• Stars
    star
    307
  • Rank 136,109 (Top 3 %)
  • Language
    OCaml
  • License
    BSD 3-Clause "New...
  • Created almost 6 years ago
  • Updated 5 months ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

An HTTP/2 implementation written in pure OCaml

h2

h2 is an implementation of the HTTP/2 specification entirely in OCaml. It is based on the concepts in http/af, and therefore uses the Angstrom and Faraday libraries to implement the parsing and serialization layers of the HTTP/2 standard. It also preserves the same API as http/af wherever possible.

Installation

Install the library and its dependencies via OPAM:

opam install h2

Usage

Resources

First of all, the generated documentation lives here. It is recommended to browse it and get to know the API exposed by H2.

There are also some examples in the examples folder. Most notably, the ALPN example provides an implementation of a common real-world use case:

It sets up a server that listens on 2 ports:

  1. port 8080: redirects all incoming traffic to https://localhost:9443

  2. port 9443: negotiates which protocol to use over the TLS Application-Layer Protocol Negotiation (ALPN) extension. It supports 2 protocols (in order of preference): h2 and http/1.1.

    If h2 is negotiated, the example sets up a connection handler using h2-lwt-unix. Otherwise the connection handler will serve HTTP/1.1 traffic using http/af.

The ALPN example also provides a unikernel implementation with the same functionality that runs on MirageOS.

A server example

We present an annotated example below that responds to any GET request and returns a response body containing the target of the request.

open H2

(* This is our request handler. H2 will invoke this function whenever the
 * client send a request to our server. *)
let request_handler _client_address reqd =
  (* `reqd` is a "request descriptor". Conceptually, it's just a reference to
   * the request that the client sends, which allows us to do two things:
   *
   * 1. Get more information about the request that we're handling. In our
   *    case, we're inspecting the method and the target of the request, but we
   *    could also look at the request headers.
   *
   * 2. A request descriptor is also what allows us to respond to this
   *    particular request by passing it into one of the response functions
   *    that we will look at below. *)
  let { Request.meth; target; _ } = Reqd.request reqd in
  match meth with
  | `GET ->
    let response_body =
      Printf.sprintf "You made a request to the following resource: %s\n" target
    in
    (* Specify the length of the response body. Two notes to make here:
     *
     * 1. Specifying the content length of a response is optional since HTTP/2
     *    is a binary protocol based on frames which carry information about
     *    whether a frame is the last for a given stream.
     *
     * 2. In HTTP/2, all header names are required to be lowercase. We use
     *    `content-length` instead of what might be commonly seen in HTTP/1.X
     *    (`Content-Length`). *)
    let headers =
      Headers.of_list
        [ "content-length", string_of_int (String.length response_body) ]
    in
    (* Respond immediately with the response body we constructed above,
     * finishing the request/response exchange (and the unerlying HTTP/2
     * stream).
     *
     * The other functions in the `Reqd` module that allow sending a response
     * to the client are `Reqd.respond_with_bigstring`, that only differs from
     * `Reqd.respond_with_string` in that the response body should be a
     * bigarray, and `Reqd.respond_with_streaming` (see
     * http://anmonteiro.com/ocaml-h2/h2/H2/Reqd/index.html#val-respond_with_streaming)
     * which starts streaming a response body which can be written to
     * asynchronously to the client. *)
    Reqd.respond_with_string reqd (Response.create ~headers `OK) response_body
  | meth ->
    let response_body =
      Printf.sprintf
        "This server does not respond to %s methods.\n"
        (Method.to_string meth)
    in
    Reqd.respond_with_string
      reqd
      (* We don't include any headers in this case. The HTTP/2 framing layer
       * knows that these will be last frames in the exchange. *)
      (Response.create `Method_not_allowed)
      response_body

(* This is our error handler. Everytime H2 sees a malformed request or an
 * exception on a specific stream, it will invoke this function to send a
 * response back to the misbehaving client. Because there might not be a
 * request for the stream (handing malformed requests to the application is
 * strongly discouraged), there is also no request descriptor like we saw in
 * the request handler above. In this case, one of the arguments to this
 * function is a function that will start the response. It has the following
 * signature:
 *
 *   val start_response : H2.headers.t -> [`write] H2.Body.t
 *
 * This is also where we first encounter the concept of a `Body` (which were
 * briefly mentioned above) that can be written to (potentially
 * asynchronously). *)
let error_handler _client_address ?request:_ _error start_response =
  (* We start the error response by calling the `start_response` function. We
   * get back a response body. *)
  let response_body = start_response Headers.empty in
  (* Once we get the response body, we can immediately start writing to it. In
   * this case, it might be sufficient to say that there was an error. *)
  Body.Writer.write_string
    response_body
    "There was an error handling your request.\n";
  (* Finally, we close the streaming response body to signal to the underlying
   * HTTP/2 framing layer that we have finished sending the response. *)
  Body.Writer.close response_body

