Polling on sockets in kernel space and struct file

Polling on sockets in kernel space and struct file

[Martin Sustrik]

Hi all,

As part of implementing a new experimental protocol family, we are 
trying to create a socket in kernel. This seems to be easy, just use 
sock_create_kern(). However, the socket returned by this function does 
not have associated file structure; thus it cannot be polled on using 
poll_initwait() and friends.

We have tried to create the appropriate struct file using sock_map_fd(), 
but this has two problems:

1) We do not want our internal socket to be visible in the process 
context, i.e. it should not have a file descriptor.

2) During process exit, we get a kernel BUG from iput() in fs/inode.c:1260.

We then tried another approach using anon_inode_getfile() to get a 
struct file, but this still produces the problem 2) above.

Any help/advice on how to proceed would be appreciated; for reference 
our work in progress can be seen at

http://github.com/sustrik/linux-2.6/blob/sp-v2.6.36/net/sp/af_sp.c

Thanks,

Martin

Re: Polling on sockets in kernel space and struct file

[Eric Dumazet]

Le mercredi 08 décembre 2010 à 15:33 +0100, Martin Sustrik a écrit :
> Hi all,
> 
> As part of implementing a new experimental protocol family, we are 
> trying to create a socket in kernel. This seems to be easy, just use 
> sock_create_kern(). However, the socket returned by this function does 
> not have associated file structure; thus it cannot be polled on using 
> poll_initwait() and friends.
> 
> We have tried to create the appropriate struct file using sock_map_fd(), 
> but this has two problems:
> 
> 1) We do not want our internal socket to be visible in the process 
> context, i.e. it should not have a file descriptor.



> 
> 2) During process exit, we get a kernel BUG from iput() in fs/inode.c:1260.
> 

you could call sock_map_fd() then :


	int fd = sock_map_fd(sock, flags);
	struct file *file = NULL;

	if (fd != -1) {
		file = fget(fd);
		sys_close(fd);  /* still racy */
	}

	if (file) ...

Take a look at net/9p/trans_fd.c

> We then tried another approach using anon_inode_getfile() to get a 
> struct file, but this still produces the problem 2) above.
> 
> Any help/advice on how to proceed would be appreciated; for reference 
> our work in progress can be seen at
> 
> http://github.com/sustrik/linux-2.6/blob/sp-v2.6.36/net/sp/af_sp.c

Re: Polling on sockets in kernel space and struct file

[Martin Lucina]

Hi Eric,

oddly enough our BUG at fs/inode.c seems to have gone away; may be a race
of some sort...

> you could call sock_map_fd() then :
> 
> 
> 	int fd = sock_map_fd(sock, flags);
> 	struct file *file = NULL;
> 
> 	if (fd != -1) {
> 		file = fget(fd);
> 		sys_close(fd);  /* still racy */
> 	}
> 
> 	if (file) ...
> 
> Take a look at net/9p/trans_fd.c

Interesting approach. 

We're currently doing the following -- we add this function to
net/socket.c:

int sock_map_anon(struct socket *sock, const char *name, int flags)
{
	struct file *newfile;
	
	newfile = anon_inode_getfile(name, &socket_file_ops, sock, flags);
	sock->file = newfile;
	return 0;
}

and then call it instead of sock_map_fd() from our sock_create():

	/* Create peer socket and associated file structure */
	rc = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sp->peer);
	if (rc < 0)
		goto out;
	rc = sock_map_anon(sp->peer, "[sp]", 0);
	if (rc < 0)
		goto out_release;

Is this a valid approach to achieve the same thing?

It's not clear what, if anything special we need to do to correctly release
the anonymous inode when we are releasing our socket...  is
sock_release(sp->peer) sufficient?

Also, I see in 9p/trans_fd.c that it's using get_file() to increment the
file refcount; do we need to do this for all sockets/files we create in
kernel space or is this only for the sys_close() hack?

Thanks,

-mato