From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>

Sadly, there is quite a bit of technical debt related to the
kernel's ACPI support subsystem and one of the most significant
pieces of it is the existence and use of ACPI drivers represented
by struct acpi_driver objects.

Those drivers are bound directly to struct acpi_device objects, also
referred to as "ACPI device nodes", representing device objects in the
ACPI namespace defined as:

 A hierarchical tree structure in OS-controlled memory that contains
 named objects. These objects may be data objects, control method
 objects, bus/device package objects, and so on.

according to the ACPI specification [1].

The above definition implies, although rather indirectly, that the
objects in question don't really represent hardware.  They are just
"device package objects" containing some information on the devices
present in the given platform that is known to the platform firmware.

Although the platform firmware can be the only source of information on
some devices, the information provided by it alone may be insufficient
for device enumeration in general.  If that is the case, binding a
driver directly to a given ACPI device node clearly doesn't make sense.
If the device in question is enumerated through a hardware interface, it
will be represented by a device object matching that interface, like
a struct pci_dev, and the ACPI device node corresponding to it will be
treated as its "ACPI companions" whose role is to amend the "native"
enumeratiom mechanism.

For the sake of consistency and confusion avoidance, it is better to
treat ACPI device nodes in general as ACPI companions of other device
objects representing hardware.  In some cases though it appeared easier
to take a shortcut and use an ACPI driver binding directly to an ACPI
device node.  Moreover, there were corner cases in which that was the
only choice, but they all have been addressed now.

In all cases in which an ACPI driver might be used, the ACPI device
node it might bind to is an ACPI companion of another device object
representing a piece of hardware.  It is thus better to use a driver
binding to the latter than to use an ACPI driver and leave the other
device object alone, not just because doing so is more consistent and
less confusing, but also because using ACPI drivers may lead to
potential functional deficiencies, like possible ordering issues
related to power management.

Unfortunately, there are quite a few ACPI drivers in use and, as a rule,
they bind to ACPI device nodes that are ACPI companions of platform
devices, so in fact they play the role of platform drivers although in
a kind of convoluted way.  An effort has been under way to replace them
with platform drivers, which is relatively straightforward in the vast
majority of cases, but it has not been pursued very aggressively so far,
mostly due to the existence of the corner cases mentioned above.
However, since those corner cases are gone now, it makes sense to spend
more time on driver conversions with the ultimate goal to get rid of
struct acpi_driver and the related code from the kernel.

To that end, add a document explaining why using ACPI drivers is not
a good idea, so it need not be explained from scratch on every attempt
to convert an ACPI driver to a platform one.

Link: https://uefi.org/specs/ACPI/6.6/02_Definition_of_Terms.html#term-ACPI-Namespace [1]
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Reviewed-by: Armin Wolf <W_Armin@gmx.de>
---

v1 -> v2:
   * Fixed a typo in the changelog (Andy).
   * Addressed two review comments from Randy ("is questionable either" ->
     "is also questionable" and "eg." -> "e.g.").
   * Added R-bys from Andy and Armin.

---

Although this patch can be applied independently, it actually depends on
some ACPI changes in linux-next and on

https://lore.kernel.org/linux-acpi/12824456.O9o76ZdvQC@rafael.j.wysocki/

so it is better to handle it along with that material.

---
 Documentation/driver-api/acpi/acpi-drivers.rst |   80 +++++++++++++++++++++++++
 Documentation/driver-api/acpi/index.rst        |    1 
 2 files changed, 81 insertions(+)

