When doing the audit for this change, it was obvious that several drivers
allocate but never free the net_device.

This fixes these drivers.  This patch is riskier than the earlier
ones, because it isn't just a simple substitution and maybe there
is a reason they never free.

Drivers that use kfree in destructor are easily converted
to use free_netdev.

Replace kfree with free_netdev in cleanup routines
of protocols and network pseudo drivers.

s/kfree/free_netdev/ in cleanup routines of a bunch
of different network drivers.

Replace kfree with free_netdev

s/kfree/free_netdev/ as appropriate for tokenring drivers.

Simple one line substitution of kfree with free_netdev
for the bulk of the network drivers.

This patch adds the free_netdev function and associated
changes so that net_device structures are not freed until
last reference to the network device class is released.

This gets it potentially closer to working, if somebody could
just test it...

into home.osdl.org:/home/torvalds/v2.5/linux

into home.osdl.org:/home/torvalds/v2.5/linux

1) icmpv6_send() and icmpv6_echo_reply() never release dst.
2) ip6_{push,flush}_pending_frames() leak np->cork.rt.

1) Make sym2_template.module get setup properly.  It was
   the scsi_module.c code taking care of this for us previously.
2) Zero out on-stack structures, in particular sym_dev not being
   zeroed out was causing sometimes-boots-sometimes-doesnt problems
   for me on sparc64 with sym2 non-modular.

Based upon a patch from Herbert Xu (herbert@gondor.apana.org.au).

into raven.il.steeleye.com:/home/jejb/BK/scsi-for-linus-2.6

into nuts.ninka.net:/home/davem/src/BK/net-2.5

into nuts.ninka.net:/home/davem/src/BK/net-2.5

into nuts.ninka.net:/home/davem/src/BK/sparc-2.5

This should fix the MCA problems.

I moved the name field to the struct mca_device because it was in such
extensive use, and this approach caused the least impact.