On Mon, 19 Mar 2018, Eric Dumazet wrote:
> I noticed high latencies caused by a daemon periodically reading
> various MSR on all cpus. KASAN kernels would see ~10ms latencies
> simply reading one MSR. Even without KASAN, sending IPI to CPU
> in deep sleep state or blocking hard IRQ in a a long section,
> then waiting for the answer can consume hundreds of usec.
>
> This patch adds rdmsr_safe_on_cpu_resched() which does not spin.
>
> I use this function from msr_read() but future patches might
> convert other callers to use this variant as well.
>
> Overall daemon cpu usage was reduced by 35 %,
> and latencies caused by msr_read() disappeared.
Looking at all call sites. None of them is performance critical and all of
them are in preemptible context.
So we simply can switch the rdmsr_safe_on_cpu() implementation over to wait
mode completely.
> +/* Note: This version spins in smp_call_function_single().
> + * Consider using rdmsr_safe_on_cpu_resched() variant instead.
Bah. This is not networking code. x86 uses sensible comment style :)
Thanks,
tglx
