[v2] ACPI: scan: Reduce overhead related to devices with dependencies

  From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Subject: [PATCH] ACPI: scan: Reduce overhead related to devices with dependencies

Notice that all of the objects for which the acpi_scan_check_dep()
return value is greater than 0 are present in acpi_dep_list as consumers
(there may be multiple entries for one object, but that is not a
problem), so after carrying out the initial ACPI namespace walk in which
devices with dependencies are skipped, acpi_bus_scan() can simply walk
acpi_dep_list and enumerate all of the unique consumer objects from
there and their descendants instead of walking the entire target branch
of the ACPI namespace and looking for device objects that have not been
enumerated yet in it.

Because walking acpi_dep_list is generally less overhead than walking
the entire ACPI namespace, use the observation above to reduce the
system initialization overhead related to ACPI, which is particularly
important on large systems.

Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
---

-> v2: Hold acpi_dep_list_lock around the acpi_fetch_acpi_dev() invocation in
       acpi_scan_postponed() (Hans).

---
 drivers/acpi/scan.c     |   81 ++++++++++++++++++++++++++++++++++++------------
 include/acpi/acpi_bus.h |    2 +
 2 files changed, 63 insertions(+), 20 deletions(-)
  

Hi Rafael,

On 5/16/23 12:25, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> Subject: [PATCH] ACPI: scan: Reduce overhead related to devices with dependencies
> 
> Notice that all of the objects for which the acpi_scan_check_dep()
> return value is greater than 0 are present in acpi_dep_list as consumers
> (there may be multiple entries for one object, but that is not a
> problem), so after carrying out the initial ACPI namespace walk in which
> devices with dependencies are skipped, acpi_bus_scan() can simply walk
> acpi_dep_list and enumerate all of the unique consumer objects from
> there and their descendants instead of walking the entire target branch
> of the ACPI namespace and looking for device objects that have not been
> enumerated yet in it.
> 
> Because walking acpi_dep_list is generally less overhead than walking
> the entire ACPI namespace, use the observation above to reduce the
> system initialization overhead related to ACPI, which is particularly
> important on large systems.
> 
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> ---
> 
> -> v2: Hold acpi_dep_list_lock around the acpi_fetch_acpi_dev() invocation in
>        acpi_scan_postponed() (Hans).

This looks good to me now:

Reviewed-by: Hans de Goede <hdegoede@redhat.com>

with the remark that this obviously is going to need a lot
of testing in case this somehow breaks some weird corner case.

I'll add it to my personal git kernel tree with all my
pending work right away, so that it will get tested on
various devices as I test other patches on those devices.

I'll let you know if I hit any problems which seem to be
caused by this.

