Re: [PATCH] SLUB: use have_arch_cmpxchg()

[Christoph Lameter <clameter@sgi.com>]

On Mon, 27 Aug 2007, Mathieu Desnoyers wrote:

> Hrm, actually, I don't think such have_arch_cmpxchg() macro will be
> required at all because of the small performance hit disabling
> preemption will have on the slow and fast paths. Let's compare, for each
> of the slow path and fast path, what locking looks like on architecture
> with and without local cmpxchg:
> 
> What we actually do here is:
> 
> fast path:
> with local_cmpxchg:
>   preempt_disable()
>   preempt_enable()
> without local_cmpxchg:
>   preempt_disable()
>   local_irq_save
>   local_irq_restore
>   preempt_enable()
>     (we therefore disable preemption _and_ disable interrupts for
>     nothing)

Hmmm..... This is a performance issue for preempt kernels.

> slow path:
> both with and without local_cmpxchg():
>   preempt_disable()
>   preempt_enable()

Here we potentially loose our per cpu structure since the process may be 
rescheduled.

>   local_irq_save()
>   local_irq_restore()
> 
> 
> Therefore, we would add preempt disable/enable to the fast path of
> architectures where local_cmpxchg is emulated with irqs off. But since
> preempt disable/enable is just a check counter increment/decrement with
> barrier() and thread flag check, I doubt it would hurt performances
> enough to justify the added complexity of disabling interrupts for the
> whole fast path in this case.

One additional cacheline referenced in the hot path. Ok the counter may be 
hot as well if this is a preemptible kernel. Nevertheless a cause for 
concern.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/