Re: [RFC PATCH 11/12] thermal: introduce cooling state arbitrator

[Zhang Rui <rui.zhang@intel.com>]

Hi, Eduardo,

On 日, 2012-06-24 at 18:19 +0300, Valentin, Eduardo wrote:
> Hello Rui,
> 
> On Wed, Jun 20, 2012 at 9:47 AM, Zhang Rui <rui.zhang@intel.com> wrote:
> > On 三, 2012-06-13 at 17:23 +0300, Eduardo Valentin wrote:
> >> Hello Rui,
> >>
> >> On Mon, Jun 11, 2012 at 11:20:39AM +0800, Zhang Rui wrote:
> >> >
> >> > Introduce simple arbitrator for setting device cooling state,
> >> > to fix the problem that a cooling device may be referenced by
> >> > by multiple trip points in multiple thermal zones.
> >> >
> >> >
> >> > With this patch, we have two stages for updating a thermal zone,
> >> > 1. check if a thermal_instance needs to be updated or not
> >> > 2. update the cooling device, based on the target cooling state
> >> >    of all its instances.
> >> >
> >> > Note that, currently, the cooling device is set to the deepest
> >> > cooling state required.
> >> >
> >> > Signed-off-by: Zhang Rui <rui.zhang@intel.com>
> >> > ---
> >> >  drivers/thermal/thermal_sys.c |   41 ++++++++++++++++++++++++++++++++++++++---
> >> >  include/linux/thermal.h       |    1 +
> >> >  2 files changed, 39 insertions(+), 3 deletions(-)
> >> >
> >> > Index: rtd3/drivers/thermal/thermal_sys.c
> >> > ===================================================================
> >> > --- rtd3.orig/drivers/thermal/thermal_sys.c
> >> > +++ rtd3/drivers/thermal/thermal_sys.c
> >> > @@ -54,6 +54,7 @@ struct thermal_instance {
> >> >     int trip;
> >> >     unsigned long upper;    /* Highest cooling state for this trip point */
> >> >     unsigned long lower;    /* Lowest cooling state for this trip point */
> >> > +   unsigned long target;   /* expected cooling state */
> >> >     char attr_name[THERMAL_NAME_LENGTH];
> >> >     struct device_attribute attr;
> >> >     struct list_head tz_node; /* node in tz->instances */
> >> > @@ -812,6 +813,7 @@ int thermal_zone_bind_cooling_device(str
> >> >     dev->trip = trip;
> >> >     dev->upper = upper;
> >> >     dev->lower = lower;
> >> > +   dev->target = -1;
> >> >
> >> >     result = get_idr(&tz->idr, &tz->lock, &dev->id);
> >> >     if (result)
> >> > @@ -949,6 +951,7 @@ thermal_cooling_device_register(char *ty
> >> >     strcpy(cdev->type, type);
> >> >     INIT_LIST_HEAD(&cdev->instances);
> >> >     cdev->ops = ops;
> >> > +   cdev->updated = 1;
> >> >     cdev->device.class = &thermal_class;
> >> >     cdev->devdata = devdata;
> >> >     dev_set_name(&cdev->device, "cooling_device%d", cdev->id);
> >> > @@ -1040,6 +1043,32 @@ void thermal_cooling_device_unregister(s
> >> >  }
> >> >  EXPORT_SYMBOL(thermal_cooling_device_unregister);
> >> >
> >> > +static void thermal_zone_do_update(struct thermal_zone_device *tz)
> >> > +{
> >> > +   struct thermal_instance *instance1, *instance2;
> >> > +   struct thermal_cooling_device *cdev;
> >> > +   int target;
> >> > +
> >> > +   list_for_each_entry(instance1, &tz->instances, tz_node) {
> >> > +           cdev = instance1->cdev;
> >> > +
> >> > +           /* cooling device has already been updated*/
> >> > +           if (cdev->updated)
> >> > +                   continue;
> >> > +
> >> > +           target = 0;
> >> > +           /* Make sure cdev enters the deepest cooling state */
> >> > +           list_for_each_entry(instance2, &cdev->instances, cdev_node) {
> >> > +                   if (instance2->target == -1)
> >> > +                           continue;
> >> > +                   if (instance2->target > target)
> >> > +                           target = instance2->target;
> >> > +           }
> >> > +           cdev->ops->set_cur_state(cdev, target);
> >> > +           cdev->updated = 1;
> >> > +   }
> >> > +}
> >>
> >> I think the above arbitrator solution does not escalate. As I can see,
> >> the arbitrator takes care of cooling device instances in the same thermal zone.
> >>
> > why?
> > It parses the thermal instances in one thermal zone to find the cooling
> > devices, and then parse all the thermal instances for the cooling
> > devices, no matter which thermal zone it belongs to.
> 
> I think I got confused with the nomenclature change. So, yes, it does
> go through all zones and cooling devices. But still, ideally it should
> be a plist per device, don't you think? This way your search space is
> reduced.
> 
I'm not aware of this before, but here is a problem when I tried to use
plist.
First, we should use plist_node->prio as the thermal_instance->target
instead, right?
But the problem is that thermal_instance->target needs to be updated
when temperature is changed. But plist does not support dynamic updating
the priority, right?
So when changing the cooling state target of a thermal instance, I need
to delete and re-insert the plist_node? This does not seems efficient to
me.