Regards,

Hans





> 
> ---
>  drivers/acpi/scan.c     |   81 ++++++++++++++++++++++++++++++++++++------------
>  include/acpi/acpi_bus.h |    2 +
>  2 files changed, 63 insertions(+), 20 deletions(-)
> 
> Index: linux-pm/include/acpi/acpi_bus.h
> ===================================================================
> --- linux-pm.orig/include/acpi/acpi_bus.h
> +++ linux-pm/include/acpi/acpi_bus.h
> @@ -289,6 +289,8 @@ struct acpi_dep_data {
>  	acpi_handle supplier;
>  	acpi_handle consumer;
>  	bool honor_dep;
> +	bool met;
> +	bool free_when_met;
>  };
>  
>  /* Performance Management */
> Index: linux-pm/drivers/acpi/scan.c
> ===================================================================
> --- linux-pm.orig/drivers/acpi/scan.c
> +++ linux-pm/drivers/acpi/scan.c
> @@ -2029,8 +2029,6 @@ static u32 acpi_scan_check_dep(acpi_hand
>  	return count;
>  }
>  
> -static bool acpi_bus_scan_second_pass;
> -
>  static acpi_status acpi_bus_check_add(acpi_handle handle, bool check_dep,
>  				      struct acpi_device **adev_p)
>  {
> @@ -2050,10 +2048,8 @@ static acpi_status acpi_bus_check_add(ac
>  			return AE_OK;
>  
>  		/* Bail out if there are dependencies. */
> -		if (acpi_scan_check_dep(handle, check_dep) > 0) {
> -			acpi_bus_scan_second_pass = true;
> +		if (acpi_scan_check_dep(handle, check_dep) > 0)
>  			return AE_CTRL_DEPTH;
> -		}
>  
>  		fallthrough;
>  	case ACPI_TYPE_ANY:	/* for ACPI_ROOT_OBJECT */
> @@ -2301,6 +2297,12 @@ static bool acpi_scan_clear_dep_queue(st
>  	return true;
>  }
>  
> +static void acpi_scan_delete_dep_data(struct acpi_dep_data *dep)
> +{
> +	list_del(&dep->node);
> +	kfree(dep);
> +}
> +
>  static int acpi_scan_clear_dep(struct acpi_dep_data *dep, void *data)
>  {
>  	struct acpi_device *adev = acpi_get_acpi_dev(dep->consumer);
> @@ -2311,8 +2313,10 @@ static int acpi_scan_clear_dep(struct ac
>  			acpi_dev_put(adev);
>  	}
>  
> -	list_del(&dep->node);
> -	kfree(dep);
> +	if (dep->free_when_met)
> +		acpi_scan_delete_dep_data(dep);
> +	else
> +		dep->met = true;
>  
>  	return 0;
>  }
> @@ -2406,6 +2410,55 @@ struct acpi_device *acpi_dev_get_next_co
>  }
>  EXPORT_SYMBOL_GPL(acpi_dev_get_next_consumer_dev);
>  
> +static void acpi_scan_postponed_branch(acpi_handle handle)
> +{
> +	struct acpi_device *adev = NULL;
> +
> +	if (ACPI_FAILURE(acpi_bus_check_add(handle, false, &adev)))
> +		return;
> +
> +	acpi_walk_namespace(ACPI_TYPE_ANY, handle, ACPI_UINT32_MAX,
> +			    acpi_bus_check_add_2, NULL, NULL, (void **)&adev);
> +	acpi_bus_attach(adev, NULL);
> +}
> +
> +static void acpi_scan_postponed(void)
> +{
> +	struct acpi_dep_data *dep, *tmp;
> +
> +	mutex_lock(&acpi_dep_list_lock);
> +
> +	list_for_each_entry_safe(dep, tmp, &acpi_dep_list, node) {
> +		acpi_handle handle = dep->consumer;
> +
> +		/*
> +		 * In case there are multiple acpi_dep_list entries with the
> +		 * same consumer, skip the current entry if the consumer device
> +		 * object corresponding to it is present already.
> +		 */
> +		if (!acpi_fetch_acpi_dev(handle)) {
> +			/*
> +			 * Even though the lock is released here, tmp is
> +			 * guaranteed to be valid, because none of the list
> +			 * entries following dep is marked as "free when met"
> +			 * and so they cannot be deleted.
> +			 */
> +			mutex_unlock(&acpi_dep_list_lock);
> +
> +			acpi_scan_postponed_branch(handle);
> +
> +			mutex_lock(&acpi_dep_list_lock);
> +		}
> +
> +		if (dep->met)
> +			acpi_scan_delete_dep_data(dep);
> +		else
> +			dep->free_when_met = true;
> +	}
> +
> +	mutex_unlock(&acpi_dep_list_lock);
> +}
> +
>  /**
>   * acpi_bus_scan - Add ACPI device node objects in a given namespace scope.
>   * @handle: Root of the namespace scope to scan.
> @@ -2424,8 +2477,6 @@ int acpi_bus_scan(acpi_handle handle)
>  {
>  	struct acpi_device *device = NULL;
>  
> -	acpi_bus_scan_second_pass = false;
> -
>  	/* Pass 1: Avoid enumerating devices with missing dependencies. */
>  
>  	if (ACPI_SUCCESS(acpi_bus_check_add(handle, true, &device)))
> @@ -2438,19 +2489,9 @@ int acpi_bus_scan(acpi_handle handle)
>  
>  	acpi_bus_attach(device, (void *)true);
>  
> -	if (!acpi_bus_scan_second_pass)
> -		return 0;
> -
>  	/* Pass 2: Enumerate all of the remaining devices. */
>  
> -	device = NULL;
> -
> -	if (ACPI_SUCCESS(acpi_bus_check_add(handle, false, &device)))
> -		acpi_walk_namespace(ACPI_TYPE_ANY, handle, ACPI_UINT32_MAX,
> -				    acpi_bus_check_add_2, NULL, NULL,
> -				    (void **)&device);
> -
> -	acpi_bus_attach(device, NULL);
> +	acpi_scan_postponed();
>  
>  	return 0;
>  }
> 
> 
>
  

