Message-ID: <f05a7f45fa6b88d58502fbdccacd0d64d04762ae.camel@linux.intel.com>
Date:   Fri, 03 Feb 2023 21:29:58 -0800
From:   srinivas pandruvada <srinivas.pandruvada@...ux.intel.com>
To:     "Rafael J. Wysocki" <rafael@...nel.org>
Cc:     linux-pm@...r.kernel.org, linux-kernel@...r.kernel.org,
        daniel.lezcano@...aro.org, rui.zhang@...el.com
Subject: Re: [PATCH v5 4/4] thermal/drivers/intel_powerclamp: Add additional
 module params

On Thu, 2023-02-02 at 17:41 +0100, Rafael J. Wysocki wrote:
> First, I would say "Add two module parameters" in the subject.
> 
> On Wed, Feb 1, 2023 at 7:35 PM Srinivas Pandruvada
> <srinivas.pandruvada@...ux.intel.com> wrote:
> > 
> > In some use cases, it is desirable to only inject idle on certain
> > set
> > of CPUs. For example on Alder Lake systems, it is possible that we
> > force
> > idle only on P-Cores for thermal reasons. Also the idle percent can
> > be
> > more than 50% if we only choose partial set of CPUs in the system.
> > 
> > Introduce module parameters for setting cpumask and max_idle.
> 
> "Introduce 2 new module parameters for this purpose."
> 
> > They can be only changed when the cooling device is inactive. This
> > module
> > already have other module parameters. There is no change done for
> > those parameters.
> 
> s/have/has/
OK

> 
> "... other parameters that are not affected by this change."
> 
> > cpumask (Read/Write): A bit mask of CPUs to inject idle. The format
> > of
> > this bitmask is same as used in other subsystems like in
> > /proc/irq/*/smp_affinity. The mask is comma separated 32 bit
> > groups.
> > Each CPU is one bit. For example for 256 CPU system the full mask
> > is:
> > ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffff
> > ffff
> > The leftmost mask is for CPU 0-32.
> > 
> > max_idle (Read/Write): Maximum injected idle time to the total CPU
> > time
> > ratio in percent range from 1 to 100. Even if the cooling device
> > max_state
> > is always 100 (100%), this parameter allows to add a max idle
> > percent
> > limit. The default is 50, to match the current implementation of
> > powerclamp
> > driver. Also doesn't allow value more than 75, if the cpumask
> > includes
> > every CPU present in the system.
> > 
> > Signed-off-by: Srinivas Pandruvada
> > <srinivas.pandruvada@...ux.intel.com>
> > ---
> > v5
> > New patch
> > 
> >  .../driver-api/thermal/intel_powerclamp.rst   |  22 +++
> >  drivers/thermal/intel/intel_powerclamp.c      | 169
> > ++++++++++++++++--
> >  2 files changed, 173 insertions(+), 18 deletions(-)
> > 
> > diff --git a/Documentation/driver-api/thermal/intel_powerclamp.rst
> > b/Documentation/driver-api/thermal/intel_powerclamp.rst
> > index 3f6dfb0b3ea6..d805e28b7a45 100644
> > --- a/Documentation/driver-api/thermal/intel_powerclamp.rst
> > +++ b/Documentation/driver-api/thermal/intel_powerclamp.rst
> > @@ -26,6 +26,8 @@ By:
> >             - Generic Thermal Layer (sysfs)
> >             - Kernel APIs (TBD)
> > 
> > +       (*) Module Parameters
> > +
> >  INTRODUCTION
> >  ============
> > 
> > @@ -318,3 +320,23 @@ device, a PID based userspace thermal
> > controller can manage to
> >  control CPU temperature effectively, when no other thermal
> > influence
> >  is added. For example, a UltraBook user can compile the kernel
> > under
> >  certain temperature (below most active trip points).
> > +
> > +Module Parameters
> > +=================
> > +
> > +``cpumask`` (RW)
> > +       A bit mask of CPUs to inject idle. The format of the
> > bitmask is same as
> > +       used in other subsystems like in /proc/irq/*/smp_affinity.
> > The mask is
> > +       comma separated 32 bit groups. Each CPU is one bit. For
> > example for a 256
> > +       CPU system the full mask is:
> > +      
> > ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffff
> > ffff
> > +
> > +       The leftmost mask is for CPU 0-32.
> > +
> > +``max_idle`` (RW)
> > +       Maximum injected idle time to the total CPU time ratio in
> > percent range
> > +       from 1 to 100. Even if the cooling device max_state is
> > always 100 (100%),
> > +       this parameter allows to add a max idle percent limit. The
> > default is 50,
> > +       to match the current implementation of powerclamp driver.
> > Also doesn't
> > +       allow value more than 75, if the cpumask includes every CPU
> > present in
> > +       the system.
> 
> I'm not sure if this is driver-api.  It's admin-guide rather IMO.
> 
This whole file can be moved to admin_guide as there is no API.


