AuroraRuntime/README.md
2022-03-17 09:04:58 +00:00

21 KiB

IN DEVELOPMENT

AuroraRuntime

The Aurora Runtime is a low level platform abstraction layer for cross-platform C++ development targeting embedded and PC systems. Simply fetch a binary package for your toolchain or integrate the buildscripts into your applications build pipeline to get started.

console picture

Features

  • Little to no C++ standard template library dependence (^1)
  • High performance threading and synchronization primitives (os userland sched optimized)
  • Async threading primitives, including WaitMultipleObjects paradigm (local IPC)
  • Asynchronous and synchronous IO (network, character, file, and buffered)
  • Optional event driven async programming paradigm
  • Consoles; graphical and standard, file archives
  • Logging; UTF-8 logger, common sink backends
  • Debug and Telementry; asserts, exception logging, fio, nio backends
  • Crypto ECC/[25519, P-384, P-256], [AES, RSA, X509], [common digests]
  • Basic cmdline parsing from any module
  • Exit and fatal save condition callbacks
  • IPC [WIP]
  • Network [WIP]
  • Random; secure and fast
  • Hardware Info; memory and cpu info
  • Software Stack Info
  • FIO settings registry
  • Compression
  • Locale and encoding
  • C++ utility templates and macros
  • Follows all strings are UTF-8 convention

^1 bring your own types auROXTL

API:

Examples: Hello Aurora
Doxygen:
Tests: Hello Aurora
Cmake-stable:
Build Pipeline: https://git.reece.sx/AuroraPipeline/Build

Performance

Performance should be that of native performance, ideally that of the best implementation on the platform, and no worse than the STL. Due to heavyweight requirements and spiral model defined objectives, a handful of portable C libraries have been brought into achieve compression, crypto, alloc, and str formatting objectives using known good industry standard libraries. Footprint is expected to be on the heavier side with optimal performance (incl toll for C++ pluses tradeoffs).

Runtime as of 2022-02-01 with the entire wxWidgets toolkit and compression libraries on Windows LTSC 2019:

DLL Disk: 4.2MB
Size Of Image: 0x50B000 (5MB)
Real Commit Charge of a Console App: 9.7MB
...the task manager lie: 3,168K (3.1MB -> less than our DLL)
...HWInfo reports

[13:07:38] [Info] | RamInfo Private Allocation: 10215424/71987290112 (^1)
[13:07:38] [Info] | RamInfo Address Space: 11669504/71987290112

^1 ...on LTSC 2019. Modern Windows 10 and 11 will return the exact task manager value of 3,168K (3.1MB).

Defer to benchmarks

Utilities

Aurora Sugar: https://git.reece.sx/AuroraSupport/auROXTL/src/branch/master/Include/auROXTLUtils.hpp
Aurora Macro Sugar: https://git.reece.sx/AuroraSupport/auROXTL/src/branch/master/Include/auROXTL/AU_MACROS.hpp
Aurora Overloadable Type Declerations: https://git.reece.sx/AuroraSupport/auROXTL/src/branch/master/Include/auROXTLTypedefs.hpp

Logging

Logging is implemented through 2 subsystems, console and logging. Console provides IO abstraction to logger subsystem sinks. Sinks are user implementable interfaces that can be either synchronous or asynchronous. Loggers are defined as an internal object the takes logger messages and dumps them to the relevant subscriber.

Flushing occurs at a fixed rate on a low prio secondary thread without any configuration requirement. The resources spent on the thread is shared with the telemetry and debug subsystems to reduce the thread overhead of the runtime. Flushes also occur during panic events and other relevant problematic points.

Asynchronous logger sinks may double buffer log lines between the asynchronous callback and on flush, where the latter call is guaranteed after one or more delegated dispatch. The Windows Event Log backend takes advantage of this to multi-line group messages by approx time and log level.

Additionally, in the console subsystem, consoles that provide an input stream can be used in conjunction with the parse subsystem to provide basic command-based deserialization, tokenization, and dispatch of UTF-8 translated strings regardless of the system locale. Command processing comes under the namespace of consoles, not logging.

Exceptions

ICanHasStackTraces

Through the use of compiler internal overloads, ELF hooking, and Win32 AddVectoredExceptionHandler, Aurora Runtime hooks exceptions at the time of throw, including some out of ecosystem exceptions, providing detailed telemetry of the object type, object string, and backtrace. In addition, the AuDebug namespace provides TLS based last-error and last-backtrace methods.

