Subject: Re: Intel gigabit card with wm driver
To: Allen Briggs <briggs@netbsd.org>
From: Greg A. Woods <woods@weird.com>
List: tech-net
Date: 01/13/2005 03:46:27
[ On Wednesday, January 12, 2005 at 11:06:16 (-0500), Allen Briggs wrote: ]
> Subject: Re: Intel gigabit card with wm driver
>
> On Wed, Jan 12, 2005 at 03:03:35AM -0500, Greg A. Woods wrote:
> > I've been trying verious models of Intel PRO 1000 cards on an
> > AlphaServer ES40 running NetBSD.
> > 
> > I haven't had a whole lot of luck getting them to really perform
> > properly, but apparently that's due do deficiencies in the way the DMA
> > is handled for the alpha when there's a lot of physical memory present
> > (the system has a mere 16GB of RAM, half its maximum :-)
> 
> That's largely to deal with 32-bit PCI trying to access a >32-bit
> address space, right?

That's what I would have assumed....

> Assuming that's the case, it seems like there should be a way around
> using bounce buffers for cards that do have 64-bit address support.
> Does the alpha PCI bus have direct access to all of physical RAM?

Yes, sort of, I think.  The 21272 chipset (Typhoon revision in the ES40
model 2) uses window registers (five windows per P-chip) to translate
PCI memory address space into system address space, and is also capable
of doing scatter-gather-mapped DMA.  There's also a "monster window"
which apparently direct-maps PCI address bits 34:0 (presumably 35:0 for
Typhoon revisions) to the same system memory address bits with no window
offset.  The latter is what I would assume is used for full 64-bit PCI
cards.

My understanding is that NetBSD's bus_dma(9) does implement all the
features necessary to make full use of this chipset, including the
scatter-gather capability, though I'm not sure about the "monster
window", but of course it's just another form of direct-mapped DMA.

At least that's what I guess after reading Jason's "A Machine-
Independent DMA Framework for NetBSD" paper a few times through and
after skimming through the 21272 chipset hardware reference manual.

I can't say I have any firm grasp of all the details though!  ;-)

-- 
						Greg A. Woods

H:+1 416 218-0098  W:+1 416 489-5852 x122  VE3TCP  RoboHack <woods@robohack.ca>
Planix, Inc. <woods@planix.com>          Secrets of the Weird <woods@weird.com>