Stephen Hemminger wrote:

Does the following get rid of the hang?
--------
Recode the transmit completion handling to avoid races between the hardware
status report mechanism and the interrupt handler. Rather than relying on
the index value in the status ring, read the chip register and cleanup
all completed transmits. 

Reduce the transmit lock window smaller to allow more parallelism.

[ patch ]

I'm afraid not. :-(

1) The number of bytes before the hang occurs seems to have _decreased_
    rather than the opposite. This observation is, however, a question
    of statistics, and I have better such with the previous version.

2) Previously, the sequence /sbin/ifdown eth0 - /sbin/ifup eth0 made
    the driver recover. Now, the latter of these commands hangs the
    whole system. The recovery is now /sbin/ifdown eth0 - rmmod sky2
    - modprobe sky2


-- Jon