This is a resync of the last patch for 2.5.34 that resulted from this discussion 
(not the original patch):
http://lwn.net/Articles/1642/
The only change was readding the reference to task_info in the beginning of 
common_interrupt.  It had been dropped when we stopped messing with 
preempt_count there.

I've beaten this thing with my normal array of Specweb tests and it is behaving 
so far.  I've booted on an 8-way with and without SMP.

part of Ben's original message:
> Below is a patch against 2.5.20 that implements 4KB stacks for tasks, 
> plus a seperate 4KB irq stack for use by interrupts.  There are a couple 
> of reasons for doing this: 4KB stacks put less pressure on the VM 
> subsystem, reduces the overall memory usage for systems with large 
> numbers of tasks, and increases the reliability of the system when 
> under heavy irq load by provide a fixed stack size for interrupt 
> handlers that other kernel code will not eat into.
> 
> The interrupt stacks are stackable, so we could use multiple 
> 4KB irq stacks.  The thread_info structure is included in each 
> interrupt stack, and has the current pointer copied into it upon 
> entry.

-- 
Dave Hansen
haveblue@us.ibm.com