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

[Ingo Molnar <mingo@elte.hu>]

* Linus Torvalds <torvalds@osdl.org> wrote:

> 
> 
> On Wed, 15 Dec 2004, Ingo Molnar wrote:
> > 
> > i ran the stresstest overnight with the 10 KHz NMI, and not a single
> > time did the new branch trigger, out of hundreds of millions of IRQs and
> > NMIs. I think this suggests that the race doesnt exist in current CPUs.
> 
> That may well be true, but I'm not convinced your test is meaningful
> or shows anything.
> 
> The thing is, either the CPU is busy, or it's idle. If it's busy,
> you'll never see this. And if it's idle, it will always be _in_ the
> "halt" instruction.

i deliberately started a test where there was roughly 50% idle time.

> The only way to see the case is in the borderline cases, and if/when
> there are multiple different interrupts (first non-NMI interrupt takes
> it out of the hlt, and then the NMI happens to catch the sti). And
> quite frankly, I don't see how you would stress-test it. A 1kHz timer
> interrupt with a 10kHz NMI interrupt is still very infrequent
> interrupts...

i started an infinite loop that generated disk IRQs, and started a
network test that generated network IRQs. The IRQ rate was roughly
10K/sec - this combined with the 10K/sec NMI rate should be an adequate
mix. (I also made sure that it's really default_idle that is used.)

	Ingo