[systemd-devel] Creating executable device nodes in /dev?

[Topi Miettinen]

On 8.12.2020 23.30, Andy Lutomirski wrote:
> 
>> On Dec 8, 2020, at 12:45 PM, Topi Miettinen <toiwoton at gmail.com> wrote:
>>
>> On 8.12.2020 20.07, Andy Lutomirski wrote:
>>>> On Thu, Nov 19, 2020 at 10:05 AM Topi Miettinen <toiwoton at gmail.com> wrote:
>>>>
>>>> On 19.11.2020 18.32, Zbigniew Jędrzejewski-Szmek wrote:
>>>>> On Thu, Nov 19, 2020 at 08:17:08AM -0800, Andy Lutomirski wrote:
>>>>>> Hi udev people-
>>>>>>
>>>>>> The upcoming Linux SGX driver has a device node /dev/sgx.  User code
>>>>>> opens it, does various setup things, mmaps it, and needs to be able to
>>>>>> create PROT_EXEC mappings.  This gets quite awkward if /dev is mounted
>>>>>> noexec.
>>>>>>
>>>>>> Can udev arrange to make a device node executable on distros that make
>>>>>> /dev noexec?  This could be done by bind-mounting from an exec tmpfs.
>>>>>> Alternatively, the kernel could probably learn to ignore noexec on
>>>>>> /dev/sgx, but that seems a little bit evil.
>>>>>
>>>>> I'd be inclined to simply drop noexec from /dev by default.
>>>>> We don't do noexec on either /tmp or /dev/shm (because that causes immediate
>>>>> problems with stuff like Java and cffi). And if you have those two at your
>>>>> disposal anyway, having noexec on /dev doesn't seem important.
>>>>
>>>> I'd propose to not enable exec globally, but if a service needs SGX, it
>>>> could use something like MountOptions=/dev:exec only in those cases
>>>> where it's needed. That way it's possible to disallow writable and
>>>> executable file systems for most services (which typically don't need
>>>> /tmp or /dev/shm either). Of course the opposite
>>>> (MountOptions=/dev:noexec) would be also possible, but I'd expect that
>>>> this would be needed to be used more often.
>>>>
>>> I imagine the opposite would be more sensible.  It seems odd to me
>>> that we would want any SGX-using service to require both special mount
>>> options and regular ACL permissions.
>>
>> How common are thes SGX-using services? Will every service start using it without any special measures taken on it's behalf, or perhaps only a special SGX control tool needs access? What about unprivileged user applications, do they ever want to access SGX? Could something like Widevine deep in a browser need to talk to SGX in a DRM scheme?
> 
> I honestly don’t know. Widevine is probably some unholy mess of SGX and ME crud. But regular user programs may well end up using SGX for little non-evil enclaves, e.g. storing their keys securely.  It would be nice if unprivileged enclaves just work as long as the use has appropriate permissions on the device nodes.

Maybe, it would be also great if the access could be limited to those 
users or services which actually need it, by principle of least privilege.

> SGX adoption has been severely hampered by the massive series of recent vulnerabilities and by Intel’s silly licensing scheme. The latter won’t be supported upstream.
> 
>>
>>> As  a further argument, I just did this on a Fedora system:
>>> $ find /dev -perm /ugo+x -a \! -type d -a \! -type l
>>> No results.  So making /dev noexec doesn't seem to have any benefit.
>>
>> It's no surprise that there aren't any executables in /dev since removing MAKEDEV ages ago. That's not the issue, which is that /dev is a writable directory (for UID=0 but no capabilities are needed) and thus a potential location for constructing unapproved executables if it is also mounted exec (W^X).
> 
> UID 0 can just change mount options, though, unless SELinux or similar is used. And SELinux can protect /dev just fine without noexec.

Well, mounting would need CAP_SYS_ADMIN in addition to UID 0. Also 
SELinux is not universal and the policies might not contain all users or 
services.

-Topi