People from time to time point out a wart in ethernet initialization:

The net_device is allocated and registered to the system in
init_etherdev, which is usually one of the first things an ethernet
driver probe function does.  The net_device's final members are setup at
some time between then and the exit of the probe function.  There is
never a clear point where the net device is available to the system for
use.

Our API already supports a solution -- setup the device, then call
register_netdev.  The patch below adds a helper, alloc_etherdev, to
eliminate duplicate code in drivers.  Ethernet device initialization,
after the patch, should now look like

	dev = alloc_etherdev(sizeof(struct netdev_private));
	... initialize device ...
	... set up net_device struct members ...
	rc = register_netdevice(dev);
	if (rc) /* handle error */
	netif_start_queue(dev);

This makes the ethernet driver look and behave similar to other APIs in
the kernel, and presents a nice and atomic
present-this-netdev-to-the-system operation.

It should be noted that there is a net_device::init function.  IIRC in
the last discussion of this issue, ::init() was mentioned as a possible
solution.  I agree that init() can be used, but I do not like the idea
of using init as a probe function, as the ISA drivers do it.  This means
the ethernet device has the potential of holdling rtnl_lock for a long
time, and staying in the probe function, inside register_netdevice, for
a long time.  And...  I agree that init() can be used as a constructor
for filling in dev->xxx and dev->priv->xxx values, but I think doing so
is pointless: you must fill in certain values before your "constructor"
is called, just so your constructor has enough information to operate.

Patch description:

* Add alloc_etherdev, alloc_trdev, alloc_hippi_dev, alloc_fcdev, ...
* Use declarator macros to create init_etherdev, init_trdev, etc., and
remove duplicate code.
* Move net_init EXPORT_SYMBOL from net/netsyms.c to net_init.c.
* Convert drivers/net/8139too.c to use alloc_etherdev, as an example.

Final word, mostly PCI drivers need this change.  ISA drivers use the
net_device::init constructor approach.  IMHO they shouldn't stay in
register_netdevice so long, but that is a wart.  This patch fixes a bug,
the huge span of time between init_etherdev and the indeterminant point
in the future when the net device is ready for use.

Comments?

-- 
Jeff Garzik       | "You see, in this world there's two kinds of
Building 1024     |  people, my friend: Those with loaded guns
MandrakeSoft      |  and those who dig. You dig."  --Blondie