The kernel could record what pages are in use by the current
process, plus the previous process if the current process was
just scheduled. It is best to just printk() the address space.
After several NMIs have happened, the sysadmin can use a
statistical tool can examine the log file for patterns.
If a pattern shows up, then some of the worst pages can be
avoided. It would be easy to add a command line option to
avoid bad pages, but it is better if there is no need to reboot.
Then klogd could automate the whole solution.
>> I believe the Machine Check Architecture implemented in the P6
>> and P5 CPUs is what needs to be looked into. If the Machine
>> Check Exception is enabled, information is placed into special
>> registers detailing memory errors that have occurred.
>
> Of course, this would be even better :-)
I will guess: That information is in appendix H, and it has
not been reverse engineered for www.x86.org yet.