I've found a way to reproduce this bug, so this isn't as critical now.

A couple minor corrections:

> 3.  Running "/usr/lib/openisr/readstats 
> /sys/class/openisr/openisra/states" shows all of the cache lines in 
> LOAD_META and STORE_META states, and the output does not change if you 
> run the command several times.

This is only true if the deadlock occurs on write traffic; if it happens
during a read then fewer cache lines (perhaps only one) will be stuck in
LOAD/STORE_META.

> 2.  Run "echo w > /proc/sysrq-trigger" as root.

Older kernels don't support this, and will say so in their dmesg output. 
  In this case use "t" instead of "w".  (SysRq may also be configured 
out of the kernel, in which case /proc/sysrq-trigger won't exist.)

--Benjamin Gilbert