toggle quoted messageShow quoted text
Good day Nate,
As you seem to be mostly in charge of the GSoC side of things, I direct this at you, but of course everyone is welcome to comment.
I think I finished my first round of investigations just in time for the deadline. You can find a list of BZs attached. Please note that 1) some are declared security issues and may not be publicly accessible, and 2) that this list is far from complete. I only added items that are a) not implicitly fixed by "simply" deploying the new Image Loader (*some* important prerequisites are listed), and b) I am confident are actual issues (or things to consider) I believe to know how to approach.
I have taken notes about more things, e.g. the existence of the security architectural protocols, which I could not find a rationale for. I can prepare something for this matter, but it really needs an active discussion with some of the core people. I'm not sure delayed e-mail discussion is going to be enough, but there is an official IRC I suppose. :) I hope we can work something out for this.
I also hope this makes it clearer why I don't believe that we need to "fill" 10 weeks, but rather the opposite. This is not a matter of replacing a library instance, but the whole surrounding ecosystem needs to follow for the changes to make sense. And as I tried to make clear in my discussion with Michael Brown, I am not keen on preserving backwards-compatibility with platform code (i.e. PEI, DXE, things we consider "internal"), as most of it should be controlled by EDK II already. This of course does *not* include user code (OROMs, bootloaders, ...), for which I want to provide the *option* to lock some of them out for security, but with sane defaults that will ensure good compatibility.
I'd like to thank Michael Brown for his cooperation and support, because we recently landed changes in iPXE to allow for the strictest image format and permission constraints currently possible.
I will have to rework the submitted proposal to reflect the new knowledge. To be honest, seeing how the BZs kept rolling in, I am not convinced an amazing amount of mainlining can be accomplished during the 10 weeks. It may have to suffice to have a publicly accessible prototype (e.g. OVMF) and a subset of the planned patches on the list. I hope you can manage to provide some feedback before the deadline passes tomorrow.
Thank you in advance!
On 06.04.21 11:41, Nate DeSimone wrote:
Great to meet you and welcome back! Glad you hear you are interested! Completing a formal verification of a PE/COFF loader is certainly impressive. Was this done with some sort of automated theorem proving? It would seem a rather arduous task doing an inductive proof for an algorithm like that by hand! I completely agree with you that getting a formally verified PE/COFF loader into mainline is undoubtably valuable and would pay security dividends for years to come.
Admittedly, this is an area of computer science that I don't have a great deal of experience with. The furthest I have gone on this topic is writing out proofs for simple algorithms on exams in my Algorithms class in college. Regardless you have a much better idea of what the current status is of the work that you and Vitaly have done. I guess my only question is do you think there is sufficient work remaining to fill the 10 week GSoC development window? Certainly we can use some of that time to perform the code reviews you mention and write up formal ECRs for the UEFI spec changes that you believe are needed.
Thank you for sending the application and alerting us to the great work you and Vitaly have done! I'll read your paper more closely and come back with any questions I still have.
With Best Regards,
From: email@example.com <firstname.lastname@example.org> On Behalf Of Marvin
Sent: Sunday, April 4, 2021 4:02 PM
To: email@example.com; Laszlo Ersek <firstname.lastname@example.org>; Andrew Fish
<email@example.com>; Kinney, Michael D <firstname.lastname@example.org>
Subject: [edk2-devel] [GSoC proposal] Secure Image Loader
Good day everyone,
I'll keep the introduction brief because I've been around for a while now. :) I'm
Marvin Häuser, a third-year Computer Science student from TU Kaiserslautern,
Germany. Late last year, my colleague Vitaly from ISP RAS and me introduced a
formally verified Image Loader for UEFI usage at ISP RAS Open due to various
defects we outlined in the corresponding paper. Thank you once again Laszlo
for your *incredible* review work on the publication part.
I now want to make an effort to mainline it, preferably as part of the current
Google Summer of Code event. To be clear, my internship at ISP RAS has
concluded, and while Vitaly will be available for design discussion, he has other
priorities at the moment and the practical part will be on me. I have previously
submitted a proposal via the GSoC website for your review.
There are many things to consider:
1. The Image Loader is a core component, and there needs to be a significant
level of quality and security assurance.
2. Being consumed by many packages, the proposed patch set will take a lot of
time to review and integrate.
3. During my initial exploration, I discovered defective PPIs and protocols (e.g.
returning data with no corresponding size) originating from the UEFI PI and
UEFI specifications. Changes need to be discussed, settled on, and submitted to
the UEFI Forum.
4. Some UEFI APIs like the Security Architecture protocols are inconveniently
abstract, see 5.
5. Some of the current code does not use the existing context, or accesses it
outside of the exposed APIs. The control flow of the dispatchers may need to be
adapted to make the context available to appropriate APIs.
But obviously there are not only unpleasant considerations:
A. The Image Loader is mostly formally verified, and only very few changes will
be required from the last proven state. This gives a lot of trust in its correctness
B. All outlined defects that are of critical nature have been fixed successfully.
C. The Image Loader has been tested with real-world code loading real-world
OSes on thousands of machines in the past few months, including rejecting
malformed images (configurable by PCD).
D. The new APIs will centralise everything PE, reducing code duplication and
potentially unsafe operations.
E. Centralising and reduced parse duplication may improve overall boot
F. The code has been coverage-tested to not contain dead code.
G. The code has been fuzz-tested including sanitizers to not invoke undefined
H. I already managed to identify a malformed image in OVMF with its help
(incorrectly reported section alignment of an Intel IPXE driver). A fix will be
I. I plan to support PE section permissions, allowing for read-only data
segments when enabled.
There are likely more points for both lists, but I hope this gives a decent
starting point for discussion. What are your thoughts on the matter? I strongly
encourage everyone to read the section regarding defects of our publication
to better understand the motivation. The vague points above can of course be
elaborated in due time, as you see fit.
Thank you for your time!