Re: [patch 2/3] scheduler: add full memory barriers upon task switch at runqueue lock/unlock

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

On Mon, 1 Feb 2010, Mathieu Desnoyers wrote:
> 
> The two event pairs we are looking at are:
> 
> Pair 1)
> 
> * memory accesses (load/stores) performed by user-space thread before
>   context switch.
> * cpumask_clear_cpu(cpu, mm_cpumask(prev));
> 
> Pair 2)
>
> * cpumask_set_cpu(cpu, mm_cpumask(next));
> * memory accessses (load/stores) performed by user-space thread after
>   context switch.

So explain why does that smp_mb() in between the two _help_?

The user of this will do a

	for_each_cpu(mm_cpumask)
		send_IPI(cpu, smp_mb);

but that's not an atomic op _anyway_. So you're reading mm_cpumask 
somewhere earlier, and doing the send_IPI later. So look at the whole 
scenario 2:

	cpumask_set_cpu(cpu, mm_cpumask(next));
	memory accessses performed by user-space

and think about it from the perspective of another CPU. What does an 
smp_mb() in between the two do?

I'll tell you - it does NOTHING. Because it doesn't matter. I see no 
possible way another CPU can care, because let's assume that the other CPU 
is doing that

	for_each_cpu(mm_cpumask)
		send_ipi(smp_mb);

and you have to realize that the other CPU needs to read that mm_cpumask 
early in order to do that.

So you have this situation:

	CPU1			CPU2
	----			----

	cpumask_set_cpu
				read mm_cpumask
	smp_mb
				smp_mb
	user memory accessses
				send_ipi

and exactly _what_ is that "smp_mb" on CPU1 protecting against?

Realize that CPU2 is not ordered (because you wanted to avoid the 
locking), so the "read mm_cpumask" can happen before or after that 
cpumask_set_cpu. And it can happen before or after REGARDLESS of that 
smp_mb. The smp_mb doesn't make any difference to CPU2 that I can see. 

So the question becomes one of "How can CPU2 care about whether CPU1 is in 
the mask"? Considering that CPU2 doesn't do any locking, I don't see any 
way you can get a "consistent" CPU mask _regardless_ of any smp_mb's in 
there. When it does the "read mm_cpumask()" it might get the value 
_before_ the cpumask_set_cpu, and it might get the value _after_, and 
that's true regardless of whether there is a smp_mb there or not. 

See what I'm asking for? I'm asking for why it matters that we have a 
memory barrier, and why that mm_cpumask is so magical that _that_ access 
matters so much. 

Maybe I'm dense. But If somebody puts memory barriers in the code, I want 
to know exactly what the reason for the barrier is. Memory ordering is too 
subtle and non-intuitive to go by gut feel.

				Linus