--- /dev/null
+++ b/Documentation/driver-api/acpi/acpi-drivers.rst
@@ -0,0 +1,80 @@
+.. SPDX-License-Identifier: GPL-2.0
+.. include:: <isonum.txt>
+
+=========================================
+Why using ACPI drivers is not a good idea
+=========================================
+
+:Copyright: |copy| 2026, Intel Corporation
+
+:Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
+
+Even though binding drivers directly to struct acpi_device objects, also
+referred to as "ACPI device nodes", allows basic functionality to be provided
+at least in some cases, there are problems with it, related to general
+consistency, sysfs layout, power management operation ordering, and code
+cleanliness.
+
+First of all, ACPI device nodes represent firmware entities rather than
+hardware and in many cases they provide auxiliary information on devices
+enumerated independently (like PCI devices or CPUs).  It is therefore generally
+questionable to assign resources to them because the entities represented by
+them do not decode addresses in the memory or I/O address spaces and do not
+generate interrupts or similar (all of that is done by hardware).
+
+Second, as a general rule, a struct acpi_device can only be a parent of another
+struct acpi_device.  If that is not the case, the location of the child device
+in the device hierarchy is at least confusing and it may not be straightforward
+to identify the piece of hardware providing functionality represented by it.
+However, binding a driver directly to an ACPI device node may cause that to
+happen if the given driver registers input devices or wakeup sources under it,
+for example.
+
+Next, using system suspend and resume callbacks directly on ACPI device nodes
+is also questionable because it may cause ordering problems to appear.  Namely,
+ACPI device nodes are registered before enumerating hardware corresponding to
+them and they land on the PM list in front of the majority of other device
+objects.  Consequently, the execution ordering of their PM callbacks may be
+different from what is generally expected.  Also, in general, dependencies
+returned by _DEP objects do not affect ACPI device nodes themselves, but the
+"physical" devices associated with them, which potentially is one more source
+of inconsistency related to treating ACPI device nodes as "real" device
+representation.
+
+All of the above means that binding drivers to ACPI device nodes should
+generally be avoided and so struct acpi_driver objects should not be used.
+
+Moreover, a device ID is necessary to bind a driver directly to an ACPI device
+node, but device IDs are not generally associated with all of them.  Some of
+them contain alternative information allowing the corresponding pieces of
+hardware to be identified, for example represeted by an _ADR object return
+value, and device IDs are not used in those cases.  In consequence, confusingly
+enough, binding an ACPI driver to an ACPI device node may even be impossible.
+
+When that happens, the piece of hardware corresponding to the given ACPI device
+node is represented by another device object, like a struct pci_dev, and the
+ACPI device node is the "ACPI companion" of that device, accessible through its
+fwnode pointer used by the ACPI_COMPANION() macro.  The ACPI companion holds
+additional information on the device configuration and possibly some "recipes"
+on device manipulation in the form of AML (ACPI Machine Language) bytecode
+provided by the platform firmware.  Thus the role of the ACPI device node is
+similar to the role of a struct device_node on a system where Device Tree is
+used for platform description.
+
+For consistency, this approach has been extended to the cases in which ACPI
+device IDs are used.  Namely, in those cases, an additional device object is
+created to represent the piece of hardware corresponding to a given ACPI device
+node.  By default, it is a platform device, but it may also be a PNP device, a
+CPU device, or another type of device, depending on what the given piece of
+hardware actually is.  There are even cases in which multiple devices are
+"backed" or "accompanied" by one ACPI device node (e.g. ACPI device nodes
+corresponding to GPUs that may provide firmware interfaces for backlight
+brightness control in addition to GPU configuration information).
+
+This means that it really should never be necessary to bind a driver directly to
+an ACPI device node because there is a "proper" device object representing the
+corresponding piece of hardware that can be bound to by a "proper" driver using
+the given ACPI device node as the device's ACPI companion.  Thus, in principle,
+there is no reason to use ACPI drivers and if they all were replaced with other
+driver types (for example, platform drivers), some code could be dropped and
+some complexity would go away.
--- a/Documentation/driver-api/acpi/index.rst
+++ b/Documentation/driver-api/acpi/index.rst
@@ -7,3 +7,4 @@ ACPI Support
 
    linuxized-acpica
    scan_handlers
+   acpi-drivers

On 1/6/26 4:27 AM, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
> Reviewed-by: Armin Wolf <W_Armin@gmx.de>

Reviewed-by: Randy Dunlap <rdunlap@infradead.org>

Thanks.

