2.4.18 kernels: no DMA w/ServerWorks LE chipsets
timm at fnal.gov
Fri Aug 2 09:44:51 EDT 2002
On Fri, 2 Aug 2002, Josip Loncaric wrote:
> I've just noticed that our Supermicro 370DLE motherboards with
> ServerWorks LE chipset and dual UltraDMA/33 EIDE interfaces do NOT
> enable DMA on ide-disk devices (under kernel 2.4.18-5smp, RedHat
> distribution). This is puzzling, because DMA worked just fine until
> recently (I think). Unfortunately, enabling DMA manually under the
> 2.4.18-5smp kernel ("hdparm -d1 /dev/hda") can lock up the system.
> Apparently, others have had this problem too
> (http://www.uwsg.iu.edu/hypermail/linux/kernel/0206.0/0518.html thread
> speculates that 2.4.18 drops out of DMA due to a hardware bug in
> ServerWorks OSB4 which at least some older kernels did not trigger).
You can set the DMA to be on at startup by using the file
/etc/sysconfig/harddisks --that is what we do.
The problems with the OSB4 (and later) chipsets are well documented,
although Serverworks has never acknowledged, as far as I know, that
there is a problem, some board manufacturers that use this chipset
(Tyan, for instance) have put a disclaimer on their site saying
that ultraDMA won't work on the IDE ports. The problem is
that the chipset thinks the DMA is still running even though the
drive has turned off DMA. This is a hardware bug and can affect
Windows systems as well as Linux.
What we found is that DMA performance on the earlier kernels
depended strongly on what make of drive was attached. With Seagate
we experienced wholesale file system corruption. With Western Digital
there was no filesystem corruption but we did have almost every
node out of 64 crash and hang over a 3 month period. With IBM
things were fine under the 2.4.9 kernel.
Now with 2.4.18-4 kernel which is what we are running now
we have seen 3 or four of our 64 machines disable the DMA once it
got started. For all but one of them, we have been able to
reboot the machine and it came back fine. For the other one,
we set the DMA on the system disk to MWDMA mode 2 (hdparm -X34).
If I understand right from the kernel maintainers, in the 2.4.18 kernel
they decided to take the safe route and shut off the DMA if there
was any chance of file corruption. Evidently the problem happens
most frequently if you are writing to the system disk and trying
to swap at the same time.
> As a consequence, we get only 3.5-4.5 MB/s on these disks instead of
> 12-22 MB/s which should be the norm.
> It appears that this new kernel problem has not been fixed yet...
> Dr. Josip Loncaric, Research Fellow mailto:josip at icase.edu
> ICASE, Mail Stop 132C PGP key at http://www.icase.edu./~josip/
> NASA Langley Research Center mailto:j.loncaric at larc.nasa.gov
> Hampton, VA 23681-2199, USA Tel. +1 757 864-2192 Fax +1 757 864-6134
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
More information about the Beowulf