This includes a custom BufReader, since the one in libstd doesn't allow
reading additional data into the buffer without consuming it. This is
required because some connections may send shorter packets, and so we
need to perform multiple reads. After each read, the contents of the
buffer are passed to httparse to see if have a valid message. If so, the
proper amount of bytes are consumed. The additional bytes are left in
the buffer since they are the beginning of the body.
The buffer in this BufReader also grows in size, compared to the libstd
which is sized once. This is because we start with a smaller buffer,
since the majority of messages will be able to include their head in a
packet or 2. Therefore, it's a wasteful performance hit to allocate the
maximum size for every message. However, some headers can be quite big,
and to allow for many of them to be set, we include a maximum size. Once
we've hit the maximum buffer size, and still haven't determined the end
of the headers, a HttpTooLargeError will be returned.
Closes#389
httparse is a http1 stateless push parser. This not only speeds up
parsing right now with sync io, but will also be useful for when we get
async io, since it's push based instead of pull.
BREAKING CHANGE: Several public functions and types in the `http` module
have been removed. They have been replaced with 2 methods that handle
all of the http1 parsing.
BREAKING CHANGE: added requirement that all HeaderFormat implementations
must also be fmt::Debug. This likely as easy as slapping
#[derive(Debug)] on to any custom headers.
All instances of `old_io` and `old_path` were switched to use the new
shiny `std::io`, `std::net`, and `std::path` modules. This means that
`Request` and `Response` implement `Read` and `Write` now.
Because of the changes to `TcpListener`, this also takes the opportunity
to correct the method usage of `Server`. As with other
languages/frameworks, the server is first created with a handler, and
then a host/port is passed to a `listen` method. This reverses what
`Server` used to do.
Closes#347
BREAKING CHANGE: Check the docs. Everything was touched.
This is a modified and specialized thread pool meant for
managing an acceptor in a multi-threaded way. A single handler
is provided which will be invoked on each stream.
Unlike the old thread pool, this returns a join guard which
will block until the acceptor closes, enabling friendly behavior
for the listening guard.
The task pool itself is also faster as it only pays for message passing
if sub-threads panic. In the optimistic case where there are few panics,
this saves using channels for any other communication.
This improves performance by around 15%, all the way to 105k req/sec
on my machine, which usually gets about 90k.
BREAKING_CHANGE: server::Listening::await is removed.
It turns out, we don't need capabilities of MuCell (or even RefCell).
We don't to have sophisticated interior mutability. We don't ever lend
out references that may be mutated later. Instead, if there is no value
to lend, then we generate the value, require interior mutability to save
the value internally, and then we return a reference. We never ever
change a value once it's been generated. It can be changed, but only via
&mut self methods, where we can safely reason about mutability.
This means we don't need keep borrow checking code at runtime, which
helps performance. We also are able to reduce the amount of unsafe
transmutes. The internal API more safely constrains the usage of unsafe,
enforcing our invariants, instead of shotgunning unsafe usage with Refs
and promising we're doing it correctly.
On a machine with lower memory (1GB):
Before:
test bench_mock_hyper ... bench: 251772 ns/iter (+/- 128445)
After:
test bench_mock_hyper ... bench: 152603 ns/iter (+/- 25928)
On a machine with more memory, weaker CPU:
Before:
test bench_mock_hyper ... bench: 129398 ns/iter (+/- 51740)
After:
test bench_mock_hyper ... bench: 115935 ns/iter (+/- 28555)
Closes#252
- Some stray deriving -> derive changes
- use::{mod} -> use::{self}
- fmt.write -> fmt.write_str
This does not catch the last case of fmt.write_str in the
Show impl of a Header Item. This will need to be changed
separately.
- Includes ergonomic traits like IntoUrl and IntoBody, allowing easy
usage.
- Client can have a RedirectPolicy.
- Client can have a SslVerifier.
Updated benchmarks for client. (Disabled rust-http client bench since it
hangs.)
Internals have been shuffled around such that Request and Reponse are
now given only a mutable reference to the stream, instead of being
allowed to consume it. This allows the server to re-use the streams if
keep-alive is true.
A task pool is used, and the number of the threads can currently be
adjusted by using the `listen_threads()` method on Server.
[breaking-change]
Intertwining was a nice feature, but it slows down hyper significantly,
so it is being removed.
There is some fallout from this, mainly that Incoming has had its type
parameter changed to `<A = HttpAcceptor>` and Handler receiving one
bounded with `A: NetworkAcceptor`.
[breaking-change]
Fixes#112