> ---
> 
> v1 -> v2:
>    * Fixed a typo in the changelog (Andy).
>    * Addressed two review comments from Randy ("is questionable either" ->
>      "is also questionable" and "eg." -> "e.g.").
>    * Added R-bys from Andy and Armin.
> 
> ---
> 
> Although this patch can be applied independently, it actually depends on
> some ACPI changes in linux-next and on
> 
> https://lore.kernel.org/linux-acpi/12824456.O9o76ZdvQC@rafael.j.wysocki/
> 
> so it is better to handle it along with that material.
> 
> ---
>  Documentation/driver-api/acpi/acpi-drivers.rst |   80 +++++++++++++++++++++++++
>  Documentation/driver-api/acpi/index.rst        |    1 
>  2 files changed, 81 insertions(+)

-- 
~Randy

On 1/6/2026 6:27 AM, Rafael J. Wysocki wrote:
> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> 
> Sadly, there is quite a bit of technical debt related to the
> kernel's ACPI support subsystem and one of the most significant
> pieces of it is the existence and use of ACPI drivers represented
> by struct acpi_driver objects.
> 
> Those drivers are bound directly to struct acpi_device objects, also
> referred to as "ACPI device nodes", representing device objects in the
> ACPI namespace defined as:
> 
>   A hierarchical tree structure in OS-controlled memory that contains
>   named objects. These objects may be data objects, control method
>   objects, bus/device package objects, and so on.
> 
> according to the ACPI specification [1].
> 
> The above definition implies, although rather indirectly, that the
> objects in question don't really represent hardware.  They are just
> "device package objects" containing some information on the devices
> present in the given platform that is known to the platform firmware.
> 
> Although the platform firmware can be the only source of information on
> some devices, the information provided by it alone may be insufficient
> for device enumeration in general.  If that is the case, binding a
> driver directly to a given ACPI device node clearly doesn't make sense.
> If the device in question is enumerated through a hardware interface, it
> will be represented by a device object matching that interface, like
> a struct pci_dev, and the ACPI device node corresponding to it will be
> treated as its "ACPI companions" whose role is to amend the "native"
> enumeratiom mechanism.
> 
> For the sake of consistency and confusion avoidance, it is better to
> treat ACPI device nodes in general as ACPI companions of other device
> objects representing hardware.  In some cases though it appeared easier
> to take a shortcut and use an ACPI driver binding directly to an ACPI
> device node.  Moreover, there were corner cases in which that was the
> only choice, but they all have been addressed now.
> 
> In all cases in which an ACPI driver might be used, the ACPI device
> node it might bind to is an ACPI companion of another device object
> representing a piece of hardware.  It is thus better to use a driver
> binding to the latter than to use an ACPI driver and leave the other
> device object alone, not just because doing so is more consistent and
> less confusing, but also because using ACPI drivers may lead to
> potential functional deficiencies, like possible ordering issues
> related to power management.
> 
> Unfortunately, there are quite a few ACPI drivers in use and, as a rule,
> they bind to ACPI device nodes that are ACPI companions of platform
> devices, so in fact they play the role of platform drivers although in
> a kind of convoluted way.  An effort has been under way to replace them
> with platform drivers, which is relatively straightforward in the vast
> majority of cases, but it has not been pursued very aggressively so far,
> mostly due to the existence of the corner cases mentioned above.

This is the same as Danilo's comment; but could you leave a few examples 
of conversions that have been done successfully?  Commit hashes that can 
demonstrate what it actually takes to convert an acpi driver to a 
platform driver and might make it easier for people to reference when 
this comes up.

