From bc18461816bd3a6a141e472f68c886bb932a6c2e Mon Sep 17 00:00:00 2001 From: Len Brown Date: Wed, 30 Nov 2016 23:00:30 -0500 Subject: ACPI: Document _OSI and _REV for Linux BIOS writers Based on a recent session at the Linux Plumber's Conference, we need to be more clear about how a BIOS should use _OSI to properly support Linux. Signed-off-by: Len Brown Reviewed-by: Lukas Wunner Signed-off-by: Rafael J. Wysocki --- Documentation/acpi/osi.txt | 187 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 187 insertions(+) create mode 100644 Documentation/acpi/osi.txt (limited to 'Documentation/acpi') diff --git a/Documentation/acpi/osi.txt b/Documentation/acpi/osi.txt new file mode 100644 index 000000000000..50cde0ceb9b0 --- /dev/null +++ b/Documentation/acpi/osi.txt @@ -0,0 +1,187 @@ +ACPI _OSI and _REV methods +-------------------------- + +An ACPI BIOS can use the "Operating System Interfaces" method (_OSI) +to find out what the operating system supports. Eg. If BIOS +AML code includes _OSI("XYZ"), the kernel's AML interpreter +can evaluate that method, look to see if it supports 'XYZ' +and answer YES or NO to the BIOS. + +The ACPI _REV method returns the "Revision of the ACPI specification +that OSPM supports" + +This document explains how and why the BIOS and Linux should use these methods. +It also explains how and why they are widely misused. + +How to use _OSI +--------------- + +Linux runs on two groups of machines -- those that are tested by the OEM +to be compatible with Linux, and those that were never tested with Linux, +but where Linux was installed to replace the original OS (Windows or OSX). + +The larger group is the systems tested to run only Windows. Not only that, +but many were tested to run with just one specific version of Windows. +So even though the BIOS may use _OSI to query what version of Windows is running, +only a single path through the BIOS has actually been tested. +Experience shows that taking untested paths through the BIOS +exposes Linux to an entire category of BIOS bugs. +For this reason, Linux _OSI defaults must continue to claim compatibility +with all versions of Windows. + +But Linux isn't actually compatible with Windows, and the Linux community +has also been hurt with regressions when Linux adds the latest version of +Windows to its list of _OSI strings. So it is possible that additional strings +will be more thoroughly vetted before shipping upstream in the future. +But it is likely that they will all eventually be added. + +What should an OEM do if they want to support Linux and Windows +using the same BIOS image? Often they need to do something different +for Linux to deal with how Linux is different from Windows. +Here the BIOS should ask exactly what it wants to know: + +_OSI("Linux-OEM-my_interface_name") +where 'OEM' is needed if this is an OEM-specific hook, +and 'my_interface_name' describes the hook, which could be a +quirk, a bug, or a bug-fix. + +In addition, the OEM should send a patch to upstream Linux +via the linux-acpi@vger.kernel.org mailing list. When that patch +is checked into Linux, the OS will answer "YES" when the BIOS +on the OEM's system uses _OSI to ask if the interface is supported +by the OS. Linux distributors can back-port that patch for Linux +pre-installs, and it will be included by all distributions that +re-base to upstream. If the distribution can not update the kernel binary, +they can also add an acpi_osi=Linux-OEM-my_interface_name +cmdline parameter to the boot loader, as needed. + +If the string refers to a feature where the upstream kernel +eventually grows support, a patch should be sent to remove +the string when that support is added to the kernel. + +That was easy. Read on, to find out how to do it wrong. + +Before _OSI, there was _OS +-------------------------- + +ACPI 1.0 specified "_OS" as an +"object that evaluates to a string that identifies the operating system." + +The ACPI BIOS flow would include an evaluation of _OS, and the AML +interpreter in the kernel would return to it a string identifying the OS: + +Windows 98, SE: "Microsoft Windows" +Windows ME: "Microsoft WindowsME:Millenium Edition" +Windows NT: "Microsoft Windows NT" + +The idea was on a platform tasked with running multiple OS's, +the BIOS could use _OS to enable devices that an OS +might support, or enable quirks or bug workarounds +necessary to make the platform compatible with that pre-existing OS. + +But _OS had fundamental problems. First, the BIOS needed to know the name +of every possible version of the OS that would run on it, and needed to know +all the quirks of those OS's. Certainly it would make more sense +for the BIOS to ask *specific* things of the OS, such +"do you support a specific interface", and thus in ACPI 3.0, +_OSI was born to replace _OS. + +_OS was abandoned, though even today, many BIOS look for +_OS "Microsoft Windows NT", though it seems somewhat far-fetched +that anybody would install those old operating systems +over what came with the machine. + +Linux answers "Microsoft Windows NT" to please that BIOS idiom. +That is the *only* viable strategy, as that is what modern Windows does, +and so doing otherwise could steer the BIOS down an untested path. + +_OSI is born, and immediately misused +-------------------------------------- + +With _OSI, the *BIOS* provides the string describing an interface, +and