Subject: Re: Snapshot 20030927 available
To: Izumi Tsutsui <tsutsui@ceres.dti.ne.jp>
From: Christopher SEKIYA <wileyc@rezrov.net>
List: port-mips
Date: 09/28/2003 15:45:00
Hello,

On Sun, Sep 28, 2003 at 02:40:47PM +0900, Izumi Tsutsui wrote:

> BTW, I'd suggest splitting your patch to integrate it into -current,

Sounds okay.  I was waiting for my developer application to go through and
then commit the noncontroversial portions of it myself, but either way is fine.

> - gcc-3.3.1 stuff
>    IMHO, we should commit other stuff (which doesn't require gcc-3.3.1)
>    first.

Yes.  As previously stated, the gcc-3 stuff is required only for the uncompleted
ELF64 loader work.

> - R10000 cache
>    I don't know if this code is right (I'm not a mips guy), but
>    it's "better than nothing" and harmless anyway?
>    (Um, I should try it on my R10000 O2...)

It does the right thing on my r10k O2 and doesn't seem to break anything.

> - R5000sc stuff
>    This should be checked on other ports, too?
>    Or already confirmed it makes things better?
>    (Umm, I should try it on my Tyne or news4000.. (R4600 with L2 cache))
> 
> - MI mips pmap (needs to be reviewed?)
>   - You remove #if defined(MIPS3_L2CACHE_ABSENT), but isn't it better
>     to define options MIPS3_L2CACHE_ABSENT in sgimips/conf/std.sgimips?
>   - flush_pcache_anyway variable seems a bit ugly, IMHO.
>   etc.

Neither one of these is ready to be committed.  They're admittedly very
ugly ... Rafal was going to look at these when he had a chance.

> - cosmetic or trivial one (cpuregs.h, genassym.cf, mips_machdep.c etc.)
>    Should these be committed right now?
> 
> - mace/crime stuff (and other sgimips specific code)
>   I think these diffs are O2 specific (or sgimips specific)
>   so they could be committed at this point, right?

Yes and yes.  The bits in arch/sgimips can go in as-is.

> - ahc diff
>   "#if !defined(sgimips)" in MI drivers would not be acceptable.
>    Is this really required only on sgimips?
>    (IIRC, someone used ahc on arc port without problem.)

I didn't see any other way to do it, given the way ahc_pci was written and the
brain-damaged O2 implementation.

Look at ahc_pci.c around line 1044, you'll see what I mean.  Without some
mechanism of forcing scsiseq to remain zero, you'll end up with async/8
transfers.

There's precedent for this sort of #ifdef -- if_rtk_pci.c, for example, has
a similar bit around line 85.

> - distrib stuff, kernel config files
>    We can leave it pending for now?

The distrib stuff is ready to go now; it's vastly superior to what is currently
in the tree.

Here's how I would prioritize the patches:

(things that can be committed now without breaking other platforms)
* arch/sgimips (in toto) and the cosmetic arch/mips,
* distrib and etc (for sysinst and devices, respectively)

(things that need review)
* ahc_pci

(things that must be maintained as a separate patch until they look pretty)
* arch/mips cache/pmap hacks for r5ksc

Sound good?

-- Chris
	GPG key FEB9DE7F (91AF 4534 4529 4BCC 31A5  938E 023E EEFB FEB9 DE7F)