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

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAJZ5v0haLjsM67QCVg-6i6i7GmBZjhZzFTmm3=r+8gmAfP4qxQ@mail.gmail.com>
Date: Fri, 11 Oct 2024 13:05:18 +0200
From: "Rafael J. Wysocki" <rafael@...nel.org>
To: "David E. Box" <david.e.box@...ux.intel.com>
Cc: hdegoede@...hat.com, ilpo.jarvinen@...ux.intel.com, rjw@...ysocki.net, 
	srinivas.pandruvada@...ux.intel.com, platform-driver-x86@...r.kernel.org, 
	linux-kernel@...r.kernel.org, linux-pm@...r.kernel.org
Subject: Re: [PATCH V2 2/2] platform/x86/intel/pmc: Disable C1 auto-demotion
 during suspend

On Fri, Oct 11, 2024 at 2:36 AM David E. Box
<david.e.box@...ux.intel.com> 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@...ux.intel.com>
> ---
>
> 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);
> --