EISA in GENERIC

[Warner Losh]

On Apr 16, 2014, at 10:36 AM, Matthew Rezny <matthew at reztek.cz> wrote:

>> Warner Losh <imp at bsdimp.com> writes:
>>> The time has come to trim EISA from the generic i386 kernel.
>> 
>> Can we also remove ISA network adapters?
>> 
>> DES
>> --
>> Dag-Erling Smørgrav - des at des.no
> 
> Remove from GENERIC? Go ahead.
> Remove from src tree? No.
> 
> I see no reason to need ISA NICs or storage support in GENERIC on amd64 and 
> little reason have it in GENERIC for i386. General ISA support needs to remain 
> for sio/uart and lpt. Anyone using ISA network, storage, sound, etc is 
> probably ok building a custom kernel.
> 
> My fist thought when I saw this thread was "I hope someone doesn't ask to 
> remove ISA support".

ISA support simply cannot be removed from the tree, full stop. It is our generic
bus on x86, and many things that don’t have a better attachment are attached
here. While some movement towards ACPI has happened here, we haven’t
done a good job of cutting the cord here because things work well enough
for people to focus on other things.

Most, but not all, ISA network, storage and sound drivers have loadable
modules, so could easily be loaded at boot should someone wish to run GENERIC
and still have these devices supported. So there’s a back-door to getting
a system recovered from modern boot media (assuming, of course, that
these systems could boot off modern boot media).

In addition, many of the ISA drivers are also PC Card drivers. I still get
many questions about PC Card devices and drivers. At least with PC Card,
though, we could automatically load the right .ko 95% of the time. But this
gets us back to the no generic bus-level matching code infrastructure that
could be mined to get this data. USB does this with a set of scripts that
are kludgy in the sense they are only for USB. PCI could do it, but some
drivers would need a lot of work. USB, PC Card and ISA PNP could do it,
with some special marking of the arrays. Old-school ISA, where we have
to have identify routines or hints can never be automatically loaded in
response to plug and play data from the bus (but perhaps could be loaded
based on the presence of hints).

So it is for all these reasons I don’t want to talk about ISA at this time. It
is a much messier situation, with ugly tendrils into weird places. EISA,
on the other hand, is very well contained and very easy to turn on/off
once the right build knobs are in place (since it already was close to
being trivial to get rid of).

Warner