Re: [PATCH V3] arm64: psci: Reduce waiting time for cpu_psci_cpu_kill()

[Mark Rutland <mark.rutland@arm.com>]

On Fri, Oct 18, 2019 at 07:24:14PM +0800, Yunfeng Ye wrote:
> In a case like suspend-to-disk, a large number of CPU cores need to be
> shut down. At present, the CPU hotplug operation is serialised, and the
> CPU cores can only be shut down one by one. In this process, if PSCI
> affinity_info() does not return LEVEL_OFF quickly, cpu_psci_cpu_kill()
> needs to wait for 10ms. If hundreds of CPU cores need to be shut down,
> it will take a long time.

Do we have an idea of roughly how long a CPU _usually_ takes to
transition state?

i.e. are we _just_ missing the transition the first time we call
AFFINITY_INFO?

> Normally, it is no need to wait 10ms in cpu_psci_cpu_kill(). So change
> the wait interval from 10 ms to max 1 ms and use usleep_range() instead
> of msleep() for more accurate schedule.
> 
> In addition, reduce the time interval will increase the messages output,
> so remove the "Retry ..." message, instead, put the number of waiting
> times to the sucessful message.
> 
> Signed-off-by: Yunfeng Ye <yeyunfeng@huawei.com>
> ---
> v2 -> v3:
>  - update the comment
>  - remove the busy-wait logic, modify the loop logic and output message
> 
> v1 -> v2:
>  - use usleep_range() instead of udelay() after waiting for a while
> 
>  arch/arm64/kernel/psci.c | 7 +++----
>  1 file changed, 3 insertions(+), 4 deletions(-)
> 
> diff --git a/arch/arm64/kernel/psci.c b/arch/arm64/kernel/psci.c
> index c9f72b2665f1..00b8c0825a08 100644
> --- a/arch/arm64/kernel/psci.c
> +++ b/arch/arm64/kernel/psci.c
> @@ -91,15 +91,14 @@ static int cpu_psci_cpu_kill(unsigned int cpu)
>  	 * while it is dying. So, try again a few times.
>  	 */
> 
> -	for (i = 0; i < 10; i++) {
> +	for (i = 0; i < 100; i++) {
>  		err = psci_ops.affinity_info(cpu_logical_map(cpu), 0);
>  		if (err == PSCI_0_2_AFFINITY_LEVEL_OFF) {
> -			pr_info("CPU%d killed.\n", cpu);
> +			pr_info("CPU%d killed by waiting %d loops.\n", cpu, i);

Could we please make that:

			pr_info("CPU%d killed (polled %d times)\n", cpu, i + 1);



>  			return 0;
>  		}
> 
> -		msleep(10);
> -		pr_info("Retrying again to check for CPU kill\n");
> +		usleep_range(100, 1000);

Hmm, so now we'll wait somewhere between 10ms and 100ms before giving up
on a CPU depending on how long we actually sleep for each iteration of
the loop. That should be called out in the commit message.

That could matter for kdump when you have a large number of CPUs, as in
the worst case for 256 CPUs we've gone from ~2.6s to ~26s. But tbh in
that case I'm not sure I care that much...

In the majority of cases I'd hope AFFINITY_INFO would return OFF after
an iteration or two.

Thanks,
Mark.