uACPI

A portable and easy-to-integrate implementation of the Advanced Configuration and Power Interface (ACPI).

Features

A fast and well-tested AML interpreter optimized to use very little stack space
NT-compatible on a fundamental level (see examples)
Very easy to integrate (ships with own overridable standard library implementation)
Highly flexible and configurable (optional sized frees, reduced-hw-only mode, etc.)
A fairly advanced event subsystem (GPE/fixed, wake, implicit notify, AML handlers)
Table management API (search, dynamic installation/loading, overrides, etc.)
Operation region subsystem (user handlers, support for BufferAcc opregions, builtins for common types)
Sleep state management (transition to any S state, wake vector programming)
PCI routing table retrieval & interrupt model API
Device search API
Resource subsystem supporting every resource defined by ACPI 6.6
Interface & feature management exposed via _OSI
Client-defined Notify() handlers
Firmware global lock management (_GL, locked fields, public API)
GAS read/write API
Fully thread safe
Supports both 32-bit and 64-bit platforms
A special barebones mode with only table API (see config.h)

Why would I use this over ACPICA?

1. Performance

uACPI shows a consistent speedup of about 3.5x over ACPICA in synthetic AML tests.

More details

Code that was tested:

Method (TOMS, 2, NotSerialized) {
    Return ((Arg1 - Arg0) / 10000)
}

Method (ADDM, 1, NotSerialized) {
    Local0 = 0

    While (Arg0) {
        Local0 += Arg0 + Arg0
        Arg0--
    }

    Return (Local0)
}

// Record start time
Local0 = Timer

// Run 10 million additions
Local1 = ADDM(10000000)

// Make sure the answer matches expected
If (Local1 != 0x5AF31112D680) {
    Printf("Bad test result %o", Local1)
    Return (1)
}

// Record end time
Local2 = Timer

Printf("10,000,000 additions took %o ms",
       ToDecimalString(TOMS(Local0, Local2)))

Compile options (acpiexec and uACPI's test-runner): -O3 -flto -march=znver4 -mtune=znver4
CPU: AMD Ryzen 9 9950X3D
Raw test scores (~average over 10 runs):

ACPICA: 16661 ms
uACPI: 4753 ms

Raw difference: 3.5053x

Real hardware tests of the same operating system using uACPI vs ACPICA show at least a 1.75-2x speedup while measuring the time it takes to load the initial AML namespace.

More details

OS: proxima

Compile options: -O3 -flto

Test Subject 1

Specs: Gigabyte B550M S2H, AMD Ryzen 5800X, 64GB RAM
Firmware: F19d (10097 AML opcodes)

Results:

ACPICA: 3,936,953 ns
uACPI: 1,902,077 ns

Raw difference: 2.0698x

Test Subject 2

Specs: Toshiba Portege R30-A, Intel Core i5-4200M, 4GB RAM
Firmware: 4.40 (4962 AML opcodes)

Results:

ACPICA: 10,899,233 ns
uACPI: 6,227,036 ns

Raw difference: 1.7503x

2. NT-compatible from the ground up

Over the decades of development, ACPICA has accumulated a lot of workarounds for AML expecting NT-specific behaviors, and is still missing compatibility in a lot of critical aspects.

uACPI, on the other hand, is built to be natively NT-compatible without extra workarounds.

Some specific highlights include:

Reference objects, especially multi-level reference chains
Implicit cast semantics
Object mutability
Named object resolution, especially for named objects inside packages

3. Fundamental safety

uACPI is built to always assume the worst about the AML byte code it's executing, and as such, has a more sophisticated object lifetime tracking system, as well as carefully designed handling for various edge-cases, including race conditions.

Some of the standard uACPI test cases crash both ACPICA, and the NT AML interpreters.

While a permanent fuzzing solution for uACPI is currently WIP, it has already been fuzzed quite extensively and all known issues have been fixed.

4. No recursion

Running at kernel level has a lot of very strict limitations, one of which is a tiny stack size, which can sometimes be only a few pages in length.

Of course, both ACPICA and uACPI have non-recursive AML interpreters, but there are still edge cases that cause potentially unbounded recursion.

One such example are the dynamic table load operators from AML (Load/LoadTable): these cause a linear growth in stack usage per call in ACPICA, whereas in uACPI these are treated as special method calls, and as such, don't increase stack usage whatsoever.

More detailed overview

Expressions within package:

Method (TEST) {
    Local0 = 10
    Local1 = Package { Local0 * 5 }
    Return (DerefOf(Local1[0]))
}

// ACPICA: AE_SUPPORT, Expressions within package elements are not supported
// Windows, uACPI: Local0 = 50
Local0 = TEST()

Packages outside of a control method:

// ACPICA: internal error
// Windows, uACPI: ok
Local0 = Package { 1 }

Reference rebind semantics:

Local0 = 123
Local1 = RefOf(Local0)

// ACPICA: Local1 = 321, Local0 = 123
// Windows, uACPI: Local1 = reference->Local0, Local0 = 321
Local1 = 321

Increment/Decrement:

Local0 = 123
Local1 = RefOf(Local0)

// ACPICA: error
// Windows, uACPI: Local0 = 124
Local1++

Multilevel references:

Local0 = 123
Local1 = RefOf(Local0)
Local2 = RefOf(Local1)

