diff -u: What's New in Kernel Development

David Drysdale wanted to add Capsicum security features to Linux after he noticed that FreeBSD already had Capsicum support. Capsicum defines fine-grained security privileges, not unlike filesystem capabilities. But as David discovered, Capsicum also has some controversy surrounding it.

Capsicum has been around for a while and was described in a USENIX paper in 2010: http://www.cl.cam.ac.uk/research/security/capsicum/papers/2010usenix-security-capsicum-website.pdf.

Part of the controversy is just because of the similarity with capabilities. As Eric Biderman pointed out during the discussion, it would be possible to implement features approaching Capsicum's as an extension of capabilities, but implementing Capsicum directly would involve creating a whole new (and extensive) abstraction layer in the kernel. Although David argued that capabilities couldn't actually be extended far enough to match Capsicum's fine-grained security controls.

Capsicum also was controversial within its own developer community. For example, as Eric described, it lacked a specification for how to revoke privileges. And, David pointed out that this was because the community couldn't agree on how that could best be done. David quoted an e-mail sent by Ben Laurie to the cl-capsicum-discuss mailing list in 2011, where Ben said, "It would require additional book-keeping to find and revoke outstanding capabilities, which requires knowing how to reach capabilities, and then whether they are derived from the capability being revoked. It also requires an authorization model for revocation. The former two points mean additional overhead in terms of data structure operations and synchronisation."

Given the ongoing controversy within the Capsicum developer community and the corresponding lack of specification of key features, and given the existence of capabilities that already perform a similar function in the kernel and the invasiveness of Capsicum patches, Eric was opposed to David implementing Capsicum in Linux.

But, given the fact that capabilities are much coarser-grained than Capsicum's security features, to the point that capabilities can't really be extended far enough to mimic Capsicum's features, and given that FreeBSD already has Capsicum implemented in its kernel, showing that it can be done and that people might want it, it seems there will remain a lot of folks interested in getting Capsicum into the Linux kernel.

Sometimes it's unclear whether there's a bug in the code or just a bug in the written specification. Henrique de Moraes Holschuh noticed that the Intel Software Developer Manual (vol. 3A, section 9.11.6) said quite clearly that microcode updates required 16-byte alignment for the P6 family of CPUs, the Pentium 4 and the Xeon. But, the code in the kernel's microcode driver didn't enforce that alignment.

In fact, Henrique's investigation uncovered the fact that some Intel chips, like the Xeon X5550 and the second-generation i5 chips, needed only 4-byte alignment in practice, and not 16. However, to conform to the documented specification, he suggested fixing the kernel code to match the spec.

Borislav Petkov objected to this. He said Henrique was looking for problems where there weren't any. He said that Henrique simply had discovered a bug in Intel's documentation, because the alignment issue clearly wasn't a problem in the real world. He suggested alerting the Intel folks to the documentation problem and moving on. As he put it, "If the processor accepts the non-16-byte-aligned update, why do you care?"

But, as H. Peter Anvin remarked, the written spec was Intel's guarantee that certain behaviors would work. If the kernel ignored the spec, it could lead to subtle bugs later on. And, Bill Davidsen said that if the kernel ignored the alignment requirement, and "if the requirement is enforced in some future revision, and updates then fail in some insane way, the vendor is justified in claiming 'I told you so'."

The end result was that Henrique sent in some patches to make the microcode driver enforce the 16-byte alignment requirement.

______________________