monoio-rs / monoio-http Goto Github PK

I have a new proposal for the http1 decoder refactor.
Now h1 decoder is a little bit too complicated, and using Framed causes it to allocate and copy memory, which can be avoided.

When Framed decode something, it calls the decoder to decode an item or tell itself that the buffered data is not sufficient. But, the Framed itself does not know at least how much data it should allocate and read, so it may grow buffer and read many times when the body is big. Also, splitting BytesMut makes it hard to reuse the buffer.
Though the Decoder provides a way(buffer.reverse) to save loop{grow, read}, it is really not intuitive and does not guarntee the buffer is filled.
The new proposal is to:

1. Make Framed/FramedRead impl AsyncReadRent like a ReadBuf. It can be impl as(the ReadBuf follows the same rule):
   1. When ReadBuf has data, copy it.
   2. When ReadBuf has no data, and there is little data to read, read to ReadBuf inner buffer and copy it. -> With a big buffer there will be less read syscall.
   3. When ReadBuf has no data, and there is a lot data to read, read from ReadBuf inner IO. -> Saves memcpy.
2. When we need fixed length data, read directly from it in async way.
3. Do not split data from Framed inner buffer in Decoder impl.

The decoder that treats Framed/FramedRead as AsyncReadRent can impl AsyncDecoder of monoio-codec crate. Also, the AsyncDecoder can be improved by changing
fn decode<IO>(&mut self, src: IO) -> Self::DecodeFuture<'_>;
to
fn decode<IO>(&mut self, src: &mut IO) -> Self::DecodeFuture<'_>;

And another thing we can do to monoio-codec is to define a new sync decoder trait, which can tell Framed a more precise hint including the data length.

@ihciah can you pls help implement a tls for http1.1 or http2?

i cant seem to find anything about this for monoio. really looking forward to this implementation.