Why don’t I implement a nice monadic parser combinator library in Rust? That’s what my thought was when after implementing low-level mock-HTTP server in MIO and actually needed to parse the bytes received by server.

What I wanted is a declative way to define sequence of strings to be matched and/or extracted. Example of a parser for a string representing key-value pair (<string>: <string>) would look like this:

#[derive(Default)]
struct KV {
    k: String,
    v: String,
}

let p: Parser<KV> = Parser::init(KV::default())
    .then(until(':'))
    .save(|target, key| target.k = key)
    .then(exact(": "))
    .save(|target, val| target.v = val);

let stream = "env: prod".to_stream();
let kv: KV = p.parse(stream).unwrap();
// returns KV { k: "env", v: "prod" }

The re-invented stream abstraction is useful when it is required to pull bytes or chunks of bytes from a buffer, with a possibility to reset the position some bytes are read (this allows re-trying matching if needed).

The implementation has a few key points. The Matcher type - just a function that can be applied to a ByteStream. ByteStream itself is just a wrapper over Vec<u8>, nothing fancy.

pub type Matcher<T> = dyn Fn(&mut ByteStream) -> Result<T, MatchError> + 'static;

The single matcher that matches a single character from input ByteStream looks like this:

pub fn single(chr: char) -> Box<Matcher<char>> {
    Box::new(move |bs| {
        let pos = bs.pos();
        bs.next()
            .map(|b| b as char)
            .filter(|c| *c == chr)
            .ok_or(MatchError::unexpected(
                pos,
                format!("EOF"),
                format!("{}", chr),
            ))
    })
}

The repeat matcher is a more generic one, that allows matching zero or more occurrences of an input matcher:

pub fn repeat<T: 'static>(this: Box<Matcher<T>>) -> Box<Matcher<Vec<T>>> {
    Box::new(move |bs| {
        let mut acc: Vec<T> = vec![];
        loop {
            let mark = bs.mark();
            match (*this)(bs) {
                Err(_) => {
                    bs.reset(mark);
                    return Ok(acc);
                }
                Ok(t) => acc.push(t),
            }
        }
    })
}

There are more useful matchers: one, maybe, until, before, token, string, space and bytes. I believe it is possible to infer underlying functionality from the name of a matcher.

To chain two matchers into a new matcher, that represents sequential match made by the first one and then by the second one, there is a chain operator:

// Given Matcher<T> and Matcher<U>, make a Matcher<(T, U)>.
pub fn chain<T: 'static, U: 'static>(
    this: Box<Matcher<T>>,
    next: Box<Matcher<U>>,
) -> Box<Matcher<(T, U)>> {
    Box::new(move |bs| {
        let t = (*this)(bs)?;
        let u = (*next)(bs)?;
        Ok((t, u))
    })
}

To manipulate the matched content, there is a map operator:

// Just apply 'f' on result of a match
pub fn map<T: 'static, U: 'static, F: Fn(T) -> U + 'static>(
    this: Box<Matcher<T>>,
    f: F,
) -> Box<Matcher<U>> {
    Box::new(move |bs| {
        let t = (*this)(bs)?;
        let u = f(t);
        Ok(u)
    })
}

Other operators, like apply, expose and unit follow similar approach.

In current form it is not too convenient to keep bunch of matchers around, so to avoid it I can just “glue” them together for composability into Parser<T>:

pub struct Parser<T> {
    f: Box<Matcher<T>>,
}

impl<T: 'static> Parser<T> {
    pub fn unit(f: Box<Matcher<T>>) -> Parser<T> {
        Parser { f }
    }

    pub fn init<F: Fn() -> T + 'static>(f: F) -> Parser<T> {
        Parser::unit(unit(f))
    }

    pub fn then<U: 'static>(self, that: Box<Matcher<U>>) -> Parser<(T, U)> {
        Parser::unit(chain(self.f, that))
    }

    pub fn map<U: 'static, F: Fn(T) -> U + 'static>(self, f: F) -> Parser<U> {
        Parser::unit(map(self.f, f))
    }
    ...
}

Same operators, but wrapped in Parser::unit actually allow convenient chaining.