> However, since those corner cases are gone now, it makes sense to spend
> more time on driver conversions with the ultimate goal to get rid of
> struct acpi_driver and the related code from the kernel.
> 
> To that end, add a document explaining why using ACPI drivers is not
> a good idea, so it need not be explained from scratch on every attempt
> to convert an ACPI driver to a platform one.
> 
> Link: https://uefi.org/specs/ACPI/6.6/02_Definition_of_Terms.html#term-ACPI-Namespace [1]
> Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
> Reviewed-by: Armin Wolf <W_Armin@gmx.de>
> ---
> 
> v1 -> v2:
>     * Fixed a typo in the changelog (Andy).
>     * Addressed two review comments from Randy ("is questionable either" ->
>       "is also questionable" and "eg." -> "e.g.").
>     * Added R-bys from Andy and Armin.
> 
> ---
> 
> Although this patch can be applied independently, it actually depends on
> some ACPI changes in linux-next and on
> 
> https://lore.kernel.org/linux-acpi/12824456.O9o76ZdvQC@rafael.j.wysocki/
> 
> so it is better to handle it along with that material.
> 
> ---
>   Documentation/driver-api/acpi/acpi-drivers.rst |   80 +++++++++++++++++++++++++
>   Documentation/driver-api/acpi/index.rst        |    1
>   2 files changed, 81 insertions(+)
> 
> --- /dev/null
> +++ b/Documentation/driver-api/acpi/acpi-drivers.rst
> @@ -0,0 +1,80 @@
> +.. SPDX-License-Identifier: GPL-2.0
> +.. include:: <isonum.txt>
> +
> +=========================================
> +Why using ACPI drivers is not a good idea
> +=========================================
> +
> +:Copyright: |copy| 2026, Intel Corporation
> +
> +:Author: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> +
> +Even though binding drivers directly to struct acpi_device objects, also
> +referred to as "ACPI device nodes", allows basic functionality to be provided
> +at least in some cases, there are problems with it, related to general
> +consistency, sysfs layout, power management operation ordering, and code
> +cleanliness.
> +
> +First of all, ACPI device nodes represent firmware entities rather than
> +hardware and in many cases they provide auxiliary information on devices
> +enumerated independently (like PCI devices or CPUs).  It is therefore generally
> +questionable to assign resources to them because the entities represented by
> +them do not decode addresses in the memory or I/O address spaces and do not
> +generate interrupts or similar (all of that is done by hardware).
> +
> +Second, as a general rule, a struct acpi_device can only be a parent of another
> +struct acpi_device.  If that is not the case, the location of the child device
> +in the device hierarchy is at least confusing and it may not be straightforward
> +to identify the piece of hardware providing functionality represented by it.
> +However, binding a driver directly to an ACPI device node may cause that to
> +happen if the given driver registers input devices or wakeup sources under it,
> +for example.
> +
> +Next, using system suspend and resume callbacks directly on ACPI device nodes
> +is also questionable because it may cause ordering problems to appear.  Namely,
> +ACPI device nodes are registered before enumerating hardware corresponding to
> +them and they land on the PM list in front of the majority of other device
> +objects.  Consequently, the execution ordering of their PM callbacks may be
> +different from what is generally expected.  Also, in general, dependencies
> +returned by _DEP objects do not affect ACPI device nodes themselves, but the
> +"physical" devices associated with them, which potentially is one more source
> +of inconsistency related to treating ACPI device nodes as "real" device
> +representation.
> +
> +All of the above means that binding drivers to ACPI device nodes should
> +generally be avoided and so struct acpi_driver objects should not be used.
> +
> +Moreover, a device ID is necessary to bind a driver directly to an ACPI device
> +node, but device IDs are not generally associated with all of them.  Some of
> +them contain alternative information allowing the corresponding pieces of
> +hardware to be identified, for example represeted by an _ADR object return
> +value, and device IDs are not used in those cases.  In consequence, confusingly
> +enough, binding an ACPI driver to an ACPI device node may even be impossible.
> +
> +When that happens, the piece of hardware corresponding to the given ACPI device
> +node is represented by another device object, like a struct pci_dev, and the
> +ACPI device node is the "ACPI companion" of that device, accessible through its
> +fwnode pointer used by the ACPI_COMPANION() macro.  The ACPI companion holds
> +additional information on the device configuration and possibly some "recipes"
> +on device manipulation in the form of AML (ACPI Machine Language) bytecode
> +provided by the platform firmware.  Thus the role of the ACPI device node is
> +similar to the role of a struct device_node on a system where Device Tree is
> +used for platform description.
> +
> +For consistency, this approach has been extended to the cases in which ACPI
> +device IDs are used.  Namely, in those cases, an additional device object is
> +created to represent the piece of hardware corresponding to a given ACPI device
> +node.  By default, it is a platform device, but it may also be a PNP device, a
> +CPU device, or another type of device, depending on what the given piece of
> +hardware actually is.  There are even cases in which multiple devices are
> +"backed" or "accompanied" by one ACPI device node (e.g. ACPI device nodes
> +corresponding to GPUs that may provide firmware interfaces for backlight
> +brightness control in addition to GPU configuration information).
> +
> +This means that it really should never be necessary to bind a driver directly to
> +an ACPI device node because there is a "proper" device object representing the
> +corresponding piece of hardware that can be bound to by a "proper" driver using
> +the given ACPI device node as the device's ACPI companion.  Thus, in principle,
> +there is no reason to use ACPI drivers and if they all were replaced with other
> +driver types (for example, platform drivers), some code could be dropped and
> +some complexity would go away.
> --- a/Documentation/driver-api/acpi/index.rst
> +++ b/Documentation/driver-api/acpi/index.rst
> @@ -7,3 +7,4 @@ ACPI Support
>   
>      linuxized-acpica
>      scan_handlers
> +   acpi-drivers
> 
> 
>