thanks,
rui

> >
> >> What if you have a device which expose cooling device instances in different
> >> thermal zones?
> >>
> >> Besides, the thermal constraint might collide with pm QoS or with settings
> >> coming from user space interfaces.
> >>
> >
> >> That's actually why I was suggesting to have this 'arbitrator' or constraint
> >> management outside the thermal framework. And managed per device at some
> >> other abstraction layer.
> >>
> > Yeah, but I think we can have an incremental patch to handle the pm Qos
> > constrains. Especially Durga already has the governor patches to move
> > this out of the thermal framework.
> > what do you think?
> 
> I think it sounds like a plan. I also like incremental approach.
> 
> >
> > thanks,
> > rui
> >
> >> > +
> >> >  /*
> >> >   * Cooling algorithm for active trip points
> >> >   *
> >> > @@ -1086,19 +1115,24 @@ static void thermal_zone_trip_update(str
> >> >                             cur_state = cur_state > instance->lower ?
> >> >                                 (cur_state - 1) : instance->lower;
> >> >                     }
> >> > -                   cdev->ops->set_cur_state(cdev, cur_state);
> >> > +                   instance->target = cur_state;
> >> > +                   cdev->updated = 0; /* cooling device needs update */
> >> >             }
> >> >     } else {        /* below trip */
> >> >             list_for_each_entry(instance, &tz->instances, tz_node) {
> >> >                     if (instance->trip != trip)
> >> >                             continue;
> >> >
> >> > +                   /* Do not use the deacitve thermal instance */
> >> > +                   if (instance->target == -1)
> >> > +                           continue;
> >> >                     cdev = instance->cdev;
> >> >                     cdev->ops->get_cur_state(cdev, &cur_state);
> >> >
> >> >                     cur_state = cur_state > instance->lower ?
> >> > -                               (cur_state - 1) : instance->lower;
> >> > -                   cdev->ops->set_cur_state(cdev, cur_state);
> >> > +                               (cur_state - 1) : -1;
> >> > +                   instance->target = cur_state;
> >> > +                   cdev->updated = 0; /* cooling device needs update */
> >> >             }
> >> >     }
> >> >
> >> > @@ -1159,6 +1193,7 @@ void thermal_zone_device_update(struct t
> >> >             }
> >> >     }
> >> >
> >> > +   thermal_zone_do_update(tz);
> >> >     if (tz->forced_passive)
> >> >             thermal_zone_device_passive(tz, temp, tz->forced_passive,
> >> >                                         THERMAL_TRIPS_NONE);
> >> > Index: rtd3/include/linux/thermal.h
> >> > ===================================================================
> >> > --- rtd3.orig/include/linux/thermal.h
> >> > +++ rtd3/include/linux/thermal.h
> >> > @@ -86,6 +86,7 @@ struct thermal_cooling_device {
> >> >     struct device device;
> >> >     void *devdata;
> >> >     const struct thermal_cooling_device_ops *ops;
> >> > +   int updated; /* 1 if the cooling device does not need update */
> >> >     struct list_head instances;
> >> >     struct list_head node;
> >> >  };
> >> >
> >> >
> >
> >
> 
> 
>