This blog now supports the HTTP/2 protocol, courtesy of nginx 1.9.5 (PDF).
In the process, I was stymied by an “ERR_SPDY_INADEQUATE_TRANSPORT_SECURITY” error from Google Chrome. HTTP/2 mandates TLS de facto, if not in the strict letter of the specification, and it also forbids a number of obsolete or weaker SSL/TLS ciphers to only permit ones that are truly secure. After some considerable digging, I found out the issue is Google Chrome on Mac and Android (presumably Windows as well) does not support 256-bit AES in HTTP/2, and my server was set up to only accept 256-bit encryption (only the best will do for my readers!). The error message was misleading: it’s not the server but Chrome’s crypto which is lacking.
It seems the cryptographers at Google feel 128-bit AES in Galois Counter Mode is good enough, and they did not want to be too far apart from Firefox (which does not support it either, and just fails without even the courtesy of an error message). In contrast, Safari on Yosemite supports AES-256-CBC (not ideal, I know, but that’s also what Chrome supports if HTTP/2 is turned off) and AES-256-GCM on El Capitan and iOS 9. Here are the settings your browser uses:
This is disappointing. AES-256-GCM is supported in hardware on most Intel hardware nowadays (all but lowest-end chips have the AES-NI instructions) and in the ARMv8-A architecture supported by most smartphones and mobile devices today, where the extra CPU load would matter most. I wonder how much of this is driven by Google’s fondness for Dan Bernstein’s ChaCha20+Poly1305 algorithms. Excellent as they may be, they are not implemented in hardware on the most common platforms, nor implemented at all in OpenSSL. It is quite disconcerting that my phone has better crypto than my desktop browser.
I ended up resolving the issue by loosening my cipher list from
EECDH+AESGCM:EDH+AESGCM:AES256+EECDH:AES256+EDH, but Chrome really should catch up and not let itself be hobbled by the increasingly irrelevant Firefox and its hoary NSS crypto.
I probably sound harsher than I intended towards the Google crypto team. The backward compatibility issues they have to deal with, from poorly designed TLS standards to broken web server software, intrusive anti-virus or corporate proxy servers mean a lot of their energy goes into exception cases, rather than implementing the latest and greatest in crypto algorithms.