Re: [Linux-ia64] Re: [PATCH] early machvec init for 2.5

On Wed, Jun 04, 2003 at 05:47:10PM -0700, Grant Grundler wrote:
> More often than not HP chip designers know when they are violating
> the PCI Local bus spec before the design hits silicon.
> I'd expect SGI HW folks to been doing this long enough to know as well.

Well, in talking to them, they didn't seem to think that they were
violating the spec in any way.  They said that since our device wasn't a
PCI to PCI bridge, DMA coherence vs. reads wasn't strictly necessary (it
also allows us to get high performance from drivers that we have control
over, i.e. those in IRIX).

> > Have you done measurements that indicate that the read machine
> > vector (or any machine vector for that matter) incurs a significant
> > overhead?
> 
> It's butt ugly to need this on a (fairly) new CPU architecture that is
> (was?) supposed to be "clean". I guess that was bound to happen
> anyway.

Yeah, I agree, but for awhile I wasn't sure if we could accomodate the
Linux driver API semantics at all, so I'm just happy that we can.  Even
if it is ugly and kludgy, it works.

> I was thinking anyone who could afford SN2 machine would not bother
> running CONFIG_IA64_GENERIC kernels except to install.  And existing
> customers will have to upgrade kernels regardless of which flavor
> kernel+modules they are running to pick up this change.

But Red Hat only supplies generic kernels, right?  So if someone wants
to run stock Red Hat (required by some app vendors), then they'll be
running a generic kernel and will expect it to work.

> But I can't judge how exposed the install process is to this problem.

Probably not much in itself.

Jesse