> > diff --git a/drivers/thermal/intel/intel_powerclamp.c
> > b/drivers/thermal/intel/intel_powerclamp.c
> > index 850195ebe5e0..68830b726da2 100644
> > --- a/drivers/thermal/intel/intel_powerclamp.c
> > +++ b/drivers/thermal/intel/intel_powerclamp.c
> > @@ -37,7 +37,7 @@
> >  #include <asm/mwait.h>
> >  #include <asm/cpu_device_id.h>
> > 
> > -#define MAX_TARGET_RATIO (50U)
> > +#define MAX_TARGET_RATIO (100U)
> >  /* For each undisturbed clamping period (no extra wake ups during
> > idle time),
> >   * we increment the confidence counter for the given target ratio.
> >   * CONFIDENCE_OK defines the level where runtime calibration
> > results are
> > @@ -109,6 +109,135 @@ static const struct kernel_param_ops
> > duration_ops = {
> >  module_param_cb(duration, &duration_ops, &duration, 0644);
> >  MODULE_PARM_DESC(duration, "forced idle time for each attempt in
> > msec.");
> > 
> > +#define DEFAULT_MAX_IDLE       50
> > +#define MAX_ALL_CPU_IDLE       75
> > +
> > +static u8 max_idle = DEFAULT_MAX_IDLE;
> > +
> > +static cpumask_var_t idle_injection_cpu_mask;
> > +
> > +static int allocate_idle_injection_mask(void)
> > +{
> > +       /* This mask is allocated only one time and freed during
> > module exit */
> > +       if (!idle_injection_cpu_mask) {
> 
> I would do
> 
> if (idle_injection_cpu_mask)
>          return 0;
OK

> 
> here and rearrange the rest of the function accordingly.
> 
> > +               if (!zalloc_cpumask_var(&idle_injection_cpu_mask,
> > GFP_KERNEL))
> > +                       return -ENOMEM;
> > +
> > +               cpumask_copy(idle_injection_cpu_mask,
> > cpu_present_mask);
> > +       }
> > +
> > +       return 0;
> > +}
> > +
> > +static int cpumask_set(const char *arg, const struct kernel_param
> > *kp)
> > +{
> > +       int ret;
> > +
> > +       mutex_lock(&powerclamp_lock);
> > +
> > +       /* Can't set mask when cooling device is in use */
> > +       if (powerclamp_data.clamping) {
> > +               ret = -EAGAIN;
> > +               goto skip_cpumask_set;
> > +       }
> > +
> > +       /*
> > +        * When module parameters are passed from kernel command
> > line
> > +        * during insmod, the module parameter callback is called
> > +        * before powerclamp_init(), so we can't assume that some
> > +        * cpumask can be allocated before here.
> > +        */
> > +       ret = allocate_idle_injection_mask();
> 
> Could it be allocated by powerclamp_init(), though?  It is not useful
> before that function runs anyway.
> 
It is required when passed as module param. For example:

#insmod drivers/thermal/intel/intel_powerclamp.ko cpumask=f

If you return error, module will not be loaded at all.

	insmod: ERROR: could not insert
module	drivers/thermal/intel/intel_powerclamp.ko: Invalid parameters

You have to return success then module parameters will be silently
ignored.


> > +       if (ret)
> > +               goto skip_cpumask_set;
> > +
> > +       ret = bitmap_parse(arg, strlen(arg),
> > cpumask_bits(idle_injection_cpu_mask),
> 
> So this would replace the existing idle_injection_cpu_mask even if it
> is going to fail later.

