adding if_dev member to struct ifnet

[Poul-Henning Kamp]

In message <XFMail.20030930103348.jhb at FreeBSD.org>, John Baldwin writes:

>>>Yes, if it helps to remove if_name/if_unit, it is a thing to do. Moreover it 
>>>sounds a good idea to have the if_dev field into the ifnet structure.
>> 
>> Somebody please explain how this would work for non-hardware
>> interfaces like if_loop, if_tun, if_tap etc ?
>> 
>> device_t is what we use to hitch drivers to hardware.
>> 
>> ifnet is what we use to hitch drivers to the netstack.  
>> 
>> They should not be tangled.
>
>You mean like dev_t and device_t shouldn't be tangled like we do
>with si_drv1?  Oh, wait...

I don't think any correctly written driver stores it's device_t in
a dev_t.  It should store it's softc structure, which should contain
pointers to both.  Even if there is a driver which does do that,
it happens inside the device driver, and it does not handicap the
remaining device drivers with its choice.

If you stick a newbus requirement on "struct ifnet" you suddenly
make demands that a lot of our network drivers cannot satisfy.

The problem with propagating newbus above the device drivers is
that we start postulating a specific relationship between logical
devices and physical (ie: a ifnet has exactly one device_t).

There is nothing in the "data-model" of the kernel that says that
a network interface corresponds to exactly one hardware device
and more importantly: there shouldn't be either.

If_tun _has_ no physical device, and it would be totally insane to
invent a device_t for if_tun, considering that it would not serve
any purpose at all, apart from satisfying some peoples craving to
have device_t in all data structures in the system.

Demanding such a relationship will only make our life more difficult
when we get to deal with all the non-standard devices like if_sl,
if_ppp, if_ng, if_tun, if_tap which have no device_t, or musycc, a
network card where you have to juggle two device_t's, one for the
framer and one for the line encoder.

As it is now, device_t and newbus provides a good model for our
attachment of device drivers to hardware, it's not quite perfect,
but it is good enough that nobody can be bothered to sit down and
write something perfect.

We have "struct ifnet" which describes a network interface, it
describes it based on the access model, and it does in fact not
care a hoot what implements that interface, hardware, software or
bongo drums, it doesn't matter.

Similarly, we have "dev_t" to descripe filesystem accessed devices,
and it describes those in the terms of the access model, not in terms
of what is behind them, hardware, software or bongo drums.

If all you want is an extra field in "struct ifnet" to hang driver
information on, then by all means add that field.  As long as you
give it type "void *" and make it private to the driver I have no
problem with that.

-- 
Poul-Henning Kamp       | UNIX since Zilog Zeus 3.20
phk at FreeBSD.ORG         | TCP/IP since RFC 956
FreeBSD committer       | BSD since 4.3-tahoe    
Never attribute to malice what can adequately be explained by incompetence.