Re: HEADS UP: KBI breakage for Ethernet modules

On Tue, May 29, 2007 at 09:18:39AM -0600, Scott Long wrote:
> Yar Tikhiy wrote:
> >On Sun, May 27, 2007 at 03:51:29PM +0400, Yar Tikhiy wrote:
> >>As discussed earlier on -net, I'd like to commit the following
> >>patch.  It will bring ether_ioctl() into accord with ioctl() WRT
> >>the type of the command argument.  In our ioctl(), command became
> >>an u_long ages ago, but ether_ioctl() has never been fixed.  With
> >>int and u_long being of different widths on 64-bit arch'es, the
> >>discrepancy can get us in trouble sooner or later.
> >>
> >>In fact, ioctl command coding is very unlikely to change, so it
> >>will continue to fit in 32 bits.  OTOH, the C compiler should be
> >>uneasy about squeezing u_long into int when ether_ioctl() is called
> >>from an if_ioctl handler, so this patch will be a little step on
> >>the way to a warning-free kernel.
> >>
> >>This change will inevitably break the kernel interface to network
> >>modules, so all of them will need rebuilding.
> >
> >I received several positive replies and no negative ones, so the
> >change has just been committed.  In fact, it breaks KBI on 64-bit
> >platforms only.  (Thanks to Ruslan Ermilov for reminding me about
> >that.)  Many thanks to those folks who encouraged the change.
> >
> >Now all Ethernet-related kernel modules need to be rebuilt on 64-bit
> >platforms.  The conventional "make buildkernel" procedure will take
> >care of stock modules, so only 3rd-party modules need some attention.
> >
> 
> How does this affect 32-bit compatibility on amd64?

It shouldn't at all.  The ether_ioctl() function is to be called
by Ethernet drivers only, and they get the ioctl command as u_long
via their ifnet.if_ioctl handlers.  I guess that 32-bit compatibility
is provided by an upper layer, such as the socket layer.

-- 
Yar