Michael, if you need networking then just solve this problem and the you can use it also for exporting stuff. I'm available privately to try help you, just ping me. El mié., 20 oct. 2021 8:31, Indan Zupancic escribió:

Hello Michael,

On 2021-10-19 22:46, Michael Neises wrote:

...

I want to be able to retrieve data from seL4's virtual Linux machine, in order to store it in a persistent way. [...] If you have a different idea how I might achieve my goal, I would be similarly effusive in my thanks.

If it is just one time, one way blob of data passing, you could add it to Linux's command line, or put it into the device tree before starting the VM.

Or use shared memory between your seL4 application and the VM. You can tell Linux about it via device tree and then map it via /dev/mem or your own Linux (UIO) kernel driver. If using camkes, it may be easiest to create a virtual PCI UIO device (I have no experience with camkes).

Or use HW specific device registers/SRAM to stash the values in.

The best way forward depends on what other communication you need to do between seL4 and Linux.

Greetings,

Indan _______________________________________________ Devel mailing list -- devel@sel4.systems To unsubscribe send an email to devel-leave@sel4.systems

Hi Michael Indan Zupancic suggestion about seL4 and Linux communication via shared memory by reserving a memory block in the memory map passed to the linux kernel sounds like an good initial step ( not sure if seL4 has some native mechanism to do so), you can have some telemetry code running inside the kernel by using ebpf ( https://www.brendangregg.com/ebpf.html), ebpf execution mechanism is already in the kernel and probably can be used to directly have access to the memory region shared with your seL4. basically ebpf could be used to inspect any runtime inside the linux system without trusting any tool available within linux user space. this is just a suggestion, an seL4 expert probably can give a more reliable option. -Jorge A. Garcia On Sun, Oct 24, 2021 at 12:46 PM Michael Neises wrote:

Hello seL4 developers,

Thank you for the replies.

For the sake of clarity, the system works like this: At compile time, some expected digest values are made available only to a distinct CAmkES component. At the time of first-Linux-boot, a kernel module takes several measurements of the other kernel modules present (including itself). It reports these digests outwards to CAmkES, where they are compared against the expected values. It is the "pre-compile-time provisioning of these expected digests" in which I am interested. At this time, I can simulate the system and compute these digests, but the only way I have to extract them is to copy them by hand off the screen. To be totally explicit, I want to extract these values in order to re-compile them into a system that knows its expected digest values. I want to have an initial simulation where I extract these digests, so that in the subsequent compilation and simulations, the system is aware what values these digests are required to take.

Please correct me if I am wrong, but I think if the very first thing the Linux kernel does is measure itself, before it is even connected to a network, then there is simply no attack surface. Of course I'm very happy to be wrong, but I don't see who the attacker is in this situation. Certainly, there remains an open question of how to extend these measurements meaningfully into the space where there is a viable attack surface (after enabling a network adapter), but I consider that question to be beside the point for now (some future work). If there is some way for me to inspect the run-time data of the Linux system without relying somewhat on a tool inside the Linux instance, I would very much like to know about it. My strategy follows the same path as the vm-introspect example app (which I'm under the impression was created for this explicit purpose), which itself trusts implicitly the Linux instance. Again, to be entirely explicit, there does not appear to be any information anywhere on a way to meaningfully inspect a virtualized Linux system without trusting it even the slightest bit. I would be elated to be corrected; if someone can show me how to scrape (from the seL4 side exclusively) all the bits from a virtual Linux system and reassemble those bits into semantically-valuable information, I'm sure you will not hear from me for several months while I rebuild everything I have. At this time that does not appear to be possible.

Cheers, Michael Neises

On Tue, Oct 19, 2021 at 3:46 PM Michael Neises wrote:

...

Hello seL4 developers,

I want to be able to retrieve data from seL4's virtual Linux machine, in order to store it in a persistent way. Namely, I want to be able to simulate a seL4 kernel, boot its Linux virtual machine, compute some hash digests, and then export those hash digests. These digests are valuable because they represent the "clean room" runtime-state of the linux machine. Currently I can export these digests by way of hand-eye coordination, but I consider this unusable as a piece of software.

To date I've taken two main approaches: CAmkES FileServer or virtual networking. I'm under the impression that the FileServer changes are not persistent through reboot, and even if they were, to change the boot image after compile-time would seem to fly in the face of seL4's principles. Virtual networking seems to promise I can host my digests on a webpage that is visible to my "root host" machine; that is, the simulated seL4's linux instance hosts a site available on my 192.168.x.x network. I know there is a seL4webserver app as part of the seL4 repositories which claims to do this, but unfortunately its prose is unhelpful and it doesn't seem to work even when it compiles and simulates.

I've taken two distinct strategies to investigate the virtual network approach. First, I tried to get it to work on my normal stack: Windows 10 using WSL2 using a Docker container to simulate the seL4 image. The problem with this approach is that it appears I'm required to blindly thread 3 or 4 needles all at once, without getting feedback more descriptive than "you didn't do it." In other words, there does not appear to be a partial success available, and without ICMP ping, I honestly have no idea how to debug these "virtual" networks.

Next, I tried simplifying my stack by installing the dependencies natively on a Debian 10 machine, which should bypass several layers of the virtual network I was suggesting in my first strategy. Unfortunately, I met with the same "AttributeError: module 'yaml' has no attribute 'FullLoader'" error that inspired me to begin using Docker several years ago. Of course I should note that "pip/pip2/pip3 install pyyaml" all report that pyyaml is already installed, so I would be in debt to anyone who has an idea about that error.

To conclude, I find virtual networks opaque, and I would be grateful for any guidance. If you have a different idea how I might achieve my goal, I would be similarly effusive in my thanks.

Cheers, Michael Neises

_______________________________________________ Devel mailing list -- devel@sel4.systems To unsubscribe send an email to devel-leave@sel4.systems