On Tue, Jan 6, 2026 at 4:47 PM Mario Limonciello (AMD) (kernel.org)
<superm1@kernel.org> wrote:
>
>
>
> On 1/6/2026 6:27 AM, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> >
> > Sadly, there is quite a bit of technical debt related to the
> > kernel's ACPI support subsystem and one of the most significant
> > pieces of it is the existence and use of ACPI drivers represented
> > by struct acpi_driver objects.
> >
> > Those drivers are bound directly to struct acpi_device objects, also
> > referred to as "ACPI device nodes", representing device objects in the
> > ACPI namespace defined as:
> >
> >   A hierarchical tree structure in OS-controlled memory that contains
> >   named objects. These objects may be data objects, control method
> >   objects, bus/device package objects, and so on.
> >
> > according to the ACPI specification [1].
> >
> > The above definition implies, although rather indirectly, that the
> > objects in question don't really represent hardware.  They are just
> > "device package objects" containing some information on the devices
> > present in the given platform that is known to the platform firmware.
> >
> > Although the platform firmware can be the only source of information on
> > some devices, the information provided by it alone may be insufficient
> > for device enumeration in general.  If that is the case, binding a
> > driver directly to a given ACPI device node clearly doesn't make sense.
> > If the device in question is enumerated through a hardware interface, it
> > will be represented by a device object matching that interface, like
> > a struct pci_dev, and the ACPI device node corresponding to it will be
> > treated as its "ACPI companions" whose role is to amend the "native"
> > enumeratiom mechanism.
> >
> > For the sake of consistency and confusion avoidance, it is better to
> > treat ACPI device nodes in general as ACPI companions of other device
> > objects representing hardware.  In some cases though it appeared easier
> > to take a shortcut and use an ACPI driver binding directly to an ACPI
> > device node.  Moreover, there were corner cases in which that was the
> > only choice, but they all have been addressed now.
> >
> > In all cases in which an ACPI driver might be used, the ACPI device
> > node it might bind to is an ACPI companion of another device object
> > representing a piece of hardware.  It is thus better to use a driver
> > binding to the latter than to use an ACPI driver and leave the other
> > device object alone, not just because doing so is more consistent and
> > less confusing, but also because using ACPI drivers may lead to
> > potential functional deficiencies, like possible ordering issues
> > related to power management.
> >
> > Unfortunately, there are quite a few ACPI drivers in use and, as a rule,
> > they bind to ACPI device nodes that are ACPI companions of platform
> > devices, so in fact they play the role of platform drivers although in
> > a kind of convoluted way.  An effort has been under way to replace them
> > with platform drivers, which is relatively straightforward in the vast
> > majority of cases, but it has not been pursued very aggressively so far,
> > mostly due to the existence of the corner cases mentioned above.
>
> This is the same as Danilo's comment; but could you leave a few examples
> of conversions that have been done successfully?  Commit hashes that can
> demonstrate what it actually takes to convert an acpi driver to a
> platform driver and might make it easier for people to reference when
> this comes up.

The purpose of this posting and the new document is to grow awareness
rather than to tell people how to convert drivers.

I'll start posting driver conversion patches at one point and the
motivation for all of them is basically the same, so I thought it
would be better to document it in on place and then refer to it
instead of repeating the same information every time a conversion
patch is posted.

