This is an implementation of HTTP/2 protocol as described in RFC9113 (and RFC7540 before) and RFC7541 (HPACK).
The core part of the library implements writing http2 frames, reading them, and
reacting on read frames. What to do with the frames is managed by application
specific classes derived from CONNECTION
and HTTP2-STREAM
. This part should be
relatively fixed.
In addition, helper functions to create server and to make client calls is provided. Based on these helpers, demo client and demo server is also provided, see client.lisp or server.lisp. The best way to make servers and clients includes decisions on how to handle asynchronicity, window management, etc, and may be still subject to change.
In quicklisp. Tested primarily on sbcl, occasionally on ecl.
Almost all parts of the listed standards implemented, see below for exceptions.
The library definitely allows to implement a client or server, and a simple ones are attached to it as examples.
Currently, all changes happen in v2-main branch that will eventually become version 2 of this. The changes include switch from frames over Lisp stream to frames as an array of octets; this enables - among other - poll based single thread server implementation (as can be seen in the ../tls-server-pocs).
The bundled client requests remote page and returns values of body (converted to string), result status, and response headers.
Improvements to allow for cookie management, … are out of scope of this library, as they would probably be better taken over from some existing HTTP/1.x library or incorporated there.
Improvements to query several resources on one server, and to receive pushes are considered for future, as they are http2 specific.
(ql:quickload 'http2/client)
(http2/client:retrieve-url "https://example.com")
Example https server (that indicates h2 support with ALPN) is provided ; TLS key
and certificate is needed (test script generates a pair when missing). It
handles connections one by thread by default; this is trivial to improve in
several ways (workers, polling…) with customizing *dispatch-fn*
but for now
out of scope of this project.
(ql:quickload 'http2/server/example)
(http2/server-example::run-demo-server :port 1234)
Then point your browser (not eww nor drakma, of course - must support http2) to https://localhost:1234 or
curl -k http://localhost:1234/
Then define handlers for your paths or prefixes, see the example server for a starting point and API documentation for more details.
- No handling of priorities is implemented. This is OK, as these are only suggestions, and they are more or less dropped in RFC9113 anyway.
- Push promises are not implemented in the client. This is OK, they are disabled by default (settings)
- Option to encode headers to Literal Header Field Never Indexed format is not implemented. This is simple to do, but interface would need to be though out.
- The algorithm to split data being sent out to windows and frames can be confused sometimes
- The MAX-HEADER-LIST-SIZE limit is not enforced.
- Some checks on behaviour of the peer are not enforced
See HACKING.org for the details.
The core library used trivial-gray-streams to implement streams over data frames.
Client and server require usocket and cl+ssl to talk over TLS, and bordeaux-threads for concurrency.
Client uses puri to manipulate URLs.
Server also uses cffi directly to check and confirm alpn. The sample server uses parenscript and cl-who to create html; this is an arbitrary choice and your server can use anything else.
Additionally, fiasco is used for testing (and Javascript for browser-side testing).
The code was not written with speed as primary concern. Measurements depend on many tunable factor both on the client and the server.
Having said that, rough test shows that the server on trivial page, single requestor, 10k requests, over TLS) seems to be about 2-3x faster than Hunchentoot and a bit slower than Woo. However, when multiple requests are sent in parallel from one client, it is much faster than Woo (even with the Hunchentoot-style request handling).
What is currently relatively slow is conversion between (utf-8) text and binary data.
Anyway, run or modify the speed test code for your results:
sbcl --script scaffolding/speed-test.lisp
The server should handle any number of requests on one connection (until stream ids are exhausted, but that is a client problem).
Licensed by MIT license.
Some comments are taken over from the RFCs above and copyrighted by RFC contributors. I read the copyright licenses for RFC that this is allowed.
There is an Akamai code on https://github.com/akamai/cl-http2-protocol that supported bigger parts of the drafted HTTP/2 protocol in 2014; apparently hard if not impossible to run now. It used NPN instead of ALPN.