Re: [PATCH -v2] use per cpu data for single cpu ipi calls

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, 29 Jan 2009, Peter Zijlstra wrote:
> 
> Calls with wait=1 are deadlockable when done from within irqs or irq
> disabled sections because two cpus could cross ipi each other and stay
> waiting.

Sure. But this isn't the "wait=1" case. In the waiting case, we don't need 
to allocate anything at all, we can just use the stack.

So the only case that matters for interrupts is indeed "wait=0", but 
that's very much also the case we're talking about.

> That leaves us with the !wait case, which wasn't safe because that
> kmalloc could fail, and the fallback was wait, which yields the same
> deadlock.

Yes. I agree that the _concept_ of the patch is good. No argument there. 

I'm just worried about the details.

> That leaves us with 6 cases,
> 
> root@laptop:/usr/src/linux-2.6# git grep "smp_call_function_single(.*0)"

This misses a _lot_ of cases. What about this one:

	kernel/relay.c:                 err = smp_call_function_single(i,

where you're not seeing if it's a non-waiting one or not (it's a waiting 
one, but my point is that your grep is missing cases).

Doing

	git grep -w smp_call_function_single | grep -v '1);'

is probably more likely to find the cases to look for.

But you're right, I'm still not seeing anything that looks _likely_ to be 
a problem. But I worry that it's something somebody may want to do.

		Linus