asks the OS: "YES/NO, are you compatible with this interface?" + +eg. _OSI("3.0 Thermal Model") would return TRUE if the OS knows how +to deal with the thermal extensions made to the ACPI 3.0 specification. +An old OS that doesn't know about those extensions would answer FALSE, +and a new OS may be able to return TRUE. + +For an OS-specific interface, the ACPI spec said that the BIOS and the OS +were to agree on a string of the form such as "Windows-interface_name". + +But two bad things happened. First, the Windows ecosystem used _OSI +not as designed, but as a direct replacement for _OS -- identifying +the OS version, rather than an OS supported interface. Indeed, right +from the start, the ACPI 3.0 spec itself codified this misuse +in example code using _OSI("Windows 2001"). + +This misuse was adopted and continues today. + +Linux had no choice but to also return TRUE to _OSI("Windows 2001") +and its successors. To do otherwise would virtually guarantee breaking +a BIOS that has been tested only with that _OSI returning TRUE. + +This strategy is problematic, as Linux is never completely compatible with +the latest version of Windows, and sometimes it takes more than a year +to iron out incompatibilities. + +Not to be out-done, the Linux community made things worse by returning TRUE +to _OSI("Linux"). Doing so is even worse than the Windows misuse +of _OSI, as "Linux" does not even contain any version information. +_OSI("Linux") led to some BIOS' malfunctioning due to BIOS writer's +using it in untested BIOS flows. But some OEM's used _OSI("Linux") +in tested flows to support real Linux features. In 2009, Linux +removed _OSI("Linux"), and added a cmdline parameter to restore it +for legacy systems still needed it. Further a BIOS_BUG warning prints +for all BIOS's that invoke it. + +No BIOS should use _OSI("Linux"). + +The result is a strategy for Linux to maximize compatibility with +ACPI BIOS that are tested on Windows machines. There is a real risk +of over-stating that compatibility; but the alternative has often been +catastrophic failure resulting from the BIOS taking paths that +were never validated under *any* OS. + +Do not use _REV +--------------- + +Since _OSI("Linux") went away, some BIOS writers used _REV +to support Linux and Windows differences in the same BIOS. + +_REV was defined in ACPI 1.0 to return the version of ACPI +supported by the OS and the OS AML interpreter. + +Modern Windows returns _REV = 2. Linux used ACPI_CA_SUPPORT_LEVEL, +which would increment, based on the version of the spec supported. + +Unfortunately, _REV was also misused. eg. some BIOS would check +for _REV = 3, and do something for Linux, but when Linux returned +_REV = 4, that support broke. + +In response to this problem, Linux returns _REV = 2 always, +from mid-2015 onward. The ACPI specification will also be updated +to reflect that _REV is deprecated, and always returns 2. + +Apple Mac and _OSI("Darwin") +---------------------------- + +On Apple's Mac platforms, the ACPI BIOS invokes _OSI("Darwin") +to determine if the machine is running Apple OSX. + +Like Linux's _OSI("*Windows*") strategy, Linux defaults to +answering YES to _OSI("Darwin") to enable full access +to the hardware and validated BIOS paths seen by OSX. +Just like on Windows-tested platforms, this strategy has risks. + +Starting in Linux-3.18, the kernel answered YES to _OSI("Darwin") +for the purpose of enabling Mac Thunderbolt support. Further, +if the kernel noticed _OSI("Darwin") being invoked, it additionally +disabled all _OSI("*Windows*") to keep poorly written Mac BIOS +from going down untested combinations of paths. + +The Linux-3.18 change in default caused power regressions on Mac +laptops, and the 3.18 implementation did not allow changing +the default via cmdline "acpi_osi=!Darwin". Linux-4.7 fixed +the ability to use acpi_osi=!Darwin as a workaround, and +we hope to see Mac Thunderbolt power management support in Linux-4.11. -- cgit v1.2.3 From ee39222d45839284f91e882df8a985e1af0f7e18 Mon Sep 17 00:00:00 2001 From: "Rafael J. Wysocki" Date: Wed, 30 Nov 2016 02:52:38 +0100 Subject: ACPI / property: Document usage rules for _DSD properties Following some discussions during the Kernel Summit and LPC, document what can be returned from ACPI _DSD as device properties and when it is valid to use the special PRP0001 device ID. Signed-off-by: Rafael J. Wysocki Reviewed-by: Mika Westerberg Acked-by: Andy Shevchenko Reviwed-by: Mark Brown --- Documentation/acpi/DSD-properties-rules.txt | 97 +++++++++++++++++++++++++++++ Documentation/acpi/enumeration.txt | 9 +++ 2 files changed, 106 insertions(+) create mode 100644 Documentation/acpi/DSD-properties-rules.txt (limited to 'Documentation/acpi') diff --git a/Documentation/acpi/DSD-properties-rules.txt b/Documentation/acpi/DSD-properties-rules.txt new file mode 100644 index 000000000000..3e4862bdad98 --- /dev/null +++ b/Documentation/acpi/DSD-properties-rules.txt @@ -0,0 +1,97 @@ +_DSD Device Properties Usage Rules +---------------------------------- + +Properties, Property Sets and Property Subsets +---------------------------------------------- + +The _DSD (Device Specific Data) configuration object, introduced in ACPI 5.1, +allows any type of device configuration data to be provided via the ACPI +namespace. In principle, the format of the data may be arbitrary, but