On 1/7/2026 6:22 AM, Rafael J. Wysocki wrote:
> On Tue, Jan 6, 2026 at 4:47 PM Mario Limonciello (AMD) (kernel.org)
> <superm1@kernel.org> wrote:
>>
>>
>>
>> On 1/6/2026 6:27 AM, Rafael J. Wysocki wrote:
>>> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>>>
>>> Sadly, there is quite a bit of technical debt related to the
>>> kernel's ACPI support subsystem and one of the most significant
>>> pieces of it is the existence and use of ACPI drivers represented
>>> by struct acpi_driver objects.
>>>
>>> Those drivers are bound directly to struct acpi_device objects, also
>>> referred to as "ACPI device nodes", representing device objects in the
>>> ACPI namespace defined as:
>>>
>>>    A hierarchical tree structure in OS-controlled memory that contains
>>>    named objects. These objects may be data objects, control method
>>>    objects, bus/device package objects, and so on.
>>>
>>> according to the ACPI specification [1].
>>>
>>> The above definition implies, although rather indirectly, that the
>>> objects in question don't really represent hardware.  They are just
>>> "device package objects" containing some information on the devices
>>> present in the given platform that is known to the platform firmware.
>>>
>>> Although the platform firmware can be the only source of information on
>>> some devices, the information provided by it alone may be insufficient
>>> for device enumeration in general.  If that is the case, binding a
>>> driver directly to a given ACPI device node clearly doesn't make sense.
>>> If the device in question is enumerated through a hardware interface, it
>>> will be represented by a device object matching that interface, like
>>> a struct pci_dev, and the ACPI device node corresponding to it will be
>>> treated as its "ACPI companions" whose role is to amend the "native"
>>> enumeratiom mechanism.
>>>
>>> For the sake of consistency and confusion avoidance, it is better to
>>> treat ACPI device nodes in general as ACPI companions of other device
>>> objects representing hardware.  In some cases though it appeared easier
>>> to take a shortcut and use an ACPI driver binding directly to an ACPI
>>> device node.  Moreover, there were corner cases in which that was the
>>> only choice, but they all have been addressed now.
>>>
>>> In all cases in which an ACPI driver might be used, the ACPI device
>>> node it might bind to is an ACPI companion of another device object
>>> representing a piece of hardware.  It is thus better to use a driver
>>> binding to the latter than to use an ACPI driver and leave the other
>>> device object alone, not just because doing so is more consistent and
>>> less confusing, but also because using ACPI drivers may lead to
>>> potential functional deficiencies, like possible ordering issues
>>> related to power management.
>>>
>>> Unfortunately, there are quite a few ACPI drivers in use and, as a rule,
>>> they bind to ACPI device nodes that are ACPI companions of platform
>>> devices, so in fact they play the role of platform drivers although in
>>> a kind of convoluted way.  An effort has been under way to replace them
>>> with platform drivers, which is relatively straightforward in the vast
>>> majority of cases, but it has not been pursued very aggressively so far,
>>> mostly due to the existence of the corner cases mentioned above.
>>
>> This is the same as Danilo's comment; but could you leave a few examples
>> of conversions that have been done successfully?  Commit hashes that can
>> demonstrate what it actually takes to convert an acpi driver to a
>> platform driver and might make it easier for people to reference when
>> this comes up.
> 
> The purpose of this posting and the new document is to grow awareness
> rather than to tell people how to convert drivers.
> 
> I'll start posting driver conversion patches at one point and the
> motivation for all of them is basically the same, so I thought it
> would be better to document it in on place and then refer to it
> instead of repeating the same information every time a conversion
> patch is posted.

Thanks for explaining.  Makes sense to me.  If you didn't already commit 
feel free to add my tag (no need to spin it if you already did).

Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>

On Tue, Jan 13, 2026 at 10:58 PM Mario Limonciello (AMD) (kernel.org)
<superm1@kernel.org> wrote:
>
>
>
> On 1/7/2026 6:22 AM, Rafael J. Wysocki wrote:
> > On Tue, Jan 6, 2026 at 4:47 PM Mario Limonciello (AMD) (kernel.org)
> > <superm1@kernel.org> wrote:
> >>
> >>
> >>
> >> On 1/6/2026 6:27 AM, Rafael J. Wysocki wrote:
> >>> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
> >>>
> >>> Sadly, there is quite a bit of technical debt related to the
> >>> kernel's ACPI support subsystem and one of the most significant
> >>> pieces of it is the existence and use of ACPI drivers represented
> >>> by struct acpi_driver objects.
> >>>
> >>> Those drivers are bound directly to struct acpi_device objects, also
> >>> referred to as "ACPI device nodes", representing device objects in the
> >>> ACPI namespace defined as:
> >>>
> >>>    A hierarchical tree structure in OS-controlled memory that contains
> >>>    named objects. These objects may be data objects, control method
> >>>    objects, bus/device package objects, and so on.
> >>>
> >>> according to the ACPI specification [1].
> >>>
> >>> The above definition implies, although rather indirectly, that the
> >>> objects in question don't really represent hardware.  They are just
> >>> "device package objects" containing some information on the devices
> >>> present in the given platform that is known to the platform firmware.
> >>>
> >>> Although the platform firmware can be the only source of information on
> >>> some devices, the information provided by it alone may be insufficient
> >>> for device enumeration in general.  If that is the case, binding a
> >>> driver directly to a given ACPI device node clearly doesn't make sense.
> >>> If the device in question is enumerated through a hardware interface, it
> >>> will be represented by a device object matching that interface, like
> >>> a struct pci_dev, and the ACPI device node corresponding to it will be
> >>> treated as its "ACPI companions" whose role is to amend the "native"
> >>> enumeratiom mechanism.
> >>>
> >>> For the sake of consistency and confusion avoidance, it is better to
> >>> treat ACPI device nodes in general as ACPI companions of other device
> >>> objects representing hardware.  In some cases though it appeared easier
> >>> to take a shortcut and use an ACPI driver binding directly to an ACPI
> >>> device node.  Moreover, there were corner cases in which that was the
> >>> only choice, but they all have been addressed now.
> >>>
> >>> In all cases in which an ACPI driver might be used, the ACPI device
> >>> node it might bind to is an ACPI companion of another device object
> >>> representing a piece of hardware.  It is thus better to use a driver
> >>> binding to the latter than to use an ACPI driver and leave the other
> >>> device object alone, not just because doing so is more consistent and
> >>> less confusing, but also because using ACPI drivers may lead to
> >>> potential functional deficiencies, like possible ordering issues
> >>> related to power management.
> >>>
> >>> Unfortunately, there are quite a few ACPI drivers in use and, as a rule,
> >>> they bind to ACPI device nodes that are ACPI companions of platform
> >>> devices, so in fact they play the role of platform drivers although in
> >>> a kind of convoluted way.  An effort has been under way to replace them
> >>> with platform drivers, which is relatively straightforward in the vast
> >>> majority of cases, but it has not been pursued very aggressively so far,
> >>> mostly due to the existence of the corner cases mentioned above.
> >>
> >> This is the same as Danilo's comment; but could you leave a few examples
> >> of conversions that have been done successfully?  Commit hashes that can
> >> demonstrate what it actually takes to convert an acpi driver to a
> >> platform driver and might make it easier for people to reference when
> >> this comes up.
> >
> > The purpose of this posting and the new document is to grow awareness
> > rather than to tell people how to convert drivers.
> >
> > I'll start posting driver conversion patches at one point and the
> > motivation for all of them is basically the same, so I thought it
> > would be better to document it in on place and then refer to it
> > instead of repeating the same information every time a conversion
> > patch is posted.
>
> Thanks for explaining.  Makes sense to me.  If you didn't already commit
> feel free to add my tag (no need to spin it if you already did).
>
> Reviewed-by: Mario Limonciello (AMD) <superm1@kernel.org>

I did, but adding a tag to it is not a big deal, thanks!

On Tue Jan 6, 2026 at 1:27 PM CET, Rafael J. Wysocki wrote:
> +This means that it really should never be necessary to bind a driver directly to
> +an ACPI device node because there is a "proper" device object representing the
> +corresponding piece of hardware that can be bound to by a "proper" driver using
> +the given ACPI device node as the device's ACPI companion.  Thus, in principle,
> +there is no reason to use ACPI drivers and if they all were replaced with other
> +driver types (for example, platform drivers), some code could be dropped and
> +some complexity would go away.

I think it would be good to explicitly encourage people to convert existing
drivers (maybe even list some of those) and rephrase the last sentence to list
what exact infrastructure, complexity, etc. can go away once that happened.

I think this would make it more likely to receive some contributions towards
this goal.

