Subject: [boost] [Beast] Redesigned HTTP algorithms and parser!
From: Vinnie Falco (vinnie.falco_at_[hidden])
Date: 2017-03-29 14:05:53
Websocket connections start with an HTTP UPGRADE request/response
sequence. Often people want to customize this handshake so I added
code in Beast to read and write HTTP messages. Upon reviewing the HTTP
RFC (rfc2616 and rfc7230) it seemed to be logical that the HTTP
message could be modeled as a single class. And there could be free
functions implementing the algorithms to serialize and deserialize
these messages in th HTTP/1 format.
I took this as far as I could and I was pretty happy with the result.
You have functions like read, async_read, write, and async_write which
take beast::http::message objects as parameters. Example:
The innovation here is the "Body" template type which controls not
only the type of container used to represent the body but also the
algorithms used to transfer bytes from the message body on the wire
into and out of that container. The Body concept, and its nested
Reader and Writer concepts, are documented here:
Previous efforts and HTTP implementations focused heavily on the
representation and parsing of the headers. To my knowledge Beast is
the first implementation to provide caller flexibility in the body
Now the Beast websocket implementation has a reasonable set of
primitives it could use to perform the handshake, and users could get
control of that process.
Unfortunately, while treating the message as an object that can be
read/written atomically works great for example programs and trivial
use-cases, it fails in the real world. These are issues with it:
* The caller must commit to a Body representation before receiving the
* No way to properly read and respond to Expect: 100-continue
* An efficient HTTP relay function (one that does not buffer the
entire body up front) cannot be implemented by the caller
* The caller cannot reasonably set a timeout, for example what if the
body is 2 gigabytes received over a slow connection?
* The parser interface forces algorithms to perform an unnecessary
buffer copy for body containers that are entirely stored in memory
(e.g std::string or boost::asio::streambuf).
These issues sparked quite a bit of discussion on GitHub:
In October of 2016 these problems became apparent. I tried a lot of
different ideas to improve the design to solve it, but none of them
worked. Some of my co-workers looked at it but we didn't really find
anything great. One user in particular, on GitHub, had strong ideas
involving the concept of "completion conditions" to augment the
read/parse algorithms but it didn't feel right to me.
After suffering from writer's block for 5 months inspiration hit and I
tried a combination of interface changes that got the design to where
it needs to be to solve these problems.
* The Reader concept is refined into two types: "indirect" readers,
and "direct" readers. This is described here:
* The parser exposes a small number of internal states, informing
callers of where it is at and giving the caller control over if and
when the parser advances to the next state:
* The parser is rewritten to work only on linear buffers (allowing it
to be optimized to become 20 times faster if SSE2 instructions are
allowed). To facilitate this requirement, the flat_streambuf class is
provided which works just like a basic_streambuf except it guarantees
a single buffer. If the parser is given multiple buffers, it allocates
memory to flatten it:
* New functions parse_some and async_parse_some allow the caller to
perform incremental parsing. This allows for the implementation of
relay functions, or reading just the header (for Expect:
With these changes callers can get control of the inner loop that
reads the body off the socket or stream, so they can do whatever
custom steps are necessary if they want. They can set timeouts, and
write their own loop that calls parse_some repeatedly. I think it
solves all of these issues.
I've written some test code which demonstrates how direct and indirect
body readers are implemented, and what calling code looks like. I've
also written a prototype "relay" function (read a message from one
socket and efficiently write it to another socket). And there's an
example of how Expect: 100-continue might be implemented. These
prototypes are located here:
How to Help?
These are some controversial changes and I think they could use some
polish and vetting before I merge it to the master branch. Maybe there
is something I haven't thought of, or a use-case it doesn't address.
Perhaps there is a better way, or perhaps there are tweaks to make it
even more usable and understandable.
If some kind soul(s) would look through it and offer some feedback,
positive or negative, that would be enormously helpful.
The formal review for Beast is now scheduled for the first week of
July and like an expectant father I am anxious and would like
everything to go just right! My only objective is to polish up the
library and its documentation so it is in the best possible shape for
Thanks for listening!
Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk