On Sun, Feb 16, 2003 at 05:39:02PM +0100, Frederik Meerwaldt wrote:

 > I just increased nmbclusters to 1024, hoping that this would have been 
 > the solution for the "fxp0: Device timeout" problem but unfortunately 
 > it wasn't.
 > 
 > As soon as I am causing more or less heavy NFS traffic I keep on 
 > getting this message.
 > 
 > Any further ideas/suggestions?

The Intel i82557 family of Ethernet chips depends on being able to
atomically update 16-bit words in memory, which the Alpha cannot do.

Specifically, in order for the chip to see a new transmit descriptor,
the driver must clear a bit in the *previous* command descriptor (which
has already been given to the chip).  The bit is in a 16-bit word who's
adjacent 16-bit word is the descriptor status word.  Updating the control
word clobbers the status word on platforms like the Alpha, thus creating
a race condition whereby the driver could fail to notice the "completed"
bit in the status word.

-- 
        -- Jason R. Thorpe <thorpej@wasabisystems.com>