Caesar,
On Fri, Jun 5, 2015 at 8:11 AM, Caesar Wang <wxt at rock-chips.com> wrote:
> In idle mode, core1/2/3 of Cortex-A17 should be either power off or in
> WFI/WFE state.
> we can delay 1ms to ensure the CPU enter WFI/WFE state.
>
> Signed-off-by: Caesar Wang <wxt at rock-chips.com>
> ---
>
> arch/arm/mach-rockchip/platsmp.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/arch/arm/mach-rockchip/platsmp.c b/arch/arm/mach-rockchip/platsmp.c
> index 25da16f..6672fdd 100644
> --- a/arch/arm/mach-rockchip/platsmp.c
> +++ b/arch/arm/mach-rockchip/platsmp.c
> @@ -325,6 +325,9 @@ static void __init rockchip_smp_prepare_cpus(unsigned int max_cpus)
> #ifdef CONFIG_HOTPLUG_CPU
> static int rockchip_cpu_kill(unsigned int cpu)
> {
> + /* ensure CPU can enter the WFI/WFE state */
> + mdelay(1);
This is a pretty weak assurance. Is there any stronger assurance you
can give that we're in WFI/WFE state and won't come out of it?
Do you actually see problems if you power off a CPU when it's not in
WFI/WFE state?
...so I _think_ I see the path that is happening here and what you're
trying to handle. Specifically, I see:
On dying CPU:
1. cpu_die() calls 'complete(&cpu_died)'
2. cpu_die() calls 'smp_ops.cpu_die(cpu)' AKA rockchip_cpu_die()
3. rockchip_cpu_die() does a bit more cache flushing before looping in
cpu_do_idle()
The problem is that the moment the completion happens in step #1 above
the dying CPU can be killed. ...so you're trying to make sure the
dying CPU makes it to cpu_do_idle(). In that case a fixed mdelay(1)
might be OK since the time that the CPU takes to run through a few
instructions (with no interrupts) is pretty predictable. It would be
really nice if the commit message went through all this, though.
...but is there any chance that cpu_do_idle() could somehow return?
We shouldn't send any events since we've marked the core offline, but
perhaps some per-core interrupt (arch timer?) that didn't get
migrated?
-Doug
