Re: spipe optimization: avoid socketpair(2) when unnecessary

[Colin Percival <cperciva@tarsnap.com>]

On 04/14/13 16:02, Bill Trost wrote:
> Below is a quick hack to eliminate the socket pair used to
> pushbits in the case that file descriptor 0 really is a
> socket.

Interesting -- do you have a use case where this optimization is
likely to matter?  It seems to me that this is only relevant if
you are pushing a very large number of bits through; and even then,
I'd expect the cryptographic work to overwhelm the data-copying and
keep the speedup from this to a minimum.

> I'm not convinced that this is the "best" way to
> detect the socketness of fd 0, but at least I was able to
> use the getsockopt result in a vaguely constructive way.

I believe the standards-compliant way to detect if a descriptor
is a socket is to fstat it and then use S_ISSOCK on the results.

> [...]
>         /* Set up a connection. */
> -       if (proto_conn_create(s[1], sas_t, 0, opt_f, K, opt_o,
> +       if (proto_conn_create(socket0 ? 0 : s[1], sas_t, 0, opt_f, K, opt_o,

This is a bug since proto_conn_create writes incoming data back to
the same socket as it's reading from -- you'll need to change it to
take separate "in" and "out" sockets and then pass 0 and 1 (if they
are sockets).  In some cases spipe will be called with descriptors 0
and 1 being the same socket, but not always...

-- 
Colin Percival
Security Officer Emeritus, FreeBSD | The power to serve
Founder, Tarsnap | www.tarsnap.com | Online backups for the truly paranoid