Create __alloc_ei_netdev() helper, which takes a size argument for
allocation of driver-private structures.

Use __alloc_ei_netdev in pcnet_cs, for embedded priv struct.

Move all initialization between alloc_ei_netdev() and register_netdev(),
and fix bugs on error paths.  Also s/kfree/free_netdev/

Affected drivers:  ixgb, sk98lin, ibmtr, airport, orinoco, wl3501_cs

1) Ensure alignment of both net_device and private area.
2) Introduce netdev_priv(), an inline which allows the dynamic private
   area (dev->priv) to be calculated as a constant offset from the
   base struct net_device at compile time.

This patch fixes a deadlock which occurs when a PCMCIA card is
physically removed, and the netdev interface is then downed.

The problem occurs because these drivers delay the call of
unregister_netdev until the netdev is downed.

Since 2.6 now downs the interface on unregister_netdev(), we can
call this function as soon as the card has been removed without
waiting for the netdev to be downed.

On Wed, 19 Nov 2003, Sam Creasey wrote:
> On Tue, 18 Nov 2003, Geert Uytterhoeven wrote:
> > On Mon, 17 Nov 2003, Geert Uytterhoeven wrote:
> > > On Sun, 16 Nov 2003, Jeff Garzik wrote:
> > > > Yet more updates.  Syncing with Andrew Morton, and more syncing with Al
> > > > Viro.
> > > >
> > > > No users of init_etherdev remain in the tree.  (yay!)
> > >
> > > Here are some (untested, except for cross-gcc) fixes for the m68k-related
> > > drivers:
> >
> > I forget to test the Sun-3 drivers:
> >   - sun3_82586.c:
> >       o add missing casts to iounmap() calls
> >       o fix parameter of free_netdev()
> >   - sun3lance.c: add missing casts to iounmap() calls
> >
> > Note that sun3_82586.c no longer compiles since SUN3_82586_TOTAL_SIZE is not
> > defined. Sammy, is it OK to use PAGE_SIZE for that, since that's what's passed
> > to ioremap()?
>
> Should be...  I looked back through a few versions of the code, and I'm
> not even sure what SUN3_82586_TOTAL_SIZE even was (appears I commented
> that line out long ago anyway).  (I'm also amazed just how much of that
> driver I've forgotten in the last year or two :)

OK, so here's a additional patch that fixes that:

Converts from static to dynamic allocation, in preparation for
further refcount changes.

On Mon, 17 Nov 2003, Geert Uytterhoeven wrote:
> On Sun, 16 Nov 2003, Jeff Garzik wrote:
> > Yet more updates.  Syncing with Andrew Morton, and more syncing with Al
> > Viro.
> >
> > No users of init_etherdev remain in the tree.  (yay!)
>
> Here are some (untested, except for cross-gcc) fixes for the m68k-related
> drivers:

I forget to test the Sun-3 drivers:
  - sun3_82586.c:
      o add missing casts to iounmap() calls
      o fix parameter of free_netdev()
  - sun3lance.c: add missing casts to iounmap() calls

Note that sun3_82586.c no longer compiles since SUN3_82586_TOTAL_SIZE is not
defined. Sammy, is it OK to use PAGE_SIZE for that, since that's what's passed
to ioremap()?

On Sun, 16 Nov 2003, Jeff Garzik wrote:
> Yet more updates.  Syncing with Andrew Morton, and more syncing with Al
> Viro.
>
> No users of init_etherdev remain in the tree.  (yay!)

Here are some (untested, except for cross-gcc) fixes for the m68k-related
drivers:
  - Space.c: fix incorrect prototypes for atarilance_probe() and mace_probe()
  - a2065.c: kill superfluous argument of alloc_etherdev()
  - apne.c:
      o fix incorrect prototype for apne_probe()
      o kill unused variable err
  - mac8390.c:
      o kill unused variable probed
      o fix typos ENDOEV -> ENODEV and ERR_PTE -> ERR_PTR
      o add missing variable slots
  - macmace.c: use ERR_PTR() where needed
  - macsonic.c: kill unused variable lp
  - mvme147.c:
      o kill conversion warning and kill a cast by making ram unsigned long
      o add missing variable err

Note: The use of `slots' in mac8390.c is not in my tree. Do you know where that
change comes from?

Mostly removing unneeded calls to ether_setup(), which alloc_etherdev()
already does for us.