Abstracting struct ifnet

[Luigi Rizzo]

On Thu, Feb 16, 2012 at 08:16:22PM -0800, Marcel Moolenaar wrote:
> All,
> 
> Juniper is in the final phases of creating a clean separation
> between FreeBSD and Junos, so as to make upgrades of FreeBSD
> easier. This also allows Juniper to track -current and be more
> active FreeBSD contributors.
> 
> To that end, we have a short-term and hopefully short-lived
> problem to solve, which is the ability to use FreeBSD's network
> drivers against the Junos network stack. As some may know, the
> Junos network stack has split up struct ifnet into a physical
> and logical component, called ifdev and iflogical.
> 
> We've tried a few approaches to bridge the gap between ifnet
> on the one hand and ifdev and iflogical on the other and found
> that abstracting ifnet and using accessor functions is the
> best way to allow us to use FreeBSD drivers with the Junos
> network stack, while retaining the ability to use them with
> the FreeBSD stack.
> 
> FreeBSD is also looking at breaking up ifnet and with that in
> mind, I was wondering if there would be any resistance to
> changing network drivers to use accessor functions or macros
> instead of direct pointer dereferences?
> 
> For example, do something like:
> 
> Index: if_fxp.c
> ===================================================================
> --- if_fxp.c	(revision 231178)
> +++ if_fxp.c	(working copy)
> @@ -823,13 +823,14 @@
>  	}
>  
>  	if_initname(ifp, device_get_name(dev), device_get_unit(dev));
> -	ifp->if_init = fxp_init;
> -	ifp->if_softc = sc;
> -	ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
> -	ifp->if_ioctl = fxp_ioctl;
> -	ifp->if_start = fxp_start;
> +	if_set_init(ifp, fxp_init);
> +	if_set_softc(ifp, sc);
> +	if_set_flags(ifp, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST, 0);
> +	if_set_ioctl(ifp, fxp_ioctl);
> +	if_set_start(ifp, fxp_start);
>  
> -	ifp->if_capabilities = ifp->if_capenable = 0;
> +	if_set_capabilities(ifp, 0);
> +	if_set_capenable(ifp, 0);
>  
>  	/* Enable checksum offload/TSO for 82550 or better chips */
>  	if (sc->flags & FXP_FLAG_EXT_RFA) {
> 
> Such a scheme, while initially touching a lot of driver,
> would make it easier to break up ifnet *and* also make it
> easier to hide ABI/API changes from driver vendors (esp.
> when the accessor functions are non-inlined functions and
> not macros or inlines). This is particularly useful for
> Juniper, where we have worked towards network stacks as
> (pre-)loadable modules so as to help with migration and
> validation.
> 
> Thoughts, feedback and suggestion are welcome,

I do like the idea, but the amount of changes will be massive
(see below). The thing that worries me the most is that it
will introduce huge changes between different releases, unless
we backport the accessors (while keeping the underlying struct ifnet
frozen so we preserve the kernel ABI).

To count the number of lines affected, in shell:

    $ for i in if_addrhead if_index if_multiaddrs if_input if_output \
	    if_snd if_addr_lock if_init if_softc if_flags if_ioctl if_start \
	    if_capenable if_addr if_refcount if_fib if_drv_flags \
	    if_data if_index_reserved if_l2com if_link if_xname if_dname
	do
	    printf "%-20s" $i; grep -r $i  ~/FreeBSD/head/sys | wc -l
	done

And you should get something like this

if_addrhead              111
if_index                 231
if_multiaddrs            144
if_input                 184
if_output                112
if_snd                   976
if_addr_lock              12
if_init                  356
if_softc                1462
if_flags                1785
if_ioctl                 259
if_start                 194
if_capenable            1103
if_addr                  324
if_refcount                6
if_fib                    27
if_drv_flags            1987
if_data                  202
if_index_reserved          2
if_l2com                 346
if_link                  187
if_xname                 431
if_dname                  20

cheers
luigi