RE: [PATCH][RFC] ACPI throttling: Save/restore tstate for each CPUs across suspend/resume

From: Chen, Yu C
Date: Mon Nov 28 2016 - 20:27:51 EST


Hi,
> -----Original Message-----
> From: rjwysocki@xxxxxxxxx [mailto:rjwysocki@xxxxxxxxx] On Behalf Of
> Rafael J. Wysocki
> Sent: Wednesday, November 23, 2016 7:03 AM
> To: Chen, Yu C
> Cc: ACPI Devel Maling List; Rui Wang; Linux Kernel Mailing List; Len Brown;
> Rafael J. Wysocki; Pavel Machek; Matthew Garrett; Zhang, Rui; Linux PM
> Subject: Re: [PATCH][RFC] ACPI throttling: Save/restore tstate for each CPUs
> across suspend/resume
>
> On Mon, Nov 14, 2016 at 6:44 PM, Chen Yu <yu.c.chen@xxxxxxxxx> wrote:
> > This is a trial version and any comments are appreciated.
> >
> > Previously a bug was reported that on certain Broadwell platforms,
> > after resuming from S3, the CPU is running at an anomalously low
> > speed, due to BIOS has enabled the throttling across S3. The solution
> > to this is to introduce a quirk framework to save/restore tstate MSR
> > register around suspend/resume, in Commit 7a9c2dd08ead ("x86/pm:
> > Introduce quirk framework to save/restore extra MSR registers around
> > suspend/resume").
> >
> > However more and more reports show that other platforms also
> > experienced the same issue, because some BIOSes would like to adjust
> > the tstate if he thinks the temperature is too high.
> > To deal with this situation, the Linux uses a compensation strategy
> > that, the thermal management leverages thermal_pm_notify() upon resume
> > to check if the Processors inside the thermal zone should be throttled
> > or not, thus tstate would be re-evaluated. Unfortunately on these
> > bogus platforms, none of the Processors are inside any thermal zones
> > due to BIOS's implementation. Thus tstate for Processors never has a
> > chance to be brought back to normal.
> >
> > This patch tries to save/restore tstate on receiving the
> > PM_SUSPEND_PREPARE and PM_POST_SUSPEND, to be more specific, the
> > tstate is saved after thermal_pm_notify(PM_SUSPEND_PREPARE)
> > is called, while it's restored before
> > thermal_pm_notify(PM_POST_SUSPEND),
> > in this way the thermal zone would adjust the tstate eventually and
> > also help adjust the tstate for Processors which do not have thermal
> > zone bound. Thus it does not imapct the old semantics.
> >
> > Another concern is that, each CPU should take care of the save/restore
> > operation, thus this patch uses percpu workqueue to achieve this.
> >
> > Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=90041
> > Reported-by: Matthew Garrett <mjg59@xxxxxxxxxxxxx>
> > Reported-by: Kadir <kadir@xxxxxxxxxxxx>
> > Cc: Len Brown <lenb@xxxxxxxxxx>
> > Cc: "Rafael J. Wysocki" <rafael@xxxxxxxxxx>
> > Cc: Pavel Machek <pavel@xxxxxx>
> > Cc: Matthew Garrett <mjg59@xxxxxxxxxxxxx>
> > Cc: Rui Zhang <rui.zhang@xxxxxxxxx>
> > Cc: linux-pm@xxxxxxxxxxxxxxx
> > Signed-off-by: Chen Yu <yu.c.chen@xxxxxxxxx>
> > ---
> > drivers/acpi/processor_throttling.c | 70
> > +++++++++++++++++++++++++++++++++++++
> > 1 file changed, 70 insertions(+)
> >
> > diff --git a/drivers/acpi/processor_throttling.c
> > b/drivers/acpi/processor_throttling.c
> > index d51ca1c..8ddc7d6 100644
> > --- a/drivers/acpi/processor_throttling.c
> > +++ b/drivers/acpi/processor_throttling.c
> > @@ -29,6 +29,7 @@
> > #include <linux/sched.h>
> > #include <linux/cpufreq.h>
> > #include <linux/acpi.h>
> > +#include <linux/suspend.h>
> > #include <acpi/processor.h>
> > #include <asm/io.h>
> > #include <asm/uaccess.h>
> > @@ -758,6 +759,75 @@ static int acpi_throttling_wrmsr(u64 value)
> > }
> > return ret;
> > }
> > +
> > +#ifdef CONFIG_PM_SLEEP
> > +static DEFINE_PER_CPU(u64, tstate_msr);
>
> Call it saved_tstate_msr maybe?
OK.
>
> > +
> > +static long tstate_pm_fn(void *data)
> > +{
> > + u64 value;
> > + bool save = *(bool *)data;
> > +
> > + if (save) {
> > + acpi_throttling_rdmsr(&value);
> > + this_cpu_write(tstate_msr, value);
> > + } else {
> > + value = this_cpu_read(tstate_msr);
> > + if (value)
> > + acpi_throttling_wrmsr(value);
> > + }
> > + return 0;
> > +}
>
> I would split the above into two functions, one for saving and one for restoring ->
>
OK.
> > +
> > +static void tstate_check(unsigned long mode, bool suspend) {
> > + int cpu;
> > + bool save;
> > +
> > + if (suspend && mode == PM_SUSPEND_PREPARE)
> > + save = true;
> > + else if (!suspend && mode == PM_POST_SUSPEND)
> > + save = false;
> > + else
> > + return;
> > +
> > + get_online_cpus();
> > + for_each_online_cpu(cpu)
>
> -> and decide here which one to invoke.
OK.
>
> > + work_on_cpu(cpu, tstate_pm_fn, &save);
>
> Does work_on_cpu() wait for the work to complete?
>
Yes, it might increase the suspend/resume time, a 'queue_work_on' might be better?
> > + put_online_cpus();
> > +}
> > +
> > +static int tstate_suspend(struct notifier_block *nb,
> > + unsigned long mode, void *_unused) {
> > + tstate_check(mode, true);
> > + return 0;
> > +}
> > +
> > +static int tstate_resume(struct notifier_block *nb,
> > + unsigned long mode, void *_unused) {
> > + tstate_check(mode, false);
> > + return 0;
> > +}
> > +
> > +static int __init tstate_pm_init(void) {
> > + /*
> > + * tstate_suspend should save tstate after
> > + * thermal zone's update in thermal_pm_notify,
> > + * vice versa tstate_resume restore tstate before
> > + * thermal_pm_notify, thus the thermal framework
> > + * has a chance to re-adjust tstate according to the
> > + * temperature trend.
> > + */
> > + pm_notifier(tstate_suspend, -1);
> > + pm_notifier(tstate_resume, 1);
>
> I don't think this is going to do what you really want.
>
> Each of these notifiers is going to be invoked during both suspend and resume,
Yes,
> so I guess you only need one notifier?
Here's my original thought: tstate_suspend needs to be invoked after
thermal_pm_notify, which has a priority of '0',
so the notifier of tstate_suspend should be lower than '0',
thus '-1'. And the same for tstate_resume,
it should be invoked before thermal_pm_notify,
thus priority is '1' ?
>
> > + return 0;
> > +}
> > +
> > +core_initcall(tstate_pm_init);
> > +#endif
> > #else
> > static int acpi_throttling_rdmsr(u64 *value) {
> > --
>
> Thanks,
> Rafael

Thanks,
Yu