Re: [RFC v4 4/6] virtio/vhost-vsock: Use nested lock notation

[Stefan Hajnoczi <stefanha@redhat.com>]

On Thu, Mar 31, 2016 at 01:41:36PM +0200, Claudio Imbrenda wrote:
> Since the socket was created locally and wasn't passed around, nobody can
> have a reference to it yet, so locking it shouldn't block, therefore no
> deadlock should be possible.
> I'm not sure why, if at all, it is needed to lock the child socket there,
> but this patch makes sure that lockdep is happy, and hopefully it doesn't
> mask an actual deadlock.
> 
> Signed-off-by: Claudio Imbrenda <imbrenda@linux.vnet.ibm.com>
> ---
>  net/vmw_vsock/virtio_transport_common.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
> index fa54359..6698b61 100644
> --- a/net/vmw_vsock/virtio_transport_common.c
> +++ b/net/vmw_vsock/virtio_transport_common.c
> @@ -792,7 +792,7 @@ virtio_transport_recv_listen(struct sock *sk, struct virtio_vsock_pkt *pkt)
>  
>  	sk->sk_ack_backlog++;
>  
> -	lock_sock(child);
> +	lock_sock_nested(child, SINGLE_DEPTH_NESTING);

This looks good.  The same approach is used in other network protocols
when processing an incoming connection to a listen socket.

Regarding lock ordering, there is no other case where we take another
socket B's lock while holding A's lock so deadlock is not possible.

I think the VMCI transport has the same issue in
vmci_transport_recv_listen() when there are pending connections.  I've
CCed Jorgen in case he feels it's worth fixing in the VMCI transport.

Stefan

Attachment: signature.asc
Description: PGP signature