Add the Miou implementation of TLS

[dinosaure]

This PR provides an implementation of tls with miou. It provides also a test which checks randomly different actions between a client and a server. The package is actually tested with httpcats (for http/1.1 and h2) and, if we want to take the opportunity to do some HTTP requests in parallel, this patch about mirage-crypto is needed - but it is not necessary if we only use Miou.call_cc which spawns a task concurrently.

The fuzzer

The fuzzer tests the behavior of Tls_miou when we use these actions:

type operation =
  | Send of string
  | Recv of int
  | Shutdown of [ `read | `write ]
  | Close
  | Noop

It does not test the TLS implementation (if we are able to correctly encrypt/decrypt the session). These actions are randomly chosen by the fuzzer. The idea is to produce a suit of actions for a client and a server. Note that we don't generate a situation where both read or both write "at the same time". The generation has 4 points where we are able to close, shutdown(recv) or shutdown(write). An example is:

client: [send(64); noop; recv(64); shutdown(send); recv(64); noop; noop; close; noop; close]
server: [recv(64); noop; send(64); shutdown(send); send(64); close; noop; close; noop; close]

Then, we launch a TLS server and start a new session with a client. Both do actions and fill a buffer. We catch all possible errors and terminate properly the server and the client. Finally, we compare what the server received and what the client received with an oracle. The oracle is the compile function. The rules are simple:

• we record when the client and the server shutdown/close the connection
• if the client/server want to send/recv something, we check the status of the connection. We continue if everything is normal (none of them shutdown/close their sides) or we directly return the final result (in our implementation, we should raise an exception End_of_file or Closed_by_peer).

The error cases

Actually, from an UNIX socket, read can raises ECONNRESET and write can raise EPIPE and ECONNRESET. For our perspective, these cases tell us that the connection is closed (on a specific side or totally). We handle these cases and transfer the information to the internal TLS state.

Another issues encountered is about read and close. Our fuzzer uses a UNIX socket - we can not rely on the throughput of the network. It can arrives that the server get the whole data and a close-notify before to do anything. In this situation, we must returns what we got and consider, afterwards, the connection closed even if the TLS state is actually `Closed.

Finally, the previous point enforces the fact that we must returns what the client/server received even if the connection was closed (or shutdown(recv)). However, the user can also decide to close or shutdown(recv) the connection before to read anything. We must drop pending application-data in such situation.

Parallelism

Currently, as said in the introduction, we must use a patch for mirage-crypto to be able to launch several parallel TLS sessions. The mirage-crypto-rng-miou is also needed if we don't want any data-race conditions when we generate a random value. The test actually uses only one domain and mirage-crypto-rng.unix according to the current situation of mirage-crypto.

[dinosaure]

A new branch dinosaure/ocaml-tls#miou-string exists which is this PR on top of our changes about Cstruct.t -> string but I will propose it when the Cstruct.t -> string PR will be merge. The code does not change to much. Let me know when everything will be good about tls to propose then a rebased version of this PR.

[hannesm]

I thought we'll do the cstruct -> string jump in mirage-crypto with a release soon, and then adapt x509 etc. as well, and get this PR merged when we also have tls using string/bytes. Or should this be merged earlier and released (I thought it'd depend on mirage-crypto-rng-miou).

[dinosaure]

We should merge this after the Cstruct.t - > string PR. It does not require as is mirage-crypto-rng-miou-unix but it requires fixes about parallelism to work nicely then (so it requires a release of mirage-crypto which includes the Cstruct.t -> string move).

[hannesm]

This can now that #497 is merged be rebased on the current main branch.

[hannesm]

superseeded by #503