Hi Hans,

On Tue, May 16, 2023 at 12:33 PM Hans de Goede <hdegoede@redhat.com> wrote:
>
> Hi Rafael,
>
> On 5/16/23 12:25, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > Subject: [PATCH] ACPI: scan: Reduce overhead related to devices with dependencies
> >
> > Notice that all of the objects for which the acpi_scan_check_dep()
> > return value is greater than 0 are present in acpi_dep_list as consumers
> > (there may be multiple entries for one object, but that is not a
> > problem), so after carrying out the initial ACPI namespace walk in which
> > devices with dependencies are skipped, acpi_bus_scan() can simply walk
> > acpi_dep_list and enumerate all of the unique consumer objects from
> > there and their descendants instead of walking the entire target branch
> > of the ACPI namespace and looking for device objects that have not been
> > enumerated yet in it.
> >
> > Because walking acpi_dep_list is generally less overhead than walking
> > the entire ACPI namespace, use the observation above to reduce the
> > system initialization overhead related to ACPI, which is particularly
> > important on large systems.
> >
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> > ---
> >
> > -> v2: Hold acpi_dep_list_lock around the acpi_fetch_acpi_dev() invocation in
> >        acpi_scan_postponed() (Hans).
>
> This looks good to me now:
>
> Reviewed-by: Hans de Goede <hdegoede@redhat.com>
>
> with the remark that this obviously is going to need a lot
> of testing in case this somehow breaks some weird corner case.
>
> I'll add it to my personal git kernel tree with all my
> pending work right away, so that it will get tested on
> various devices as I test other patches on those devices.
>
> I'll let you know if I hit any problems which seem to be
> caused by this.

Awesome, thanks!


