libressl sha1 performance regression vs openssl.

[mharvey-jt]

Looks like the sha1 code in libressl uses SSSE3 instructions, but not the newer SHA1 instruction extensions, which are used by openssl. This is showing up as a performance regression with 2.11 vs 2.8.

[jblomer]

Thank you! That went unnoticed. The micro benchmarks (on my machine) confirm a factor 2 slowdown of libressl vs. openssl. This is true also for the latest version of libressl. Unfortunately, the OpenSSL 3 interface changes make it unusable for cvmfs. Not sure what a good option is. Possibilities are

• Patching libressl with the openssl code Edit: not trivial, most likely a bad idea
• Using a sha-1 standalone library (to be found) Edit: possible libraries with SHA support include botan and GnuTLS

Let's see with the libressl project first. I'll open an issue.

This seems to be the relevant commit. Merged for OpenSSL 1.0.2, unfortunately briefly after LibreSSL was forked.

[jblomer]

We should also not exclude moving back to OpenSSL. At least we should note down exactly where API functionality was really removed as compared to where we only need to use it differently.

[jblomer]

Both GnuTLS and Botan make it difficult to control memory allocation/deallocation of the hash context. But nettle looks interesting. It has the same SHA-1 performance than OpenSSL. Nettle is a crypto backend of GnuTLS.