Efficient copying between sockets

[Brooks Davis]

On Mon, Nov 01, 2004 at 03:42:52AM -0600, Mike Silbersack wrote:
> 
> On Fri, 29 Oct 2004, Ollie Cook wrote:
> 
> >Good afternoon,
> >
> >I am currently writing a potentially high bandwidth (think fileserver)
> >application which will proxy data from one PF_INET socket to another (no 
> >reason
> >it has to be PF_INET, but that's how the application stands).
> 
> >In actual fact, I know in advance exactly how many bytes need to be copied 
> >from
> >one socket to the other, so if there was any way of doing something like:
> >
> > socket_redirect(sock_src, sock_dst, bytes_to_copy);
> >
> >it would be ideal. However I'd be very surprised if such a trivial way to 
> >do
> >that did actually exist.
> >
> >If anyone has any advice at all on a more efficient way to copy data 
> >between
> >sockets I'd be very glad to hear about it. The software is very much 
> >prototype
> >at the moment, but I'd like to make it as efficient as possible from the
> >beginning and this seems like a prime area for optimisation.
> 
> Splicing bytes from one socket's buffer to another socket's buffer should 
> be relatively simple to do inside the kernel, but I don't think it's 
> implemented anywhere at this point in time.
> 
> If you're really need that functionality, the best place to do it would 
> probably be in sendfile; you could just extend it so that if two sockets 
> are passed to it instead of a socket and file, it would do what you've 
> described above.

I've thought about this on and off and think there's a good case for
such a service, however, the sendfile API is a bit too focused for some
of things you might want to do.  I would encourage any intrested person
to do a socket->socket or file->file version[0] with sendfile since it's
already there and both cases would be useful, but I think an API that
allows you to specify multiple pairs is necessicary in the general case.
Think of netcat where you want to hook stdin and stdout to a socket.

-- Brooks

[0] think cp(1) in five syscalls (open, open, sendfile, close, close).

-- 
Any statement of the form "X is the one, true Y" is FALSE.
PGP fingerprint 655D 519C 26A7 82E7 2529  9BF0 5D8E 8BE9 F238 1AD4
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-net/attachments/20041101/6c408b65/attachment.bin