it has to +be identified by a UUID which must be recognized by the driver processing the +_DSD output. However, there are generic UUIDs defined for _DSD recognized by +the ACPI subsystem in the Linux kernel which automatically processes the data +packages associated with them and makes those data available to device drivers +as "device properties". + +A device property is a data item consisting of a string key and a value (of a +specific type) associated with it. + +In the ACPI _DSD context it is an element of the sub-package following the +generic Device Properties UUID in the _DSD return package as specified in the +Device Properties UUID definition document [1]. + +It also may be regarded as the definition of a key and the associated data type +that can be returned by _DSD in the Device Properties UUID sub-package for a +given device. + +A property set is a collection of properties applicable to a hardware entity +like a device. In the ACPI _DSD context it is the set of all properties that +can be returned in the Device Properties UUID sub-package for the device in +question. + +Property subsets are nested collections of properties. Each of them is +associated with an additional key (name) allowing the subset to be referred +to as a whole (and to be treated as a separate entity). The canonical +representation of property subsets is via the mechanism specified in the +Hierarchical Properties Extension UUID definition document [2]. + +Property sets may be hierarchical. That is, a property set may contain +multiple property subsets that each may contain property subsets of its +own and so on. + +General Validity Rule for Property Sets +--------------------------------------- + +Valid property sets must follow the guidance given by the Device Properties UUID +definition document [1]. + +_DSD properties are intended to be used in addition to, and not instead of, the +existing mechanisms defined by the ACPI specification. Therefore, as a rule, +they should only be used if the ACPI specification does not make direct +provisions for handling the underlying use case. It generally is invalid to +return property sets which do not follow that rule from _DSD in data packages +associated with the Device Properties UUID. + +Additional Considerations +------------------------- + +There are cases in which, even if the general rule given above is followed in +principle, the property set may still not be regarded as a valid one. + +For example, that applies to device properties which may cause kernel code +(either a device driver or a library/subsystem) to access hardware in a way +possibly leading to a conflict with AML methods in the ACPI namespace. In +particular, that may happen if the kernel code uses device properties to +manipulate hardware normally controlled by ACPI methods related to power +management, like _PSx and _DSW (for device objects) or _ON and _OFF (for power +resource objects), or by ACPI device disabling/enabling methods, like _DIS and +_SRS. + +In all cases in which kernel code may do something that will confuse AML as a +result of using device properties, the device properties in question are not +suitable for the ACPI environment and consequently they cannot belong to a valid +property set. + +Property Sets and Device Tree Bindings +-------------------------------------- + +It often is useful to make _DSD return property sets that follow Device Tree +bindings. + +In those cases, however, the above validity considerations must be taken into +account in the first place and returning invalid property sets from _DSD must be +avoided. For this reason, it may not be possible to make _DSD return a property +set following the given DT binding literally and completely. Still, for the +sake of code re-use, it may make sense to provide as much of the configuration +data as possible in the form of device properties and complement that with an +ACPI-specific mechanism suitable for the use case at hand. + +In any case, property sets following DT bindings literally should not be +expected to automatically work in the ACPI environment regardless of their +contents. + +References +---------- + +[1] http://www.uefi.org/sites/default/files/resources/_DSD-device-properties-UUID.pdf +[2] http://www.uefi.org/sites/default/files/resources/_DSD-hierarchical-data-extension-UUID-v1.1.pdf diff --git a/Documentation/acpi/enumeration.txt b/Documentation/acpi/enumeration.txt index a91ec5af52df..209a5eba6b87 100644 --- a/Documentation/acpi/enumeration.txt +++ b/Documentation/acpi/enumeration.txt @@ -415,3 +415,12 @@ the "compatible" property in the _DSD or a _CID as long as one of their ancestors provides a _DSD with a valid "compatible" property. Such device objects are then simply regarded as additional "blocks" providing hierarchical configuration information to the driver of the composite ancestor device. + +However, PRP0001 can only be returned from either _HID or _CID of a device +object if all of the properties returned by the _DSD associated with it (either +the _DSD of the device object itself or the _DSD of its ancestor in the +"composite device" case described above) can be used in the ACPI environment. +Otherwise, the _DSD itself is regarded as invalid and therefore the "compatible" +property returned by it is meaningless. + +Refer to DSD-properties-rules.txt for more information. -- cgit v1.2.3