Re: [PATCH 2/2] x86: implement multiple queues for smp function call IPIs

[Jeremy Fitzhardinge <jeremy@goop.org>]

Andi Kleen wrote:
> On Tue, Jul 29, 2008 at 04:32:57PM -0700, Jeremy Fitzhardinge wrote:
>   
>> This adds 8 queues for smp_call_function(), in order to avoid a
>>     
>
> Now that we have per CPU IDT and there's no global bottleneck anymore
> I think it would be actually fine to use
> more than 8 vectors. 32 or 64 might be a better default.
>   

Well, there's no point in having more vectors than CPUs, and a bit of 
doubling up doesn't hurt too much.  So I think 8 is a good default for 
normal sized machines.  But I can see that being able to add more 
vectors for large machines might be helpful.  I guess it really depends 
on what the fan-out is for multicast messages.

I dunno, maybe it makes sense to take numa topology into account, on the 
assumption that 1) most cross-cpu function calls will be tlb flushes now 
(or at least, sending to mm->cpu_vm_mask), and 2) most tlb flushes will 
be between cpus within one node.

>> void native_send_call_func_ipi(cpumask_t mask)
>> {
>> 	cpumask_t allbutself;
>> +	unsigned queue = smp_processor_id() % CONFIG_GENERIC_SMP_QUEUES;
>>     
>
> Does this really always run with preemption disabled?

Think so, but I'll check again.  One of my TODO list items is to check 
whether smp_call_function_mask should disable preemption for itself, or 
at least WARN_ON if its called with preemption enabled.

    J