let () =
  (* We're going to be using the `H2_lwt_unix` module from the `h2-lwt-unix`
   * library to create a server that communicates over the underlying operating
   * system socket abstraction. The first step is to create a connection
   * handler that will accept incoming connections and let our request and
   * error handlers handle the request / response exchanges in those
   * connections. *)
  let connection_handler =
    H2_lwt_unix.Server.create_connection_handler
      ?config:None
      ~request_handler
      ~error_handler
  in
  (* We'll be listening for requests on the loopback interface (localhost), on
   * port 8080. *)
  let listen_address = Unix.(ADDR_INET (inet_addr_loopback, 8080)) in
  (* The final step is to start a server that will set up all the low-level
   * networking communication for us, and let it run forever. *)
  let _server =
    Lwt_io.establish_server_with_client_socket listen_address connection_handler
  in
  let forever, _ = Lwt.wait () in
  Lwt_main.run forever

A client example

The following annotated client example performs a GET request to example.com and prints the response body as it arrives.

open H2
module Client = H2_lwt_unix.Client

(* This is our response handler. H2 will invoke this function whenever the
 * server has responded to our request. The `notify_response_received` argument
 * is explained further down. *)
let response_handler notify_response_received response response_body =
  (* `response` contains information about the response that we received. We're
   * looking at the status to know whether our request produced a successful
   * response, but we could also get the response headers, for example. *)
  match response.Response.status with
  | `OK ->
    (* If we got a successful response, we're going to read the response body
     * as it arrives, and print its fragments as we receive them. *)
    let rec read_response () =
      (* Scheduling a read of the response body registers two functions with
       * H2:
       *
       * 1. `on_read`: this function will be called upon the receipt of a
       *    response body chunk (in HTTP/2 speak, a DATA frame). Our handling
       *    of these chunks is explained inline below.
       *
       * 2. `on_eof`: this function will be called once the entire response
       *    body has arrived. In our case, this is where we fulfill the promise
       *    that we're done handling the response.
       *)
      Body.Reader.schedule_read
        response_body
        ~on_read:(fun bigstring ~off ~len ->
          (* Once a response body chunk is handed to us (as a bigarray, and an
           * offset and length into that bigarray), we'll copy it into a string
           * and print it to stdout. *)
          let response_fragment = Bytes.create len in
          Bigstringaf.blit_to_bytes
            bigstring
            ~src_off:off
            response_fragment
            ~dst_off:0
            ~len;
          print_string (Bytes.to_string response_fragment);
          (* We need to make sure that we register another read of the response
           * body after we're done handling a fragment, as it will not be
           * registered for us. This is where our recursive function comes in
           * handy. *)
          read_response ())
        ~on_eof:(fun () ->
          (* Signal to the caller of the HTTP/2 request that we are now done
           * handling the response, and the program can continue. *)
          Lwt.wakeup_later notify_response_received ())
    in
    read_response ()
  | _ ->
    (* We didn't get a successful status in the response. Just print what we
     * received to stderr and bail early. *)
    Format.eprintf "Unsuccessful response: %a\n%!" Response.pp_hum response;
    exit 1

let error_handler _error =
  (* There was an error handling the request. In this simple example, we don't
   * try too hard to understand it. Just print to stderr and exit with an
   * unsuccessful status code. *)
  Format.eprintf "Unsuccessful request!\n%!";
  exit 1