Yes, when user sets 0s in mask. Otherwise I have to keep a separate
cpumask and then copy. The size of cpumask type is 1024 bytes.

I will fix this.

> 
> > +                          nr_cpumask_bits);
> > +       if (ret)
> > +               goto skip_cpumask_set;
> > +
> > +       if (cpumask_empty(idle_injection_cpu_mask)) {
> > +               ret = -EINVAL;
> > +               goto skip_cpumask_set;
> > +       }
> > +
> > +       if (cpumask_equal(cpu_present_mask,
> > idle_injection_cpu_mask) &&
> 
> Should this check be against cpu_online_mask instead?  Arguably
> offline CPUs don't matter here.
It is present mask which takes care of case when you started idle
injection on CPUS when some CPUs were offline and then online while the
session is going on. The idle-inject core takes care of online during
session. Othewise online CPUs don't get idle injection and will run at
100%.


> 
> > +                         max_idle > MAX_ALL_CPU_IDLE) {
> > +               ret = -EINVAL;
> > +               goto skip_cpumask_set;
> > +       }
> > +
> > +       mutex_unlock(&powerclamp_lock);
> > +
> > +       return 0;
> > +
> > +skip_cpumask_set:
> > +       mutex_unlock(&powerclamp_lock);
> > +
> > +       return ret;
> > +}
> > +
> > +static int cpumask_get(char *buf, const struct kernel_param *kp)
> > +{
> > +       if (!idle_injection_cpu_mask)
> > +               return -EINVAL;
> 
> I would return -ENODEV here.
OK

> 
> > +
> > +       return bitmap_print_to_pagebuf(false, buf,
> > cpumask_bits(idle_injection_cpu_mask),
> > +                                      nr_cpumask_bits);
> > +}
> > +
> > +static const struct kernel_param_ops cpumask_ops = {
> > +       .set = cpumask_set,
> > +       .get = cpumask_get,
> > +};
> > +
> > +module_param_cb(cpumask, &cpumask_ops, NULL, 0644);
> > +MODULE_PARM_DESC(cpumask, "Mask of CPUs to use for idle
> > injection.");
> > +
> > +static int max_idle_set(const char *arg, const struct kernel_param
> > *kp)
> > +{
> > +       u8 _max_idle;
> 
> new_max_idle would be slightly better I think.
OK

> 
> > +       int ret = 0;
> > +
> > +       mutex_lock(&powerclamp_lock);
> > +
> > +       /* Can't set mask when cooling device is in use */
> > +       if (powerclamp_data.clamping) {
> > +               ret = -EAGAIN;
> > +               goto skip_limit_set;
> > +       }
> > +
> > +       ret = kstrtou8(arg, 10, &_max_idle);
> > +       if (ret)
> > +               goto skip_limit_set;
> > +
> > +       if (_max_idle > MAX_TARGET_RATIO) {
> > +               ret = -EINVAL;
> > +               goto skip_limit_set;
> > +       }
> > +
> > +       if (idle_injection_cpu_mask &&
> > cpumask_equal(cpu_present_mask, idle_injection_cpu_mask) &&
> > +           _max_idle > MAX_ALL_CPU_IDLE) {
> 
> The same check is done here and in cpumask_set().  Could it be done
> in
> a separate function called from here and from there?
OK.

