Re: [PATCH V2 2/2] platform/x86/intel/pmc: Disable C1 auto-demotion during suspend

From: Rafael J. Wysocki
Date: Fri Oct 11 2024 - 07:05:42 EST


On Fri, Oct 11, 2024 at 2:36 AM David E. Box
<david.e.box@xxxxxxxxxxxxxxx> wrote:
>
> On some platforms, aggressive C1 auto-demotion may lead to failure to enter
> the deepest C-state during suspend-to-idle, causing high power consumption.
> To prevent this, disable C1 auto-demotion during suspend and re-enable on
> resume.
>
> Signed-off-by: David E. Box <david.e.box@xxxxxxxxxxxxxxx>
> ---
>
> V2 - Remove #define DEBUG
> - Move refactor of cnl_resume() to separate patch
> - Use smp_call_function() to disable and restore C1_AUTO_DEMOTE
> - Add comment that the MSR is per core, not per package.
> - Add comment that the online cpu mask remains unchanged during
> suspend due to frozen userspace.
>
> drivers/platform/x86/intel/pmc/cnp.c | 53 ++++++++++++++++++++++++++++
> 1 file changed, 53 insertions(+)
>
> diff --git a/drivers/platform/x86/intel/pmc/cnp.c b/drivers/platform/x86/intel/pmc/cnp.c
> index 513c02670c5a..f12d4f0f9e93 100644
> --- a/drivers/platform/x86/intel/pmc/cnp.c
> +++ b/drivers/platform/x86/intel/pmc/cnp.c
> @@ -8,6 +8,8 @@
> *
> */
>
> +#include <linux/smp.h>
> +#include <linux/suspend.h>
> #include "core.h"
>
> /* Cannon Lake: PGD PFET Enable Ack Status Register(s) bitmap */
> @@ -206,8 +208,52 @@ const struct pmc_reg_map cnp_reg_map = {
> .etr3_offset = ETR3_OFFSET,
> };
>
> +
> +/*
> + * Disable C1 auto-demotion
> + *
> + * Aggressive C1 auto-demotion may lead to failure to enter the deepest C-state
> + * during suspend-to-idle, causing high power consumption. To prevent this, we
> + * disable C1 auto-demotion during suspend and re-enable on resume.
> + *
> + * Note that, although MSR_PKG_CST_CONFIG_CONTROL has 'package' in its name, it
> + * is actually a per-core MSR on client platforms, affecting only a single CPU.
> + * Therefore, it must be configured on all online CPUs. The online cpu mask is
> + * unchanged during the phase of suspend/resume as user space is frozen.
> + */
> +
> +static DEFINE_PER_CPU(u64, pkg_cst_config);
> +
> +static void disable_c1_auto_demote(void *unused)
> +{
> + int cpunum = smp_processor_id();
> + u64 val;
> +
> + rdmsrl(MSR_PKG_CST_CONFIG_CONTROL, val);
> + per_cpu(pkg_cst_config, cpunum) = val;
> + val &= ~NHM_C1_AUTO_DEMOTE;
> + wrmsrl(MSR_PKG_CST_CONFIG_CONTROL, val);
> + pr_debug("%s: cpu:%d cst %llx\n", __func__, cpunum, val);
> +}

If you decide to retain the debug statements, I generally prefer to
separate them from the rest of the code with empty lines, for clarity.

> +
> +static void restore_c1_auto_demote(void *unused)
> +{
> + int cpunum = smp_processor_id();
> +
> + pr_debug("%s: cpu:%d cst %llx\n", __func__, cpunum,
> + per_cpu(pkg_cst_config, cpunum));
> + wrmsrl(MSR_PKG_CST_CONFIG_CONTROL, per_cpu(pkg_cst_config, cpunum));
> +}
> +

I would add something like

static void s2idle_cpu_quirk(smp_call_func_t func)
{
if (pm_suspend_via_firmware())
return;

preempt_disable();
func();
smp_call_function(func, NULL, 0);
preempt_enable();
}

> void cnl_suspend(struct pmc_dev *pmcdev)
> {
> + if (!pm_suspend_via_firmware()) {
> + preempt_disable();
> + disable_c1_auto_demote(NULL);
> + smp_call_function(disable_c1_auto_demote, NULL, 0);
> + preempt_enable();
> + }
> +

And here I'd just do

s2idle_cpu_quirk(disable_c1_auto_demote);

and analogously below.

> /*
> * Due to a hardware limitation, the GBE LTR blocks PC10
> * when a cable is attached. To unblock PC10 during suspend,
> @@ -218,6 +264,13 @@ void cnl_suspend(struct pmc_dev *pmcdev)
>
> int cnl_resume(struct pmc_dev *pmcdev)
> {
> + if (!pm_suspend_via_firmware()) {
> + preempt_disable();
> + restore_c1_auto_demote(NULL);
> + smp_call_function(restore_c1_auto_demote, NULL, 0);
> + preempt_enable();
> + }
> +
> pmc_core_send_ltr_ignore(pmcdev, 3, 0);
>
> return pmc_core_resume_common(pmcdev);
> --