> >
> > ---
> >  drivers/acpi/scan.c     |   81 ++++++++++++++++++++++++++++++++++++------------
> >  include/acpi/acpi_bus.h |    2 +
> >  2 files changed, 63 insertions(+), 20 deletions(-)
> >
> > Index: linux-pm/include/acpi/acpi_bus.h
> > ===================================================================
> > --- linux-pm.orig/include/acpi/acpi_bus.h
> > +++ linux-pm/include/acpi/acpi_bus.h
> > @@ -289,6 +289,8 @@ struct acpi_dep_data {
> >       acpi_handle supplier;
> >       acpi_handle consumer;
> >       bool honor_dep;
> > +     bool met;
> > +     bool free_when_met;
> >  };
> >
> >  /* Performance Management */
> > Index: linux-pm/drivers/acpi/scan.c
> > ===================================================================
> > --- linux-pm.orig/drivers/acpi/scan.c
> > +++ linux-pm/drivers/acpi/scan.c
> > @@ -2029,8 +2029,6 @@ static u32 acpi_scan_check_dep(acpi_hand
> >       return count;
> >  }
> >
> > -static bool acpi_bus_scan_second_pass;
> > -
> >  static acpi_status acpi_bus_check_add(acpi_handle handle, bool check_dep,
> >                                     struct acpi_device **adev_p)
> >  {
> > @@ -2050,10 +2048,8 @@ static acpi_status acpi_bus_check_add(ac
> >                       return AE_OK;
> >
> >               /* Bail out if there are dependencies. */
> > -             if (acpi_scan_check_dep(handle, check_dep) > 0) {
> > -                     acpi_bus_scan_second_pass = true;
> > +             if (acpi_scan_check_dep(handle, check_dep) > 0)
> >                       return AE_CTRL_DEPTH;
> > -             }
> >
> >               fallthrough;
> >       case ACPI_TYPE_ANY:     /* for ACPI_ROOT_OBJECT */
> > @@ -2301,6 +2297,12 @@ static bool acpi_scan_clear_dep_queue(st
> >       return true;
> >  }
> >
> > +static void acpi_scan_delete_dep_data(struct acpi_dep_data *dep)
> > +{
> > +     list_del(&dep->node);
> > +     kfree(dep);
> > +}
> > +
> >  static int acpi_scan_clear_dep(struct acpi_dep_data *dep, void *data)
> >  {
> >       struct acpi_device *adev = acpi_get_acpi_dev(dep->consumer);
> > @@ -2311,8 +2313,10 @@ static int acpi_scan_clear_dep(struct ac
> >                       acpi_dev_put(adev);
> >       }
> >
> > -     list_del(&dep->node);
> > -     kfree(dep);
> > +     if (dep->free_when_met)
> > +             acpi_scan_delete_dep_data(dep);
> > +     else
> > +             dep->met = true;
> >
> >       return 0;
> >  }
> > @@ -2406,6 +2410,55 @@ struct acpi_device *acpi_dev_get_next_co
> >  }
> >  EXPORT_SYMBOL_GPL(acpi_dev_get_next_consumer_dev);
> >
> > +static void acpi_scan_postponed_branch(acpi_handle handle)
> > +{
> > +     struct acpi_device *adev = NULL;
> > +
> > +     if (ACPI_FAILURE(acpi_bus_check_add(handle, false, &adev)))
> > +             return;
> > +
> > +     acpi_walk_namespace(ACPI_TYPE_ANY, handle, ACPI_UINT32_MAX,
> > +                         acpi_bus_check_add_2, NULL, NULL, (void **)&adev);
> > +     acpi_bus_attach(adev, NULL);
> > +}
> > +
> > +static void acpi_scan_postponed(void)
> > +{
> > +     struct acpi_dep_data *dep, *tmp;
> > +
> > +     mutex_lock(&acpi_dep_list_lock);
> > +
> > +     list_for_each_entry_safe(dep, tmp, &acpi_dep_list, node) {
> > +             acpi_handle handle = dep->consumer;
> > +
> > +             /*
> > +              * In case there are multiple acpi_dep_list entries with the
> > +              * same consumer, skip the current entry if the consumer device
> > +              * object corresponding to it is present already.
> > +              */
> > +             if (!acpi_fetch_acpi_dev(handle)) {
> > +                     /*
> > +                      * Even though the lock is released here, tmp is
> > +                      * guaranteed to be valid, because none of the list
> > +                      * entries following dep is marked as "free when met"
> > +                      * and so they cannot be deleted.
> > +                      */
> > +                     mutex_unlock(&acpi_dep_list_lock);
> > +
> > +                     acpi_scan_postponed_branch(handle);
> > +
> > +                     mutex_lock(&acpi_dep_list_lock);
> > +             }
> > +
> > +             if (dep->met)
> > +                     acpi_scan_delete_dep_data(dep);
> > +             else
> > +                     dep->free_when_met = true;
> > +     }
> > +
> > +     mutex_unlock(&acpi_dep_list_lock);
> > +}
> > +
> >  /**
> >   * acpi_bus_scan - Add ACPI device node objects in a given namespace scope.
> >   * @handle: Root of the namespace scope to scan.
> > @@ -2424,8 +2477,6 @@ int acpi_bus_scan(acpi_handle handle)
> >  {
> >       struct acpi_device *device = NULL;
> >
> > -     acpi_bus_scan_second_pass = false;
> > -
> >       /* Pass 1: Avoid enumerating devices with missing dependencies. */
> >
> >       if (ACPI_SUCCESS(acpi_bus_check_add(handle, true, &device)))
> > @@ -2438,19 +2489,9 @@ int acpi_bus_scan(acpi_handle handle)
> >
> >       acpi_bus_attach(device, (void *)true);
> >
> > -     if (!acpi_bus_scan_second_pass)
> > -             return 0;
> > -
> >       /* Pass 2: Enumerate all of the remaining devices. */
> >
> > -     device = NULL;
> > -
> > -     if (ACPI_SUCCESS(acpi_bus_check_add(handle, false, &device)))
> > -             acpi_walk_namespace(ACPI_TYPE_ANY, handle, ACPI_UINT32_MAX,
> > -                                 acpi_bus_check_add_2, NULL, NULL,
> > -                                 (void **)&device);
> > -
> > -     acpi_bus_attach(device, NULL);
> > +     acpi_scan_postponed();
> >
> >       return 0;
> >  }
> >
> >
> >
>