EXCEPTIONS ARE NOT CONTROL FLOW...

  • Aurora Runtime WILL attempt to mitigate exceptions in internal logic
  • Aurora Runtime WILL NOT abuse exceptions to communicate failure
  • Aurora Runtime WILL try to decouple internal exceptions from the API
  • Aurora Runtime WILL NOT use anything that automatically crashes on exception catch (no-noexcept)
  • Aurora Runtime WILL provide extended exception information to telemetry backends and through the AuDebug namespace
  • Aurora Runtime WILL NOT make any guarantees of being globally-noexcept; however, it should be a safe assumption in non-critical environments

SysPanic can be used to format a std::terminate-like exit condition, complete with telemetry data and safe cleanup.

Debug

Error Markers

SysPushError[EFailureCategory shorthand] can be used to include additional side-channel telemetry information about the execution of a program. SysPushError_(...) takes a string format sequence and a variadic sequence of substitute values - or no arguments whatsoever.

Example:

IBufferedCharacterConsumer *BufferConsumerFromProviderNew(const AuSPtr<ICharacterProvider> &provider)
{
    if (!provider)
    {
        SysPushErrorArg("Missing ICharacterProvider");
        return {};
    }

    return _new BufferedCharacterConsumer(provider);
}

Asserts

[TODO]

Example:

Debug, Release, and Ship (all) assertions:

SysAssert(AuFunction{}, "unexpected default function")

Debug and Release (debug and optimized ship-with-debug) assertions:

SysAssertDbg(AuFunction{}, "unexpected default function")

Something went wrong

You should ensure AuDebug::CheckErrors() or a SysPushError-like function is called to ensure enchanced TLS-state telemetry is captured. AuDebug::PrintErrors() will print the the errors gathered by the debug subsystem for telemetry purposes. These may include the crts errno, the last reported posix return value, the last Win32 error code, and/or last reported microkernel error.

Example

Try/Catch:

try
{
    
}
catch (...)
{
    SysPushErrorCatch(); // THIS IS NOT REQUIRED FOR EXCEPTION TELEMETRY
}

Windows System Error Messages:
CRT:


POSIX:


Loop

The Aurora Runtime implements a loop concept similar to macos's RunLoop or BSD's kevent. On Mach platforms, kevent objects can be passed to loop sources. On BSD platforms, kevent objects can be used to synchronize against a wait any condition. On Linux, io_submit with epoll semantics can be used instead. NT, surpisingly where WaitForMultipleObjects is common, has the worst abstraction. Either way, all modern platforms provide a way to join file handles, an object from which you can derive any othe thread primitive from, semaphores, mutexes, network sockets, and asynchronous file transactions together. Aurora Runtime is implementing a loop source object and wait for any utility to solve this fiddly and complex abstraction. This should allow for non-bruteforced low thread count work groups to work on queues of ipcable objects across platforms with low synchronization overhead.

Thread Primitives

The Aurora Runtime provides platform optimized threading primitives inheriting from a featureful IWaitable interface. Each method is guaranteed. Each primitive is user-land scheduler optimized.

struct IWaitable
{
   virtual bool TryLock()                            = 0;
   virtual void Lock(relativeTimeoutInMilliseconds)  = 0;
   virtual void Lock()                               = 0;
   virtual void Unlock()                             = 0;
}

Included high performance primitives

  • arbitrary condition variable ^1
  • condition mutex
  • condition variable
  • critical section ^2
  • event
  • mutex
  • semaphore
  • rwlock ^3
  • spinlocks

^1 Accepts any IWaitable as the mutex
^2 Reentrant Mutex
^3 Includes extended read to write upgrades and permits write-entrant read-routines to prevent writer deadlocks.

Fixing problems in other scheduler apis

Problem one (1):
Most STL implementations have generally awful to unnecessarily inefficient abstraction. Defer to libc++'s abuse of spin while (cond) yield loops and msvc/stl's painfully slow std::mutex and semaphore primitives.

Problem Two (2):
Moving to or from linux, macos, bsd, and win32 under varous kernels, there is no one standard (even in posix land) for the key thread primitives.

Bonus point NT (3):
The userland CriticalSection/CV set of APIs suck, lacking timeouts and try lock

Bonus point UNIX (4):
No wait multiple mechanism

1, 2, 3: Use the high performance AuThreadPrimitives objects

4: Consider using loop sources, perhaps with the async subsystem, in your async application. Performance of loop sources will vary wildly between platforms, always being generally worse than the high performance primitives. They should be used to observe kernel-level signalable resources.

