On Wed, 2023-03-22 at 15:27 -0700, Alexei Starovoitov wrote:
On Wed, Mar 22, 2023 at 5:08 AM Roberto Sassu roberto.sassu@huaweicloud.com wrote:
On Tue, 2023-03-21 at 19:23 -0700, Alexei Starovoitov wrote:
On Fri, Mar 17, 2023 at 7:53 AM Roberto Sassu roberto.sassu@huaweicloud.com wrote:
From: Roberto Sassu roberto.sassu@huawei.com
A User Mode Driver (UMD) is a specialization of a User Mode Helper (UMH), which runs a user space process from a binary blob, and creates a bidirectional pipe, so that the kernel can make a request to that process, and the latter provides its response. It is currently used by bpfilter, although it does not seem to do any useful work.
FYI the new home for bpfilter is here: https://github.com/facebook/bpfilter
Thanks. I just ensured that it worked, by doing:
getsockopt(fd, SOL_IP, IPT_SO_GET_INFO, &info, &optlen);
and accepting IPT_SO_GET_INFO in main.c.
The problem is, if other users would like to implement a UMD similar to bpfilter, they would have to duplicate the code. Instead, make an UMD management library and API from the existing bpfilter and sockopt code, and move it to common kernel code.
Also, define the software architecture and the main components of the library: the UMD Manager, running in the kernel, acting as the frontend interface to any user or kernel-originated request; the UMD Loader, also running in the kernel, responsible to load the UMD Handler; the UMD Handler, running in user space, responsible to handle requests from the UMD Manager and to send to it the response.
That doesn't look like a generic interface for UMD.
What would make it more generic? I made the API message format- independent. It has the capability of starting the user space process as required, when there is a communication.
It was a quick hack to get bpfilter off the ground, but certainly not a generic one.
True, it is not generic in the sense that it can accomodate any possible use case. The main goal is to move something that was running in the kernel to user space, with the same isolation guarantees as if the code was executed in the kernel.
They are not the same guarantees. UMD is exactly equivalent to root process running in user space. Meaning it can be killed, ptraced, priority inverted, etc
That is the starting point.
I suppose you can remove any privilege from the UMD process, it just needs to read/write from/to a pipe (and in my case to use socket() with AF_ALG to interact with the Crypto API).
Also, as I mentioned, you can enforce a very strict seccomp profile, which forces the UMD process to use a very limited number of system calls.
For the interactions of the rest of the system to the UMD process, you could deny with an LSM all the operations that you mentioned. The rest of the system would not be affected, only operations which have the UMD process as target are denied.
I have two use cases, but for sake of brevity I will propose one.
I would like to add support for PGP keys and signatures in the kernel, so that I can extend secure boot to applications, and allow/deny code execution based on the signed file digests included in RPM headers.
While I proposed a patch set a while ago (based on a previous work of David Howells), the main objection was that the PGP packet parser should not run in the kernel.
That makes a perfect example for using a UMD. If the PGP parser is moved to user space (UMD Handler), and the kernel (UMD Manager) just instantiates the key and verifies the signature on already parsed data, this would address the concern.
I don't think PGP parser belongs to UMD either. Please do it as a normal user space process and define a proper protocol for communication between kernel and user space.
UMD is better in the sense that it establishes a bidirectional pipe between the kernel and the user space process. With that, there is no need to further restrict the access to a sysfs file, for example.
If a simple pipe is good enough then you can have a kernel module that creates it and interacts with the user space process.
Few points I forgot to mention.
With the UMD approach, the binary blob is embedded in the kernel module, which means that no external dependencies are needed for integrity verification. The binary is statically compiled, and the kernel write-protects it at run-time.
Second, since DIGLIM would check the integrity of any executable, including init, the PGP signature verification needs to occur before. So, the PGP UMD should be already started by then. That is not going to be a problem, since the binary is copied to a private tmpfs mount.
Out-of-tree bpftiler can do that, so can you.
As far as I can see, the out-of-tree bpfilter works exactly in the same way as the in-tree counterpart. The binary blob is embedded in the kernel module.
PGP is not suitable for kernel git repo either as kernel code or as UMD.
Well, the asymmetric key type can be extended with new parsers, so this possibility was already taken into account. The objection that the PGP parser should not run in kernel space is fair, but I think the UMD approach fully addresses that.
Also, I agree with you that we should not just take any code and pretend that it is part of the kernel. However, in this particular case, the purpose of the PGP UMD would be simply to extract very few information from the PGP packets. The asymmetric key type and the signature verification infrastructure already take care of the rest.
PGP keys and signatures would act as an additional system trust anchor for verifying critical system data (for DIGLIM, which executables are allowed to run), similarly to how X.509 certificates are used for verifying kernel modules. RPM headers, executables digests are taken from, are signed with PGP, so there is no other way than adding this functionality.
And unfortunately, especially for features impacting the entire system, out-of-tree drivers are not really an option:
https://docs.fedoraproject.org/en-US/quick-docs/kernel/overview/#_out_of_tre...
Thanks
Roberto