Re: Hmm, IDE DMA seems broken with Etch on my PWS433a (Miata MX5).
On Sat, Sep 01, 2007 at 10:51:07AM -0400, Jay Estabrook wrote:
> Sounds like you could try the CPU swap. But I don't know if the caches
> will run at different speeds, though it's be worth the try. If it
> doesn't come up, or comes up flakey, you'll know... ;-}
Yep I think I will try that since I have two axp164ux boards with 600MHz
CPUs (which so far I haven't managed to make milo work with even though
it should), so I will try the cpu swap and see.
> I think the latest kernels' cmd64x driver is new - there was a call
> recently, on this list IIRC, for help testing the new version, as much
> of the IDE support has been totally rewritten, and there are few
> instances of the CMD646 (?) in the wild beside built into several of
> our Alphas...
Well I am trying to apply the workaround from a 2.4.x patch I found into
the 2.6.18 etch kernel and rebuild that (building as I write this).
Other than being slow in PIO mode the system is now quite stable (unlike
with DMA enabled which caused the HD to vanish as soon as a write
crossed the page boundary (so partition table write no problem, mkfs
Since only writes are affected, reads will still be full normal DMA
speed, and only writes that cross a page boundary will be affected, so
hopefully very minor performance impact if it works.
> > My solutions was to go through the installer to just before the
> > partition tool where networking is setup, return to the menu, load
> > additional installer features, and install the ssh client. I then scp'd
> > the hdparm binary (from the hdparm udeb I extracted on another machine),
> > and ran it to disable dma on the hd. After that the install when just
> > fine. I then chroot'd to /target, and installed hdparm and configured
> > /etc/hdparm.conf to disable dma on the hd on boot. System boots fine
> > with no errors now.
> Clever... :-)
Well after ide=nodma didn't work, I didn't see much other choice (other
than cross compile a "fixed" kernel and regenerate installer, which
seemed like a lot of work).
> > I guess the early miatas require a modified kernel
> > to fix the DMA bug in the cmd64x driver. I am surprised this has never
> > made it into the kernel yet. I guess most people run scsi drives on
> > these machines.
> Yes, I believe they were never sold with IDE HDDs, just CDs.
True, but I can't get the stupid scsi drives to show up. I will
probably buy a new wide scsi cable and terminator and try that, but on
the other hand an 80GB IDE drive is much more useful than a couple of
3.2GB scsi drives.
> If you should decide to try to find a wide SCSI controller, remember
> that SRM doesn't know about Adaptec boards. Best bets are QLogic 1040
> and NCR 875 and prolly 895. I don't know if there are any supported by
> SRM that can do Ultra-3, unfortunately.
Well it has a 1020 board already, other than I can't make the wide
I will probably just avoid scsi (too expensive). I will save my scsi
parts for my sparcs, SGIs and Decstation 5000s instead. Oh and I
suppose the sun3's might like scsi too (I always ran them diskless with
NFS in the past, but just as Xterms).