Re: Unbuffered socket I/O

[Kevin Ryde]

address@hidden (Ludovic Courtès) writes:
>
> Right.  That's an incompatible change if the _output_ is buffered.
> Input can be buffered, though, without this being visible by users.

Alas, that too is in an incompatible change, because recv! ignores
buffering.

> Fortunately, port buffering doesn't have to be symmetrical (although the
> API allowing to do that is internal---actually, we might want to expose
> and document `scm_fport_buffer_add ()').

Sounds good.  Perhaps `setvbuf-input' and `setvbuf-output' for the two
directions.  They could go in 1.8 too if you're careful with the
implementation.

Might have to check the SCM_BUFLINE and SCM_BUF0 flags are ok though.
Suspect the answer is yes, SCM_BUFLINE being an output side feature,
and SCM_BUF0 already merely a combination read==unbuf + write==unbuf.

> Thus, I propose the following
> change, where sockets are turned into ports whose output is left
> unbuffered and whose input is buffered.

The manual could emphasise that unbuffered is not what you want if
reading piecemeal, but the default should stay as advertised.

> BTW, do you know what the purpose of `fport_wait_for_input ()' is?

Maybe left from the 1.6 cooperative threads.

> It
> does nothing for O_NONBLOCK streams and waits for events otherwise.
> Since, for blocking streams, `read ()' does not return until either EOF
> is reached or at least one octet was read, `fport_wait_for_input ()'
> seems redundant.

Ahh, hang on, I wonder if it's a hack to do an "exit guile" while
blocked within a read(), thus allowing gc to run in other threads.

If that's true then presumably the write side is afflicted too, as
well as various other potentially blocking operations like
read-string!/partial and gethost.