4 ex: Windows developers could use loop sources as a replacement to WaitMultipleObjects with more overhead

Strings

The Aurora Runtime defines an AuString type as an std::string; however, it should be assumed this type represents a binary blob of UTF-8. Looking to switch to tiny-utf8 for UTF-8 safety.

Memory

Allocator

Objects are allocated across API/Module boundaries; and so long as the high level API design isn't horribly inefficient, the cache invalidation and indirect lookups should be minimalized. On modern hardware, indirect branching versus short jumps aren't so expensive, and in combination with a fast enough allocator, it's a model that could provide reasonable OOP performance through a C-with-classes-like API.

As for allocation, we generally expect a dependence on Microsoft's mimalloc. Linking against the Aurora Runtime in the Aurora Ecosystem will automatically replace global allocators with Aurora::Memory, which in turn, proxies any other suitable allocator with extended re[zero]allocate[aligned] with uniform deallocate APIs.

Memory Heap

Aurora provides a heap allocator for dividing up a large preallocated region of memory

Shared Pointers

By default, AuSPtr is backed by std::shared_ptr, extended by #include <Aurora/Memory/ExtendStlLikeSharedPtr>. Using this class, undefined behaviour on dereference and operator pointer is altered to guarantee an AU_THROW_STRING. Such hacks without support for native behaviour from the language drivers themselves can be rather expensive; however, it's an experiment worth trying now that modern hardware can make up for software and microcode flaws; and architecture translation. Null exception experiments can be easily disabled or configured from within your AuroraConfiguration.h file globally.

Types: 
    AuSPtr<Type_t>
    AuWPtr<Type_t>
    AuUPtr<Type_t, Deleter_t>
	
Functions:
    AuSPtr<T> AuMakeShared<T>(Args&& ...)
    AuSPtr<T> AuUnsafeRaiiToShared<T>(T *)
    AuSPtr<T> AuUnsafeRaiiToShared<T>(AuUPtr<T>) 

Macros:
    _new (pseudo no-throw new operator)
    AuSPtr<This_T> AuSharedFromThis()
    AuWPtr<This_T> AuWeakFromThis()
    AuFunction<...> AuBindThis(This_t *::?, ...)

Note

Aurora provides a bring your own container and shared pointer model overloadable in your configuration header.
User-overloadable type declerations and generic access utilities are defined under auROXTL

Binding

Aurora Runtime provides C++ APIs; however, it should be noted that two libraries are used to extend interfaces and enums to help with porting and internal utility access. One, AuroraEnums, wraps basic enumerations and provides value vectors; value strings; look up; iteration; and more. The other, AuroraInterfaces, provides TWO class types for each virtual interface. Each interface can be backed by a; C++ class method overriding a superclass's virtual ...(...) = 0; method, or a AuFunctional -based structure.

It should be noted that most language bindings and generator libraries (^swig, v8pp, nbind, luabind) work with shared pointers. Other user code may wish to stuff pointers into a machineword-sized space, whether its a C library, a FFI, or a size constraint. One handle or abstraction layer will be required to integrate the C++ API into the destination platform, and assuming we have a C++ language frontend parsing our API, we can use AuSPtr for all caller-to-method constant reference scanerios. Furthermore, AuSPtrs can be created, without a deletor, using AuUnsafeRaiiToShared(unique/raw pointer). To solve the raw pointer issue, AuSPtrs are created in the public headers with the help of exported/default visibility interface create and destroy functions. These APIs provide raw pointers to public C++ interfaces, and as such, can be binded using virtually any shim generator. Method and API mapping will likely involve manual work from the library developer to reimplement AU concepts under their language runtime instead of using the C++ platform, or at least require manual effort to shim or map each runtime prototype into something more sane across the language barrier.

Memory is generally viewed through a std::span like concept called MemoryViews. MemoryViewRead and MemoryViewWrite provide windows into a defined address range. MemoryViewStreamRead and MemoryViewStreamWrite expand upon this concept by accepting an additional offset (AuUInt &: reference) that is used by internal APIs to indicate how many bytes were written or read from a given input region. Such requirement came about from so many APIs, networking, compression, encoding, doing the exact same thing in different not-so-portable ways. Unifying memory access to 4 class types should aid with SWIG prototyping.

