Re: [PATCH] http-backend: buffer headers before sending

From: Eric Wong <e@80x24.org>
Date: Aug 10, 2016, 20:36 UTC
Message-ID: <20160810203611.GA24099@starla>
In-Reply-To: <20160810123201.ylfsnzmubpmtyoaa@sigill.intra.peff.net>

Jeff King <peff@peff.net> wrote:

Show 19 quoted lines

> On Tue, Aug 09, 2016 at 11:47:31PM +0000, Eric Wong wrote:
> 
> > Avoid waking up the readers for unnecessary context switches for
> > each line of header data being written, as all the headers are
> > written in short succession.
> > 
> > It is unlikely any HTTP/1.x server would want to read a CGI
> > response one-line-at-a-time and trickle each to the client.
> > Instead, I'd expect HTTP servers want to minimize syscall and
> > TCP/IP framing overhead by trying to send all of its response
> > headers in a single syscall or even combining the headers and
> > first chunk of the body with MSG_MORE or writev.
> > 
> > Verified by strace-ing response parsing on the CGI side.
> 
> I don't think this is wrong to do, but it does feel like it makes the
> code slightly more brittle (you have to pass around the strbuf and
> remember to initialize it and end_headers() when you're done), for not
> much benefit.

Yes, I was nervous about some of these changes and had to look it over several times. I think the problem was the code initially assumed global state while this change localizes it; so this ought to make future changes easier.

Thanks to you and Junio for the extra eyes.

Show 5 quoted lines

> Using some kind of buffered I/O would be nicer, as then you would get
> nice-sized chunks without having to impact the code. I wonder if just
> using stdio here would be that bad. The place it usually sucks is in
> complex error handling, but we don't care about that at all here (I
> think we are basically happy to write until we get SIGPIPE).

stdio to a non-blocking pipe (if so setup by a caller) could be problematic portability-wise. And reliance on SIGPIPE would hurt reusability if this were eventually factored into a standalone daemon using libmicrohttpd or similar.

> I dunno. I suspect the performance improvement from your patch is
> marginal, but it's not like the resulting code is all _that_ complex. So
> I guess I am OK either way, just not enthused.

Yes, marginal, but I still get annoyed to see extra lines from strace (maybe I trace syscalls too much :x). But I think it's also a baby step which opens up potential for the future. I have nothing planned at the moment, but who knows in a year or five...

Show 10 quoted lines

> > ---
> >   I admit I only noticed this because I was being lazy when
> >   implementing the reader-side on an HTTP server by making
> >   a single read(2) call :x
> 
> The trouble is that your HTTP server is still broken. Now it's just
> broken in an unpredictable and racy way, because the OS may still split
> the write at PIPE_BUF boundaries. (Though given that this is not in the
> commit message, I suspect you know this patch is not an excuse not to
> fix your HTTP server).

Of course :)