New alpha 5.x bug

[Bernd Walter]

On Tue, Nov 04, 2003 at 03:18:48PM -0800, Kris Kennaway wrote:
> On Wed, Nov 05, 2003 at 12:00:40AM +0100, Bernd Walter wrote:
> 
> > > Hmmm.. the package builders are Miatas right? 
> > > 
> > > Are they using a PCI add-in ethernet card?
> > > 
> > > I'm thinking of the Miata's rather fragile DMA stuff in the Pyxis
> > > core logic chip.
> 
> I don't think the corruption happened during network transit.  For
> example, they're pushed to the server via scp, so data corruption of
> the ethernet packet would cause it to bounce off scp's integrity
> protection.

Well - this should be handled at TCP level by just dropping the packet
and waiting for retransmit without the application ever to know.
But if the DMA corrupts unrelated data.
However - it sounds unlikey to not corrupt other kind of data as well.

The question is what kind of bug this is to only trigger on the port
cluster.
vm/pipe sounds reasonable to believe for this kind of corruption, but
it never happened for me.
Possibly I'm not pushing vm that much.
I wonder if setting vm.idlezero_enable=0 has an influence - this will
not change the danger of this bug, but if it changes sympotoms might
give a good indication of the direction to search.

-- 
B.Walter                   BWCT                http://www.bwct.de
ticso at bwct.de                                  info at bwct.de