Philippe Gerum wrote:

> On Thu, 2006-09-28 at 12:31 +0200, Jan Kiszka wrote:
>> Does it still make sense then to optimise like I proposed?
>>
>
> If you did measure a 5% improvement on low-end x86 hw, then I guess so.
> Moving __ipipe_unstall_root to the arch-dependent section would do, so
> that we could use the most efficient implementation on a case by case
> basis.
>

I'm not seeing 5% alone due to this change, it's a piece in the puzzle. I just
benchmarked the fixed version below against the unmodified one. There is an improvement
of the average hackbench runtime, though it's minimal. The wrong variant I posted first
doesn't show noticeable benefit compared to the correct one. I guess the "critical"
part is switching IRQs off when they were on, and not just enabling them when they
weren't disabled.

So I think the one below could be applied for now, but per-arch optimisations may
remain on the to-do list for the future if we discover archs that could do better with
hand-optimised code here. The x86 assembly looks quite good already.


--- linux-2.6.17.13.orig/kernel/ipipe/core.c
+++ linux-2.6.17.13/kernel/ipipe/core.c
@@ -166,6 +158,7 @@ void __ipipe_unstall_root(void)
 {
 	ipipe_declare_cpuid;
 
+#ifdef CONFIG_SMP
 	local_irq_disable_hw();
 
 	ipipe_load_cpuid();
@@ -174,6 +167,14 @@ void __ipipe_unstall_root(void)
 
 	if (ipipe_root_domain->cpudata[cpuid].irq_pending_hi != 0)
 		__ipipe_sync_pipeline(IPIPE_IRQMASK_ANY);
+#else	/* !CONFIG_SMP */
+	clear_bit(IPIPE_STALL_FLAG, &ipipe_root_domain->cpudata[cpuid].status);
+
+	if (unlikely(ipipe_root_domain->cpudata[cpuid].irq_pending_hi != 0)) {
+		local_irq_disable_hw();
+		__ipipe_sync_pipeline(IPIPE_IRQMASK_ANY);
+	}
+#endif	/* CONFIG_SMP */
 
 	local_irq_enable_hw();
 }