On 01/09/2019 02:59 PM, Matthew Wilcox wrote:
> On Wed, Jan 09, 2019 at 01:54:36PM -0500, Waiman Long wrote:
>> If you read patch 4, you can see that quite a bit of CPU cycles was
>> spent looking up the radix tree to locate the IRQ descriptor for each of
>> the interrupts. Those overhead will still be there even if I use percpu
>> counters. So using percpu counter alone won't be as performant as this
>> patch or my previous v1 patch.
> Hm, if that's the overhead, then the radix tree (and the XArray) have
> APIs that can reduce that overhead. Right now, there's only one caller
> of kstat_irqs_usr() (the proc code). If we change that to fill an array
> instead of returning a single value, it can look something like this:
>
> void kstat_irqs_usr(unsigned int *sums)
> {
> XA_STATE(xas, &irq_descs, 0);
> struct irq_desc *desc;
>
> xas_for_each(&xas, desc, ULONG_MAX) {
> unsigned int sum = 0;
>
> if (!desc->kstat_irqs)
> continue;
> for_each_possible_cpu(cpu)
> sum += *per_cpu_ptr(desc->kstat_irqs, cpu);
>
> sums[xas->xa_index] = sum;
> }
> }
OK, I will try something like that as a replacement of patch 4 to see
how it compares with my current patch.
Thanks,
Longman
