For the record, the valgrind trace for the crash is:

Here's another one related to this: https://lists.gnupg.org/pipermail/gcrypt-devel/2022-July/005344.html

Key length requirements for KDFs are specified in SP 800-131Ar2 (https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-131Ar2.pdf), which is linked from SP 800-140Dr1 (https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-140Dr1.pdf) in section "6.2.1 Transitions".

FIPS 140-3 (https://csrc.nist.gov/Projects/cryptographic-module-validation-program/fips-140-3-standards) points to SP 800-140Dr1 (https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-140Dr1.pdf) to list acceptable "Security Parameter Generation and Establishment Methods". From this document, RFC 5869 (i.e., HKDF with the counter at the end) can be reached via two paths:

Feedback from the lab is that they'd recommend returning a specific error code that indicates that the prime search failed and then relying on the caller to decide whether to loop or bubble up the error. I'm not sure who we would consider to be the "caller" of the relevant generation function in this case, though.

That sounds reasonable. The FIPS 186-5 draft (https://nvlpubs.nist.gov/nistpubs/FIPS/NIST.FIPS.186-5-draft.pdf) covers this in section A.1.3, although I'm not quite sure why a lower bound for p was chosen compared to q. The comment that seems to have triggered this change is published on page 68 of https://csrc.nist.gov/CSRC/media/Publications/fips/186/4/final/documents/comments-received-fips186-4-december-2015.pdf by Allen Roginsky. It only contains a suggestion of 20, presumably for both numbers.

Ah, right, I can get that added to the containers tomorrow.

That only seems to work in some configurations: https://gitlab.com/redhat-crypto/libgcrypt/libgcrypt-mirror/-/pipelines/472626834

Thanks! Maybe it would be simpler to use dl_iterate_phdr(3) for this. I wasn't aware of the function, but a colleague just implemented a proof-of-concept of what you're proposing in https://gitlab.com/dueno/integrity-notes.

This still seems to be a problem: