On Thu, Nov 20, 2003 at 12:23:44AM -0500, Jeff Garzik wrote:
> viro@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx wrote:
> >On Wed, Nov 19, 2003 at 12:33:08PM -0500, Jeff Garzik wrote:
> >
> >>Ok, Al Viro's net driver refcounting work is pretty much complete, and
> >
> >
> >The hell it is. We are through with legacy probes, we are through with
> >init_etherdev(), we are practically through with static struct net_device.
>
> hehe :) I don't mean to suggest that all is clean and pure :)
>
>
> >However, we still have weird allocators (I've got almost all of them
> >done by now, will submit in the next batch) and we still have struct
> >net_device embedded as a field of other structures in several drivers.
>
> Some of that will be interesting. ns83820 for example embedded
> net_device on purpose... Ben seemed to think at the time it gave him
> some speed, a few less pointer derefs and such.
That's fine, but 83820 should be doing that the other way round.
Note that for objects allocated by alloc_netdev() we have
(char*) dev->priv == (char *)dev + const
and constant can be found at compile time _if_ we pad in front of
net_device and add a pointer to allocated block into net_device.
So we can have exactly the same structure (modulo padding) and no extra
dereferencing. All we need is inlined void *net_priv(struct net_device *);
That, BTW, would be a win for other drivers using alloc_...() and not
reassigning ->priv (i.e. majority of those beasts).
|