Date: Sat, 30 Sep 1995 21:48:49 +1000
From: Bruce Evans <bde@zeta.org.au>
Message-Id: <199509301148.VAA00595@godzilla.zeta.org.au>
To: davidg@Root.COM, julian@ref.tfs.com
Subject: Re: correctness of isa.c
Cc: hackers@freebsd.org
Sender: owner-hackers@freebsd.org
Precedence: bulk

>The way that I envision this, the 1MB physical memory mapping that starts
>at KERNBASE should completely go away, and each driver should allocate it's
>own kernel VA using pmap_mapdev() and then use that to talk to the shared
>memory. This makes the scheme completely flexible and allows you to map cards
>that are in unusual physical memory areas (like the 15-16MB range). The only
>thing I haven't yet figured out is what to do with the console drivers; these
>need a kernel VA before the VM system is initialized, so you pretty much have
>to kludge up something, at least temporarily during the early startup.

Fake pmap_mapdev() early.  It should probably remap the address to somewhere
that can remain valid forever.  The final address could be attached in
cninit_finish() after VM is initialized, but the old address has to remain
valid until then for debugging.

Shouldn't pmap_mapdev() be declared in a machine-independent header and
used in future drivers in other systems?  It has no i386 dependencies
except for the type of a physical address.  How did old versions of BSD
handle mapping physical addresses for device drivers?  Why doesn't VM
distinguish between the types of physical and virtual addresses?

Bruce