Re: [PATCH 01/10] strbuf: Add strbuf_read_noblock

[Jeff King <peff@peff.net>]

On Thu, Sep 17, 2015 at 10:26:19AM -0700, Stefan Beller wrote:

> > If we return -1, though, we have a similar annoyance. If the caller
> > notices a -1 return value and finds EAGAIN, they still may need to check
> > sb->len to see if they made forward progress and have data they should
> > be dealing with.
> 
> If errno == EAGAIN, we know it is a non blocking fd, so we could encode
> the length read as (- 2 - length), such that
> 
> ...-2 the length read if EAGAIN was ignored
> -1 for error, check errno!
> 0 for EOF
> +1... length if we just read it or restarted it due to EINTR.
> 
> The call site should know if it is non blocking (i.e. if the <-2 case can
> happen) and handle it appropriately.

I actually wonder if callers who are _expecting_ non-blocking want to
loop in strbuf_read() at all.

strbuf_read() is really about reading to EOF, and growing the buffer to
fit all of the input. But that's not at all what you want to do with
non-blocking. There you believe for some reason that data may be
available (probably due to poll), and you want to read one chunk of it,
maybe act, and then go back to polling.

You _can_ loop on read until you hit EAGAIN, but I think in general you
shouldn't; if you get a lot of input on this fd, you'll starve all of
the other descriptors you're polling.  You're better off to read a
finite amount from each descriptor, and then check again who is ready to
read.

And then the return value becomes a no-brainer, because it's just the
return value of read(). Either you got some data, you got EOF, or you
get an error, which might be EAGAIN. You never have the case where you
got some data, but then you also got EOF and EAGAIN, and the caller has
to figure out which.

So I think you really want something like:

  ssize_t strbuf_read_once(struct strbuf *sb, int fd, size_t hint)
  {
	ssize_t cnt;

	strbuf_grow(hint ? hint : 8192);
	cnt = xread(fd, sb->buf + sb->len, sb->alloc - sb->len - 1);
	if (cnt > 0)
		strbuf_setlen(sb->len + cnt);
	return cnt;
  }

(where I'm assuming xread passes us back EAGAIN; we could also replace
it with read and loop on EINTR ourselves).

-Peff