At some point I’ve discovered that in order to define next matcher in a chain, glimpse of a current state of the accumulator is required (example: match the HTTP request body as exactly number of bytes in Content-Length header), I would need a way to look “into” the monad. Actually what I was looking for is a flat_map operator. I decided to name it then_with - just because I can.

pub fn expose<T: 'static, U: 'static, F: Fn(&T) -> Box<Matcher<U>> + 'static>(
    this: Box<Matcher<T>>,
    f: F,
) -> Box<Matcher<(T, U)>> {
    Box::new(move |bs| {
        let t = (*this)(bs)?;
        let g = f(&t);
        let u = (*g)(bs)?;
        Ok((t, u))
    })
}
impl<T: 'static> Parser<T> {
...
fn then_with<U: 'static, F: Fn(&T) -> Box<Matcher<U>> + 'static>(self, f: F) -> Parser<(T, U)> {
    Parser::unit(expose(self.f, f))
}
...

For example of usage, here is the implementation of HTTP request parser:

#[derive(Debug)]
struct Header {
    name: String,
    value: String,
}

#[derive(Debug, Default)]
struct Request {
    method: String,
    path: String,
    protocol: String,
    headers: Vec<Header>,
    content: Vec<u8>,
}

fn header_parser() -> Parser<Header> {
    Parser::init(|| vec![])
        .then(before(':'))
        .map(|(mut vec, val)| {
            vec.push(as_string(val));
            vec
        })
        .then(single(':'))
        .map(|(vec, _)| vec)
        .then(single(' '))
        .map(|(vec, _)| vec)
        .then(before('\n'))
        .map(|(mut vec, val)| {
            vec.push(as_string(val));
            vec
        })
        .then(single('\n'))
        .map(|(vec, _)| vec)
        .map(|vec| Header {
            name: vec[0].to_owned(),
            value: vec[1].to_owned(),
        })
}

fn request_parser() -> Parser<Request> {
    Parser::init(|| Request::default())
        .then(before(' '))
        .save(|req, bytes| req.method = as_string(bytes))
        .then(single(' '))
        .skip()
        .then(before(' '))
        .save(|req, bytes| req.path = as_string(bytes))
        .then(single(' '))
        .skip()
        .then(before('\n'))
        .save(|req, bytes| req.protocol = as_string(bytes))
        .then(single('\n'))
        .skip()
        .then(repeat(header_parser().into()))
        .save(|req, vec| req.headers = vec)
        .then(single('\n'))
        .skip()
        .then_with(|req| {
            let len: usize = get_header_value(req, "Content-Length")
                .map(|n| n.parse::<usize>().unwrap())
                .unwrap_or_default();
            bytes(len)
        })
        .save(|req, content| req.content = content)
}

// Checking if it actually works

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn http_request() {
        let text = r#"GET /docs/index.html HTTP/1.1
Host: www.nowhere123.com
Accept: image/gif, image/jpeg, */*
Accept-Language: en-us
Accept-Encoding: gzip, deflate
Content-Length: 8
User-Agent: Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1)

0123456
"#;
        let mut bs = text.to_string().into_stream();
        let req = request_parser().apply(&mut bs).unwrap();

        assert_eq!(req.method, "GET");
        assert_eq!(req.path, "/docs/index.html");
        assert_eq!(req.protocol, "HTTP/1.1");
        assert_eq!(req.content, b"0123456\n");
        assert_eq!(req.headers[0].name, "Host");
        assert_eq!(req.headers[0].value, "www.nowhere123.com");
        assert_eq!(req.headers[1].name, "Accept");
        assert_eq!(req.headers[1].value, "image/gif, image/jpeg, */*");
        assert_eq!(req.headers[2].name, "Accept-Language");
        assert_eq!(req.headers[2].value, "en-us");
        assert_eq!(req.headers[3].name, "Accept-Encoding");
        assert_eq!(req.headers[3].value, "gzip, deflate");
        assert_eq!(req.headers[4].name, "Content-Length");
        assert_eq!(req.headers[4].value, "8");
        assert_eq!(req.headers[5].name, "User-Agent");
        assert_eq!(req.headers[5].value, "Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1)");
    }
...
}

This is easy way to build clean, simple and composable parser combinators, glue them into monad, and define naive HTTP request parser, in pure Rust, without any dependencies. And this naive request parser event seems to work! At least test passes.

Current implementation is “stateless” in a way that if entity is not fully matched (e.g. not all chunks of a request were received), it returns error and next attempts will need to start over. For this reason, the client code should care about keeping the current read position in a stream and resetting the stream to that position if matching did not succeed.

The “stateful” parser could keep the aggregate and repeat matching from last position in a stream. I believe it would require to plug in the State monad somewhere, and allow returning something like (State, Either<Error, Parser<T>>) from parser.

The final code with implemented parsers for HTTP request and WebSocket frame are available on github.