huldufolk is an implementation of the userspace side of the Linux kernel's
CONFIG_STATIC_USERMODEHLEPER option. Using it should make your Linux more
secure!
Usermode helper binaries are always run as "real" root, and are thus an attractive place to launch code that runs with escalated privileges.
In particular, it can be an easier way to run arbitrary code as root once some other in-kernel exploit is successful, e.g.:
- It may be easier to call
run_cmdorcall_usermodehelperwith a custom binary vs. manipulating memory protections to achieve code execution, as in https://googleprojectzero.blogspot.com/2018/09/. - It's an easy way to hand flow control of a successful exploit back to userspace, as in https://www.openwall.com/lists/oss-security/2017/02/04/1 or https://www.openwall.com/lists/oss-security/2016/12/07/3
Thus, the chief threat is that the kernel may accidentally execute something it didn't intend to as root, via the usermodehelper script. There are two ways this can happen:
- Memory corruption, e.g. the exploits above, or from the original patch series by changing some of the hard coded strings in the kernel. The original series also a bunch of these strings const, but there are various mentions of tiny races which can be won. In any case, dispatching to a trusted userspace binary eliminates these.
- Someone overwrites
/proc/sys/kernel/modprobeor similar control file to cause the kernel to execute the binary they want. Usermode helper control files require root to be written to, but some clever userspace attacker may be able to figure out a way to write to this file.
In our initial design, we're mostly concerned with offering an implementation of a static usermode-helper that is general purpose enough to be used in the various distros. Thus, it needs to be configurable by administrators, and since the usermode helper infrastructure doesn't allow us to export any type of configuration via the environment, we need to use a config file.
Of course, this means that a clever attacker who can write to
/proc/sys/kernel/modprobe can instead just write to
/etc/usermode-helper.conf instead. By using a config file, we're not really
protecting against attack type two above. One could imagine an entirely static
(i.e. config file free) version of this binary down the road to solve this
problem. But also, someone who can write to arbitrary files as root already has
lots of power anyway, so it's not a high priority :)
Finally, the dispatcher can also protect against bugs in the binaries
themselves by applying the principle of least privilege. In particular,
/sbin/modprobe presumably doesn't need CAP_SYS_MKNOD or CAP_SYS_ADMIN, so
a dispatcher could drop these privileges before running the right binary,
protecting against bugs in the userspace code as well.
Linux 4.11 introduced a pair of compile time options,
CONFIG_STATIC_USERMODEHELPER and CONFIG_STATIC_USERMODEHELPER_PATH, which
when set, force all usermode helper executions to go through the path specified
(/sbin/usermode-helper by default), but passes the name of the original
executable as argv[0], so that userspace can figure out which thing was
originally intended to be executed. This configuration setting is recommended
by various Linux hardening guides (including Alexander Popov's excellent
kconfig-hardened-check
script), but vanishingly little information is available about the userspace
counterpart to this kernel setting.
While at Docker I wrote an initial implementation of this for LinuxKit, but it was not really suitable for general purpose Linux distros. If we wanted to make it useful, we'd need a configuration file. But since it's written in C, then we'd have to parse stuff with that configuration file, and that would probably have bugs.
So yes, this is written in rust. I understand that makes me a hipster, but I already have a beard so I'm dangerously close anyway. However, rust seems like a good choice for a low-level system utility such-as the static usermode-helper. But, it also means that packaging it may be a pain for distros. Today we don't have a lot of experience packaging rust binaries, and this may be a pain point. For that reason, I've tried to keep the number of dependencies small. However, it does mean that we'll hopefully have less bugs, since this thing will necessarily run as root.
The basic design is that the configuration file says what's allowed to be run, with potential additional restrictions like argument filters or dropped capabilities, so if there's a bug in the helper itself, it can't do too much damage.
There is also a default configuration file with comments indicating what paths are used where, so that you can strip out things you don't want. A general purpose distro should hopefully be able to use this config file with little fuss.
- argument filters (probably based on regexes?)
- setting NNP?
- Namespaces?
- seccomp filters? Is there some nice language for specifying these in config files? Perhaps we want to do something else?
Huldufólk are hidden elves from Icelandic folklore. One can see evidence of their "existence" when traveling Iceland today.
[Huldufólk] prevented many children from wandering away from human habitations, taught Iceland's topographical history, and instilled fear and respect for the harsh powers of nature.
-- Ólina Thorvarðardóttir
It seems only natural to want to also prevent child processes from wondering away :)
Plus, static-usermodehelper is boring.