> 
> > +               ret = -EINVAL;
> > +               goto skip_limit_set;
> > +       }
> > +
> > +       max_idle = _max_idle;
> > +
> > +skip_limit_set:
> > +       mutex_unlock(&powerclamp_lock);
> > +
> > +       return ret;
> > +}
> > +
> > +static const struct kernel_param_ops max_idle_ops = {
> > +       .set = max_idle_set,
> > +       .get = param_get_int,
> > +};
> > +
> > +module_param_cb(max_idle, &max_idle_ops, &max_idle, 0644);
> > +MODULE_PARM_DESC(max_idle, "maximum injected idle time to the
> > total CPU time ratio in percent range:1-100");
> > +
> >  struct powerclamp_calibration_data {
> >         unsigned long confidence;  /* used for calibration,
> > basically a counter
> >                                     * gets incremented each time a
> > clamping
> > @@ -342,6 +471,10 @@ static unsigned int get_run_time(void)
> >         unsigned int compensated_ratio;
> >         unsigned int runtime;
> > 
> > +       /* No compensation for non systemwide idle injection */
> > +       if (max_idle > MAX_ALL_CPU_IDLE)
> > +               return (duration * 100 /
> > powerclamp_data.target_ratio - duration);
> > +
> >         /*
> >          * make sure user selected ratio does not take effect until
> >          * the next round. adjust target_ratio if user has changed
> > @@ -460,21 +593,11 @@ static void trigger_idle_injection(void)
> >   */
> >  static int powerclamp_idle_injection_register(void)
> >  {
> > -       /*
> > -        * The idle inject core will only inject for online CPUs,
> > -        * So we can register for all present CPUs. In this way
> > -        * if some CPU goes online/offline while idle inject
> > -        * is registered, nothing additional calls are required.
> > -        * The same runtime and idle time is applicable for
> > -        * newly onlined CPUs if any.
> > -        *
> > -        * Here cpu_present_mask can be used as is.
> > -        * cast to (struct cpumask *) is required as the
> > -        * cpu_present_mask is const struct cpumask *, otherwise
> > -        * there will be compiler warnings.
> > -        */
> > -       ii_dev = idle_inject_register_full((struct cpumask
> > *)cpu_present_mask,
> > -                                          idle_inject_update);
> > +       if (cpumask_equal(cpu_present_mask,
> > idle_injection_cpu_mask))
> > +               ii_dev =
> > idle_inject_register_full(idle_injection_cpu_mask,
> > idle_inject_update);
> > +       else
> > +               ii_dev =
> > idle_inject_register(idle_injection_cpu_mask);
> > +
> >         if (!ii_dev) {
> >                 pr_err("powerclamp: idle_inject_register
> > failed\n");
> >                 return -EAGAIN;
> > @@ -510,7 +633,7 @@ static int start_power_clamp(void)
> >         ret = powerclamp_idle_injection_register();
> >         if (!ret) {
> >                 trigger_idle_injection();
> > -               if (poll_pkg_cstate_enable)
> > +               if (poll_pkg_cstate_enable && max_idle <
> > MAX_ALL_CPU_IDLE)
> 
> Why is the additional check needed here?
poll_pkg_cstate_enable is true for single package system, the
additional check prevents starting package c-state polling thread.
But poll_pkg_cstate_enable can be set to false to avoid this check.

Thanks,
Srinivas

> 
> >                        
> > schedule_delayed_work(&poll_pkg_cstate_work, 0);
> >         }
> > 
> > @@ -565,7 +688,7 @@ static int powerclamp_set_cur_state(struct
> > thermal_cooling_device *cdev,
> >         mutex_lock(&powerclamp_lock);
> > 
> >         new_target_ratio = clamp(new_target_ratio, 0UL,
> > -                               (unsigned long) (MAX_TARGET_RATIO -
> > 1));
> > +                               (unsigned long) (max_idle - 1));
> >         if (!powerclamp_data.target_ratio && new_target_ratio > 0)
> > {
> >                 pr_info("Start idle injection to reduce power\n");
> >                 powerclamp_data.target_ratio = new_target_ratio;
> > @@ -656,6 +779,13 @@ static int __init powerclamp_init(void)
> > 
> >         /* probe cpu features and ids here */
> >         retval = powerclamp_probe();
> > +       if (retval)
> > +               return retval;
> > +
> > +       mutex_lock(&powerclamp_lock);
> > +       retval = allocate_idle_injection_mask();
> > +       mutex_unlock(&powerclamp_lock);
> > +
> >         if (retval)
> >                 return retval;
> > 
> > @@ -689,6 +819,9 @@ static void __exit powerclamp_exit(void)
> > 
> >         cancel_delayed_work_sync(&poll_pkg_cstate_work);
> >         debugfs_remove_recursive(debug_dir);
> > +
> > +       if (idle_injection_cpu_mask)
> > +               free_cpumask_var(idle_injection_cpu_mask);
> >  }
> >  module_exit(powerclamp_exit);
> > 
> > --

Powered by blists - more mailing lists