open Lwt.Infix

let () =
  let host = "www.example.com" in
  Lwt_main.run
    ( (* We start by resolving the address of the host we want to connect to. *)
      Lwt_unix.getaddrinfo host "443" [ Unix.(AI_FAMILY PF_INET) ]
    >>= fun addresses ->
      (* Once the address for the host we want to contact has been resolved, we
       * need to create the socket through which the communication with the
       * remote host is going to happen. *)
      let socket = Lwt_unix.socket Unix.PF_INET Unix.SOCK_STREAM 0 in
      (* Then, we connect to the socket we just created, on the address we have
       * previously obtained through name resolution. *)
      Lwt_unix.connect socket (List.hd addresses).Unix.ai_addr >>= fun () ->
      let request =
        Request.create
          `GET
          "/"
          (* a scheme pseudo-header is required in HTTP/2 requests, otherwise
           * the request will be considered malformed. In our case, we're
           * making a request over HTTPS, so we specify "https" *)
          ~scheme:"https"
          ~headers:
            (* The `:authority` pseudo-header is a blurry line in the HTTP/2
             * specificiation. It's not strictly required but most
             * implementations treat a request with a missing `:authority`
             * pseudo-header as malformed. That is the case for example.com, so
             * we include it. *)
            Headers.(add_list empty [ ":authority", host ])
      in
      (* The H2 API relies on callbacks to allow for a single, stable core to
       * be used with different I/O runtimes. Because we're using Lwt in this
       * example, we'll create an Lwt task that is going to help us transform
       * the callback-calling style of H2 into an Lwt promise whenever we're
       * done handling the response.
       *
       * If you're not familiar with Lwt or its `Lwt.wait` function, it's
       * recommended you read at least the following bit before moving on:
       * http://ocsigen.org/lwt/4.1.0/api/Lwt#VALwait. *)
      let response_received, notify_response_received = Lwt.wait () in
      (* Partially apply the `response_handler` function that we defined above
       * to produce one that matches H2's expected signature. After this line,
       * `response_handler` now has the following signature:
       *
       *   val response_handler: Response.t -> [ `read ] Body.t -> unit
       *)
      let response_handler = response_handler notify_response_received in
      (* HTTP/2 itself does not define that the protocol must be used with TLS.
       * In practice, though, TLS is widely used in the Internet today (and
       * that's a good thing!) and no serious deployments use plaintext HTTP/2.
       * The following is a good read on why this is the case:
       * https://http2-explained.haxx.se/content/en/part8.html#844-its-use-of-tls-makes-it-slower
       *
       * For us, this means that we need to make our request over TLS. H2, and
       * more specifically `h2-lwt-unix`, provide a `TLS` module for both the
       * client and the server implementations that rely on an optional
       * dependency to ocaml-tls.
       *
       * We start by creating a connection handler. The `create_connection`
       * function takes two arguments: a connection-level error handler (you
       * can read more about the difference between connection-level and
       * stream-level in H2 and HTTP/2 in general here:
       * https://anmonteiro.com/ocaml-h2/h2/H2/Client_connection/index.html#val-create)
       * and the file descriptor that we created above. *)
      Client.TLS.create_connection_with_default ~error_handler socket
      >>= fun connection ->
      (* Once the connection has been created, we can initiate our request. For
       * that, we call the `request` function, which will send the request that
       * we created to the server, and direct its response to either the
       * response handler - in case of a successful request / response exchange
       * - or the (stream-level) error handler, in case our request was
       * malformed. *)
      let request_body =
        Client.TLS.request connection request ~error_handler ~response_handler
      in
      (* The `request` function returns a request body that we can write to,
       * but in our case just the headers are sufficient. We close the request
       * body immediately to signal to the underlying HTTP/2 framing layer that
       * we're done sending our request. *)
      Body.Writer.close request_body;
      (* Our call to `Lwt_main.run` above will wait until this promise is
       * filled  before exiting the program. *)
      response_received )

Conformance

One of h2's goals is to be 100% compliant with the HTTP/2 specification. There are currently 3 mechanisms in place to verify such conformance:

  1. Unit tests using the HPACK stories in the http2jp/hpack-test-case repository
  2. Unit tests using the test cases provided by the http2jp/http2-frame-test-case repository.
  3. Automated test runs (in CI) using the h2spec conformance testing tool for HTTP/2 implementations.
    • These test all the Reqd.respond_with_* functions for conformance against the specification.

Performance

h2 aims to be a high-performance, memory-efficient, scalable, and easily portable (with respect to different I/O runtimes) implementation. To achieve that, it takes advantage of the unbuffered parsing interface in Angstrom using off-heap buffers wherever possible, for both parsing and serialization.

Below is a plot of H2's latency profile at a sustained rate of 17000 requests per second over 30 seconds, benchmarked using the vegeta load testing tool.

ocaml-h2

Development

This source distribution provides a number of packages and examples. The directory structure is as follows:

  • examples/: contains example applications using the various I/O runtimes provided in this source distribution.
  • hpack/: contains the implementation of HPACK, the Header Compression specification for HTTP/2.
  • lib/: contains the core implementation of this library, including HTTP/2 frame parsing, serialization and state machine implementations.
  • lib_test/: contains various unit tests for modules in the core h2 package.
  • lwt/: contains an implementation of a Lwt runtime for h2 functorized over the specific input / output channel abstraction such that it can work in either UNIX-like systems or MirageOS.
  • lwt-unix/: contains an Lwt runtime adapter for h2 that communicates over UNIX file descriptors.
  • mirage/: contains a Mirage runtime adapter for h2 that allows using h2 to write unikernels that serve traffic over HTTP/2.
  • spec/: contains example implementations of servers using h2 that respond with the different provided APIs to be used for conformance testing with the h2spec tool.

