Re: Socket hack question.

Re: Socket hack question.

[Andi Kleen]

On Thu, Apr 19, 2001 at 12:28:52AM +0200, Joel Eriksson wrote:
> Hello,
> 
> I am a kernel hacking newbie and am struggling to understand the
> networking subsystem. I would like to be able to add a systemcall,
> preferably asynchronous, that connects a socket with a filedescriptor
> (proxy(srcsd, dstfd)) so that everything received on srcsd is directly
> written to dstfd. The proxy should close when srcsd is closed or when
> a zero-size packet is sent (or something like that..).

That syscall already exists -- it's called sendfile. 

-Andi

Socket hack question.

[Joel Eriksson]

Hello,

I am a kernel hacking newbie and am struggling to understand the
networking subsystem. I would like to be able to add a systemcall,
preferably asynchronous, that connects a socket with a filedescriptor
(proxy(srcsd, dstfd)) so that everything received on srcsd is directly
written to dstfd. The proxy should close when srcsd is closed or when
a zero-size packet is sent (or something like that..).

I started digging in the source yesterday, and have read a little
in "The Linux Kernel" document and a few others, but can't say I
feel much wiser. :-) I was thinking about creating a custom "data_ready"
function that I set for the socket to be forwarded (in "struct sock"),
does this make sense? How would I go about to write the data to the
destination file/socket descriptor?

I am of course fully aware that it is not of public interest and
have certainly not intended to try getting it into the mainstream
kernel. :-) I'm just going to use it for a web server I am developing
that communicates with applications via sockets (UNIX domain for
local apps and TCP for remote), portability is not an issue in this
particular case. Another obvious use is for bouncers.

I would like to avoid the kernel -> userspace -> kernel copying
of data just to forward a packet.. The web server should be able
to scale up to a couple of thousands simultaneous connections
without significant performance penalty, so I plan to use multiple
processes or perhaps threads to avoid the bad scalability of
select()/poll().

When I (or if I :-) get more proficient in kernel hacking I will
perhaps try adding support for fully event-driven I/O handling.
Or are there perhaps any similar patches available from somewhere
already?

-- 
Joel Eriksson

Re: Socket hack question.

[Joel Eriksson]

On Wed, Apr 18, 2001 at 11:31:00PM +0200, Andi Kleen wrote:
> On Thu, Apr 19, 2001 at 12:28:52AM +0200, Joel Eriksson wrote:
> > Hello,
> > 
> > I am a kernel hacking newbie and am struggling to understand the
> > networking subsystem. I would like to be able to add a systemcall,
> > preferably asynchronous, that connects a socket with a filedescriptor
> > (proxy(srcsd, dstfd)) so that everything received on srcsd is directly
> > written to dstfd. The proxy should close when srcsd is closed or when
> > a zero-size packet is sent (or something like that..).
> 
> That syscall already exists -- it's called sendfile. 

Actually, I realised the similarities with sendfile() after I made the
post. :-) But I thought sendfile() could only be used for sending data
from a "regular" file descriptor to another file- or socket descriptor..?

The syscall I would like to implement is kind of the reverse to sendfile()
since it should be used to copy data _from_ a socket descriptor to a
file descriptor.

Hmm, I made a small test program and from what I can understand it seems
to verify what I thought:

---
#include <sys/types.h>
#include <sys/socket.h>
#include <unistd.h>
#include <errno.h>
#include <stdio.h>

int main(void)
{
        pid_t pid;
        int fdarr[2];

        if (socketpair(AF_UNIX, SOCK_STREAM, 0, fdarr) ÿ-1) {
                perror("socketpair");
                return 1;
        }

        if ((pid ÿork()) ÿ-1) {
                perror("fork");
                return 1;
        } else if (pid) {
                close(fdarr[0]);
                if (sendfile(1, fdarr[1], NULL, 5) ÿ-1) {
                        perror("sendfile");
                        return 1;
                }
        } else {
                close(fdarr[1]);
                write(fdarr[0], "Test\n", 5);
        }

        return 0;
}
---

The output from the program is "sendfile: Invalid argument"..

> -Andi

-- 
Joel Eriksson

Re: Socket hack question.

[Joel Eriksson]

On Wed, Apr 18, 2001 at 07:31:55PM -0400, Mark Hahn wrote:
> > post. :-) But I thought sendfile() could only be used for sending data
> > from a "regular" file descriptor to another file- or socket descriptor..?
> 
> he said the syscall (ie, interface) already existed,
> not that it was implemented how you want it.

Well, that's right, I could still use the sendfile() interface. If I
would like to implement sendfile() for socket -> [file|socket], would
it be possible to do something like in the following pseudocode:

if srcfd is socket then
	s = sock struct for srcfd
	s->foo_member = dstfd
	s->data_ready = my_data_ready
	block until srcfd is closed

Where foo_member was added by me to the sock struct and my_data_ready
writes to dstfd before calling the default data_ready function. To
follow the sendfile() semantics I should add a max_forward or something
too. Btw, how would I accomplish the block until srcfd is closed, is
it possible (ok, anything is _possible_ but..)?

-- 
Joel Eriksson