Honestly if it weren't programming I feel like this is a movie-worthy story. To me it sounds so thrilling like a spy movie plot but am I just imagining that or was it actually this crazy / cool / integral to way bigger moving parts / things like I'm assuming? Regardless kudos. Your story definitely started my day on a happy note, thanks!
We had MasterCard end-to-end test auditor on site. This is the first time ever you get to do a transaction with real transaction system with real credit card.
Due to requirements we opted to have the only large meeting room to have outside our secure zone. This created an issue as we had no network access from there and in the end we decided to use slow GPRS terminal for the test.
The end-to-end test starts with offline transactions which by their very nature are quite fast (it is negotiated between terminal and card).
But then we went to online transaction and it finished instantly too.
The auditor, bewildered, proclaimed the test failed as he assumed it was incorrectly processed offline instead of going online. But then I pointed out to the printout to show ARQC (basically says it was certified online).
Now, the real discussion started. The terminal was very slow taking quite few seconds to establish GPRS and then even more for the SSL handshake so the auditor said it was not possible to make it work.
How it worked was that I have completely gutted OpenSSL and had entire cryptographic state stored locally (safely, using internal HSM) so the SSL session could be optimistically re-established without another handshake even after TCP connection was closed. The first message the terminal sends is already encrypted transaction message, there is no SSL handshake. I wrote an application to terminate the connection in our data center so that it stored the states of each connection in the database. The entire handshake was only done if the first message could not be decrypted successfully.
The operating system was single-threaded with no multitasking of any kind. This meant that all applications on this device did their operations sequentially. Send network message, print something, display something, etc.
I wrote a cooperative multitasking functionality into the application (using coroutines) so that it could work on multiple tasks at the same time (like talking to network and printing).
I then have segregated all data on the printouts so that it can start printing without having to already have response from network. Hopefully if everything went right, the response would come before it even came to that place on the printout effectively looking as if it was done in zero time.
FWIW, the described technique (or something roughly equivalent) is now standardized as 0-rtt early data in TLS 1.3. (you still need 1-rtt for TCP, unless you can combine this with tcp fast open, or run TLs over UDP)
But am I wrong to have my hackles raised by a) the roll-your-own security nature of this, b) the reliance on a single developer's single stack implementation as what guarantees the integrity of the system? It seems like there are a lot of assumptions baked in.
I, too, would love to see a more detailed write-up--if there's a big idea here (almost a unikernel thought), it deserves to be shared and tried by fire.