On Tue, Jan 6, 2026 at 3:01 PM Danilo Krummrich <dakr@kernel.org> wrote:
>
> On Tue Jan 6, 2026 at 1:27 PM CET, Rafael J. Wysocki wrote:
> > +This means that it really should never be necessary to bind a driver directly to
> > +an ACPI device node because there is a "proper" device object representing the
> > +corresponding piece of hardware that can be bound to by a "proper" driver using
> > +the given ACPI device node as the device's ACPI companion.  Thus, in principle,
> > +there is no reason to use ACPI drivers and if they all were replaced with other
> > +driver types (for example, platform drivers), some code could be dropped and
> > +some complexity would go away.
>
> I think it would be good to explicitly encourage people to convert existing
> drivers (maybe even list some of those) and rephrase the last sentence to list
> what exact infrastructure, complexity, etc. can go away once that happened.

I can rephrase the last sentence, but the purpose of this document is
to explain the motivation for the change rather than to make a call to
action.

> I think this would make it more likely to receive some contributions towards
> this goal.

I have prototype driver conversion patches for almost 50% of the cases
right now and I'm expecting to have them for all of the cases by the
end of the current development cycle, so I'm not sure how much there
is to gain.

I want people to not be surprised when they see those patches though.

On Wed, Jan 7, 2026 at 1:14 PM Rafael J. Wysocki <rafael@kernel.org> wrote:
>
> On Tue, Jan 6, 2026 at 3:01 PM Danilo Krummrich <dakr@kernel.org> wrote:
> >
> > On Tue Jan 6, 2026 at 1:27 PM CET, Rafael J. Wysocki wrote:
> > > +This means that it really should never be necessary to bind a driver directly to
> > > +an ACPI device node because there is a "proper" device object representing the
> > > +corresponding piece of hardware that can be bound to by a "proper" driver using
> > > +the given ACPI device node as the device's ACPI companion.  Thus, in principle,
> > > +there is no reason to use ACPI drivers and if they all were replaced with other
> > > +driver types (for example, platform drivers), some code could be dropped and
> > > +some complexity would go away.
> >
> > I think it would be good to explicitly encourage people to convert existing
> > drivers (maybe even list some of those) and rephrase the last sentence to list
> > what exact infrastructure, complexity, etc. can go away once that happened.
>
> I can rephrase the last sentence, but the purpose of this document is
> to explain the motivation for the change rather than to make a call to
> action.
>
> > I think this would make it more likely to receive some contributions towards
> > this goal.
>
> I have prototype driver conversion patches for almost 50% of the cases
> right now and I'm expecting to have them for all of the cases by the
> end of the current development cycle, so I'm not sure how much there
> is to gain.
>
> I want people to not be surprised when they see those patches though.

Actually, all of this work is almost ready to be submitted, but in the
meantime I put the patches on an experimental branch here:

https://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git/log/?h=experimental/acpi-driver-conversion

It includes some material already queued up for 6.20/7.0 and a fix
that has been posted.

The changelogs are a bit rudimentary and need some more work, but the
code changes should be complete (famous last words).

On Wed Jan 7, 2026 at 1:14 PM CET, Rafael J. Wysocki wrote:
> On Tue, Jan 6, 2026 at 3:01 PM Danilo Krummrich <dakr@kernel.org> wrote:
>>
>> On Tue Jan 6, 2026 at 1:27 PM CET, Rafael J. Wysocki wrote:
>> > +This means that it really should never be necessary to bind a driver directly to
>> > +an ACPI device node because there is a "proper" device object representing the
>> > +corresponding piece of hardware that can be bound to by a "proper" driver using
>> > +the given ACPI device node as the device's ACPI companion.  Thus, in principle,
>> > +there is no reason to use ACPI drivers and if they all were replaced with other
>> > +driver types (for example, platform drivers), some code could be dropped and
>> > +some complexity would go away.
>>
>> I think it would be good to explicitly encourage people to convert existing
>> drivers (maybe even list some of those) and rephrase the last sentence to list
>> what exact infrastructure, complexity, etc. can go away once that happened.
>
> I can rephrase the last sentence, but the purpose of this document is
> to explain the motivation for the change rather than to make a call to
> action.
>
>> I think this would make it more likely to receive some contributions towards
>> this goal.
>
> I have prototype driver conversion patches for almost 50% of the cases
> right now and I'm expecting to have them for all of the cases by the
> end of the current development cycle, so I'm not sure how much there
> is to gain.
>
> I want people to not be surprised when they see those patches though.


Ok, sounds good then. With or without rephrasing the last sentence,

Reviewed-by: Danilo Krummrich <dakr@kernel.org>