> > If you can migrate the pages. The current Linux MM doesn't support that.
>
> Yes, but it could be done. Fairly easily in fact. Bad pages can be marked
> not present (extra code), or simply "owned" by kswapd (minimal code). As
> with VAX/VMS, a non-kernel task that encountered bad RAM can be killed,
> but the rest of the world survives.
Nope, it's significant work to make page migration work. SGI did it for
different purposes for IRIX and their results are ... ``limited''. The pain
is that it basically requires to go over the whole allocations and to
deciede whether they're migratable or not.
The problem is that with the current mm we cannot find the owner of a
faulting mapped page, so don't know which process to send a SIGBUS to.
Especially the SysV-SHM case is problematic. Even more problematic are
kernel private allocations like vmalloc'ed memory or driver allocations.
Faults for these cannot necessarily be charged to the current process.
That's a pain in the lower part of the body with machines which have *huge*
amounts of memory or for some reason like bad design or out of spec operation
are suffering from a high probability of such errors.
Work to do for Non-Stop Linux (TM) ...
Ralf
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/