// ACPICA: Local3 = reference->Local0
// Windows, uACPI: Local3 = 123
Local3 = DerefOf(Local2)

Implict-cast semantics:

Name (TEST, "BAR")

// ACPICA: TEST = "00000000004F4F46"
// Windows, uACPI: TEST = "FOO"
TEST = 0x4F4F46

Buffer size mutability:

Name (TEST, "XXXX")
Name (VAL, "")

// ACPICA: TEST = "LONGSTRING"
// Windows, UACPI: TEST = "LONG"
TEST = "LONGSTRING"

// ACPICA: VAL = "FOO"
// Windows, UACPI: VAL = ""
VAL = "FOO"

Returning a reference to a local object:

Method (TEST) {
    Local0 = 123

    // Use-after-free in ACPICA, perfectly fine in uACPI
    Return (RefOf(Local0))
}

Method (FOO) {
    Name (TEST, 123)

    // Use-after-free in ACPICA, object lifetime prolonged in uACPI (node is still removed from the namespace)
    Return (RefOf(TEST))
}

CopyObject into self:

Method (TEST) {
    CopyObject(123, TEST)
    Return (1)
}

// Segfault in ACPICA, prints 1 in uACPI  
Debug = TEST()

// Unreachable in ACPICA, prints 123 in uACPI
Debug = TEST

There's even more examples, but this should be enough to demonstrate the fundamental differences in designs.

Integrating into a kernel

1. Add uACPI sources & include directories into your project

If you're using CMake

Simply add the following lines to your cmake:

include(uacpi/uacpi.cmake)

target_sources(
    my-kernel
    PRIVATE
    ${UACPI_SOURCES}
)

target_include_directories(
    my-kernel
    PRIVATE
    ${UACPI_INCLUDES}
)

If you're using Meson

Add the following lines to your meson.build:

uacpi = subproject('uacpi')

uacpi_sources = uacpi.get_variable('sources')
my_kernel_sources += uacpi_sources

uacpi_includes = uacpi.get_variable('includes')
my_kernel_includes += uacpi_includes

Any other build system

Add all .c files from source into your target sources
Add include into your target include directories

2. Implement/override platform-specific headers

uACPI defines all platform/architecture-specific functionality in a few headers inside include/uacpi/platform

All of the headers can be "implemented" by your project in a few ways:

Implement the expected helpers exposed by the headers
Replace the expected helpers by your own and override uACPI to use them by defining the respective UACPI_OVERRIDE_X variable. In this case, the header becomes a proxy that includes a corresponding uacpi_x.h header exported by your project.

Currently used platform-specific headers are:

arch_helpers.h - defines architecture/cpu-specific helpers & thread-id-related interfaces
compiler.h - defines compiler-specific helpers like attributes and intrinsics. This already works for MSVC, clang & GCC so you most likely won't have to override it.
atomic.h - defines compiler-specific helpers for dealing with atomic operations. Same as the header above, this should work out of the box for MSVC, clang & GCC.
libc.h - an empty header by default, but may be overriden by your project if it implements any of the libc functions used by uACPI (by default uACPI uses its own implementations to be platform-independent and to make porting easier). The internal implementation is just the bare minimum and not optimized in any way.
types.h - typedefs a bunch of uacpi-specific types using the stdint.h header. You don't have to override this unless you don't provide stdint.h.
config.h - various compile-time options and settings, preconfigured to reasonable defaults.

3. Implement kernel API

uACPI relies on kernel-specific API to do things like mapping/unmapping memory, writing/reading to/from IO, PCI config space, and many more things.

This API is declared in kernel_api.h and is implemented by your kernel.

4. Initialize uACPI

That's it, uACPI is now integrated into your project.

You should proceed to initialization.
Refer to the uACPI page on osdev wiki to see a snippet for basic initialization, as well as some code examples of how you may want to use certain APIs.

All of the headers and APIs defined in uacpi are public and may be utilized by your project.
Anything inside uacpi/internal is considered private/undocumented and unstable API.

Developing and contributing

Most development work is fully doable in userland using the test runner.

Setting up an IDE:

Simply open tests/runner/CMakeLists.txt in your favorite IDE.

For Visual Studio:

cd tests\runner && mkdir build && cd build && cmake ..

Then just simply open the .sln file generated by cmake.

Running the test suite:

./tests/run_tests.py

If you want to contribute:

Commits are expected to be atomic (changing one specific thing, or introducing one feature) with detailed description (if one is warranted for), an S-o-b line is welcome
Code style is 4-space tabs, 80 cols, the rest can be seen by just looking at the current code

All contributions are very welcome!

Notable projects using uACPI & performance leaderboards

License

uACPI is licensed under the MIT License.
The full license text is provided in the LICENSE file inside the root directory.

GitHub Analytics

uACPI

Quality

uACPI

Features

Why would I use this over ACPICA?

1. Performance

Test Subject 1

Test Subject 2

2. NT-compatible from the ground up

3. Fundamental safety

4. No recursion

More detailed overview

Integrating into a kernel

1. Add uACPI sources & include directories into your project

If you're using CMake

If you're using Meson

Any other build system

2. Implement/override platform-specific headers

3. Implement kernel API

4. Initialize uACPI

Developing and contributing

Setting up an IDE:

Running the test suite:

Notable projects using uACPI & performance leaderboards

License