Cloning the repository

# Use --recurse-submodules to get the test git submodules
$ git clone [email protected]:anmonteiro/ocaml-h2.git --recurse-submodules

Using OPAM

To install development dependencies, pin the package from the root of the repository:

$ opam pin add -n hpack .
$ opam pin add -n h2 .
$ opam install --deps-only h2

After this, you may install a development version of the library using the install command as usual.

Tests can be run via dune:

dune runtest

License

h2 is distributed under the 3-Clause BSD License, see LICENSE.

This source distribution includes work based on http/af. http/af's license file is included in httpaf.LICENSE

More Repositories

1

lumo

Fast, cross-platform, standalone ClojureScript environment
Clojure
1,884
star
2

reason-graphql-fullstack

Fullstack Reason + GraphQL Todo List App
OCaml
248
star
3

aws-lambda-ocaml-runtime

An OCaml custom runtime for AWS Lambda
OCaml
201
star
4

piaf

Web library for OCaml with support for HTTP/1.X / HTTP/2
OCaml
170
star
5

archi

An OCaml library for managing the lifecycle of stateful components and their dependencies
OCaml
48
star
6

gluten

Reusable runtime library for network protocols
OCaml
44
star
7

om-next-fullstack

Fullstack application demoing Om Next server-side rendering.
Clojure
37
star
8

ocaml-quic

OCaml
31
star
9

bucklescript-monorepo

Example of a BuckleScript monorepo setup
Reason
31
star
10

github-stars-lambda-demo

OCaml
28
star
11

gh-feed-reader

OCaml
26
star
12

aemette

A batteries-included Leiningen template for Om
Clojure
24
star
13

dotfiles

Vim Script
23
star
14

now-static-bin

Run arbitrary executables in Now 2.0 Lambdas
Go
22
star
15

reason-graphql-experiment

Reason + GraphQL Fullstack (Native) experiment with Now 2.0
JavaScript
21
star
16

now-custom-runtime

ZEIT Now v2.0 builder for custom AWS Lambda runtimes
TypeScript
21
star
17

eio-ssl

OpenSSL bindings to OCaml EIO
OCaml
18
star
18

devcards-om-next

Devcards extension for writing fully reloadable Om Next cards
Clojure
16
star
19

ocaml-mongodb

OCaml
13
star
20

ocaml-subscriptions-transport-ws

Websocket protocol for exchanging GraphQL requests and responses
OCaml
12
star
21

snoopy-mode

Minor mode for number row unshifted character insertion in Emacs
Emacs Lisp
10
star
22

now-rust

Rust
10
star
23

youtube

Reason
9
star
24

graphql-sdl

OCaml
9
star
25

plomber

Component instrumentation for Om Next
Clojure
9
star
26

rekeys

Example ReasonML app to find JavaScript event keycodes.
OCaml
8
star
27

craftconf-demo

7
star
28

websocket-httpaf

OCaml
6
star
29

anmonteiro.github.io

πŸ’Ύ sources for my blog
Clojure
6
star
30

reason-lwt-tutorial

OCaml
6
star
31

ocaml-cpuinfo

OCaml
5
star
32

cra-reason-template

JavaScript
5
star
33

caqti-eio

OCaml
5
star
34

deathrow

πŸ”« Random last statements by executed offenders - web client
Clojure
4
star
35

M101JS

MongoDB University MongoDB for Node.js Developers course
JavaScript
4
star
36

melange-relay

Use Relay with ReScript.
OCaml
4
star
37

nixerve

OCaml
3
star
38

nowtop

OCaml
3
star
39

js-modules-demo

JS Modules demo with Om Next
Clojure
3
star
40

netlify-staging-example

JavaScript
3
star
41

stdlib-merge

OCaml
2
star
42

kubernetes-nixos

Nix
2
star
43

chital

πŸ‚ Scrape web pages for groups of articles
JavaScript
2
star
44

talks

Clojure
2
star
45

clj-isitup

πŸ’Ί isitup.org client, from your code or command line
Clojure
2
star
46

ocaml-sendfile

OCaml
1
star
47

AroundTheCorneriOS

Swift
1
star
48

diy-lisp

my implementation of diy-lisp (from kvalle/diy-lisp)
Python
1
star
49

lumo-demo

Clojure
1
star
50

next-netlify-starter-2

JavaScript
1
star
51

dune-repro

Nix
1
star
52

netlify-graph-nextjs-example

JavaScript
1
star
53

next-netlify-starter-demo

JavaScript
1
star
54

kpop-stack

TypeScript
1
star