Unrelated note, structure interfacing with questionable C++ ABI reimplementations is somewhat sketchy in FFI projects (^ CppSharp) can lead to some memory leaks.

IO

[TODO] Summary

An important note about texting encoding. Stdin, file encoding, text decoders, and other IO resources work with codepage UTF-8 as the internal encoding scheme. String overloads and dedicated string APIs in the IO subsystem will always write BOM prefixed UTF-8 and attempt to read a BOM to translate any other input to UTF-8.

FIO

[TODO] async, fio abstraction, utf8 read/write, blob read/write, stat, dir recursion, stream abstraction

Paths

We assume all paths are messy. Incorrect splitters, double splitters, relative paths, and keywords are resolved internally. No URL or path builder, data structure to hold a tokenized URI expression, or similar concept exists in the codebase. All string 'paths' are simply expanded, similar to MSCRT's fullpath or UNIX's realpath, at time of usage.

Expression Meaning
Path[0] == '.' Current Working Directory
Path[0] == '^' Executable module's Directory
Path[0] == '~' User Profile Storage + SDK brand
Path[0] == '!' All User Shared Storage + SDK brand
.. Go up a directory
/ Agnostic Directory Splitter
\ Agnostic Directory Splitter
. [SPLITTER] Nothing

[TODO] Aurora Branding

Resources

[TODO] Aurora IO Resources

NIO

  • Worker thread delegated resolve using system resolver
  • Callback with fence-id based asynchronous write abstraction
  • Loop Source support

CIO

BIO

Aurora Async

The Aurora Runtime offers an optional asynchronous task driven model under the AuAsync namespace. Featuring promises, thread group pooling, functional-to-task wrapping, and task-completion callback-task-dispatch idioms built around 3 concepts.

Example:

Proccesses

The Aurora Runtime provides worker process monitoring, worker stdin/out stream redirection process spawning, file opening, and url opening functionality.

Locale

Encoding and decoding UTF-8, UTF-16, UTF-32, GBK, GB-2312, and SJIS is supported through platform provided decoders. Fetch system language and country backed by environment variables, the OS system configuration, the unix locale env variable, and/or the provided overload mechanism.

Dependencies

Aurora

Crypto (third party)

Compression (third party)

Utility (third party)

^1 Include-only macro library
^2 Provides core utilities and stl decoupling
^3 Provides platform information, included by default by the Aurora build pipeline
^4 C++ 20 saw another pathetic adoption attempt of an open source library, this one actually passed, but hardly
anyone implements std::format. Not to mention such is only a subset of the original library.
^5 Public Domain
^6 Potentially STL heavy, still potentially portable w/ a modern-ish toolchain

Philosophies

  • Assume C++17 language support in the language driver

  • Use AuXXX type bindings for std types, allow customers to overload the std namespace We assume some containers and utility APIs exist, but where they come from is up to you

  • Keep the code and build chain simple such that any C++ developer could maintain their own software stack built around aurora components.

  • Dependencies and concepts should be cross-platform, cross-architecture, cross-ring friendly

    It is recommended to fork and replace any legacy OS specific code with equivalent AuroraRuntime concepts, introducing a circular dependency with the Aurora Runtime

    APIs shouldn't be designed around userland, mobile computing, or desktop computing; AuroraRuntime must provide a common backbone for all applications.

    Locale and user-info APIs will be limited due to the assumption userland is not a concept

  • Dependencies, excluding core reference algorithms (eg compression), must be rewritten and phased out over time.

  • Dependencies should not be added if most platforms provide some degree of native support
    Examples:
    -> Don't depend on a pthread shim for windows; implement the best thread
    primitives that lie on the best possible api for them
    -> Don't depend on ICU when POSIX's iconv and Win32's multibyte apis cover
    everything a conservative developer cares about; chinese, utf-16, utf-8,
    utf-32 conversion, on top of all the ancient windows codepages

  • Dependencies should only be added conservatively when it saves development time and provides production hardening
    Examples:
    -> Use embedded crypto libraries; libtomcrypt, libtommath
    ->> While there are some bugs in libtomcrypt and others, none appear to
    cryptographically cripple the library. Could you do better?
    -> Use portable libraries like mbedtls, O(1) heap, mimalloc
    ->> Writing a [D]TLS/allocator stack would take too much time
    ->> Linking against external allocators, small cross-platform utilities, and
    so on is probably fine
    -> Shim libcurl instead of inventing yet another http stack