Re: [patch, 2.6.10-rc3] safe_hlt() & NMIs

[Andrea Arcangeli <andrea@suse.de>]

On Tue, Dec 14, 2004 at 11:47:06PM +0100, Ingo Molnar wrote:
> find the correct patch below. I've tested it with an NMI watchdog
> frequency artificially increased to 10 KHz, and i've instrumented the

Nice test, it'd be nice to trigger it in real life.

on the lines of the 64k movl ss, I wonder if we could create an huge
piece of memory like this:

new_htl:
cli
sti
htl
cli
sti
htl
[..]
jmp original_hlt


and to call new_htl from original_hlt instead of sti;hlt. A dozen megs
of the above should boost the probability of getting interrupted in
"hlt" quite a bit.

However even if the nmi can execute on top of the "hlt" instruction, it
doesn't necessairly mean the next pending irq will execute before
executing 'hlt' too, so it'd need a bit more of instrumentation to as
well track down the race as happening (it's not enough to see the branch
in the nmi handler to be taken). The additional instrumentation should
be quite easy though, just copying the same nmi code to the irq handler
should do the trick.