didn't fix it,
you didn't get your certificate. If you know about a problem
during the design/implementation phase it's really inexcusable not
to fix it before "completing" the code.

> Talking about this in the context of (small) side channels: The
> notification types introduced in this patch are mostly things that a
> user would be able to observe anyway if they polled /proc/self/mounts,
> right?

It's supposed to be a general mechanism. Of course it would
be simpler if is was restricted to things you can get at via
/proc/self.

>  It might make sense to align access controls based on that - if
> you don't want it to be possible to observe events happening on some
> mount points through this API, you should probably lock down
> /proc/*/mounts equivalently, by introducing an LSM hook for "is @cred
> allowed to see @mnt" or something like that - and if you want to
> compare two cred structures, you could record the cred structure that
> is responsible for the creation of the mount point, or something like
> that.

I'm not going to argue against that.

> For some of the other patches, I guess things get more tricky because
> the notification exposes new information that wasn't really available
> before.

We have to look not just at the information being available,
but the mechanism used. Being able to look at information about
a process in /proc doesn't mean I should be able to look at it
using ptrace(). Access control isn't done on data, it's done on
objects. That I can get information by looking in one object provides
no assurance that I can get it through a different object containing
the same information. This happens in /dev all over the place. A
file with hard links may be accessible by one path but not another.

>
>>>  You can transmit information through the scheduler,
>>> through hyperthread resource sharing, through CPU data caches, through
>>> disk contention, through page cache state, through RAM contention, and
>>> probably dozens of other ways that I can't think of right now.
>> Yeah, and there's been a lot of activity to reduce those,
>> which are hard to exploit, as opposed to this, which would
>> be trivial and obvious.
>>
>>> There
>>> have been plenty of papers that demonstrated things like an SSH
>>> connection between two virtual machines without network access running
>>> on the same physical host (<https://gruss.cc/files/hello.pdf>),
>>> communication between a VM and a browser running on the host system,
>>> and so on.
>> So you're saying we shouldn't have mode bits on files because
>> spectre/meltdown makes them pointless?
> spectre/meltdown are vulnerabilities that are being mitigated.
> Microarchitectural covert channels are an accepted fact and I haven't
> heard of anyone seriously considering trying to get rid of them all.