AuroraRuntime/Source/IO/Loop/LSLocalSemaphore.cpp

/***
    Copyright (C) 2023 J Reece Wilson (a/k/a "Reece"). All rights reserved.

    File: LSLocalSemaphore.cpp
    Date: 2023-10-21
    Author: Reece
***/
#include <RuntimeInternal.hpp>
#include "LSLocalSemaphore.hpp"
#include <Source/Threading/Primitives/SMTYield.hpp>

namespace Aurora::IO::Loop
{
    extern AuRWRenterableLock gWaitForMultipleALLLock;

    LSLocalSemaphore::LSLocalSemaphore()
    {

    }

    LSLocalSemaphore::~LSLocalSemaphore()
    {

    }

    bool LSLocalSemaphore::TryInit(AuUInt32 initialCount, AuUInt32 uMaxCount)
    {
        if (!LSSemaphore::TryInit(initialCount))
        {
            return false;
        }

        this->uAtomicSemaphore = initialCount;
        this->uAtomicKernelSemaphore = initialCount;
        this->uMaxCount = uMaxCount;
        return true;
    }

    void LSLocalSemaphore::DoParanoia()
    {
        if (auto uCount = AuAtomicLoad(&this->uAtomicSemaphore))
        {
            AuAtomicAdd(&this->uAtomicKernelSemaphore, uCount);
            LSSemaphore::AddMany(uCount);
        }
    }

    bool LSLocalSemaphore::OnTrigger(AuUInt handle)
    {
        auto bRet = this->TryTakeNoSpin();

        while (true)
        {
            auto uOld = AuAtomicLoad(&this->uAtomicKernelSemaphore);

            if (uOld == 0)
            {
                DoParanoia();
                break;
            }

            if (AuAtomicCompareExchange(&this->uAtomicKernelSemaphore, uOld - 1, uOld) == uOld)
            {
                auto uCount = AuAtomicLoad(&this->uAtomicSemaphore);

                if (uOld - 1 == 0)
                {
                #if defined(AURORA_PLATFORM_LINUX)
                    if (uCount == 1)
                    {
                        // Don't acknowledge?
                        // Don't write into?
                        // saves two syscalls for nothang
                        AuAtomicAdd(&this->uAtomicKernelSemaphore, 1u);
                    }
                    else if (uCount)
                    {
                        AuAtomicAdd(&this->uAtomicKernelSemaphore, uCount);
                        LSSemaphore::AddMany(uCount - 1);
                    }
                #else
                    if (uCount)
                    {
                        AuAtomicAdd(&this->uAtomicKernelSemaphore, uCount);
                        LSSemaphore::AddMany(uCount);
                    #if !defined(AURORA_IS_MODERNNT_DERIVED)
                        (void)LSSemaphore::OnTrigger(0);
                    #endif
                    }
                #endif
                    else
                    {
                        (void)LSSemaphore::OnTrigger(0);
                    }
                }
                else if (uOld || !bRet)
                {
                    (void)LSSemaphore::OnTrigger(0);
                }

                break;
            }
        }

        return bRet;
    }

    bool LSLocalSemaphore::AddOne()
    {
        AuUInt32 uNext {};

        if (auto uMaxValue = this->uMaxCount)
        {
            while (true)
            {
                auto uCurrentValue = AuAtomicLoad(&this->uAtomicSemaphore);
                uNext              = uCurrentValue + 1;

                if (uNext > uMaxValue)
                {
                    return false;
                }

                if (AuAtomicCompareExchange(&this->uAtomicSemaphore, uNext, uCurrentValue) == uCurrentValue)
                {
                    break;
                }
            }
        }
        else
        {
            uNext = AuAtomicAdd(&this->uAtomicSemaphore, 1u);
        }

        while (true)
        {
            auto uCurrentValue = AuAtomicLoad(&this->uAtomicKernelSemaphore);
            auto uNextValue    = uCurrentValue;
            bool bCanReturn    = false;

            if (uCurrentValue < uNext)
            {
                uNextValue = uNext;
            }
            else
            {
                bCanReturn = true;
            }

            if (AuAtomicCompareExchange(&this->uAtomicKernelSemaphore, uNextValue, uCurrentValue) == uCurrentValue)
            {
                if (bCanReturn)
                {
                    return true;
                }
                else
                {
                    break;
                }
            }
        }

        return LSSemaphore::AddOne();
    }

    bool LSLocalSemaphore::AddMany(AuUInt32 uCount)
    {
        AuUInt32 uNext {};

        if (auto uMaxValue = this->uMaxCount)
        {
            while (true)
            {
                auto uCurrentValue = AuAtomicLoad(&this->uAtomicSemaphore);
                uNext              = uCurrentValue + uCount;

                if (uNext > uMaxValue)
                {
                    return false;
                }

                if (AuAtomicCompareExchange(&this->uAtomicSemaphore, uNext, uCurrentValue) == uCurrentValue)
                {
                    break;
                }
            }
        }
        else
        {
            uNext = AuAtomicAdd(&this->uAtomicSemaphore, uCount);
        }

    #if 0
        if (AuAtomicLoad(&this->uAtomicKernelSemaphore) >= uNext)
        {
            return true;
        }
        else
        {
            /* if AddMany add/load/kernel-wake race condition, it's the next AddMany persons problem. */
            /* uAtomicKernelSemaphore cannot be lower than uAtomicSemaphore, at the epilogue of the last unlock/adds tick. */

            /* If it somehow is, ::OnTrigger will check that the final kernel negative increment does not occur just before (linux) after (win32) (bool(this->uAtomicSemaphore)). */
            /* Remember: this->uAtomicKernelSemaphore should only be decremented after uAtomicSemaphore and uAtomicKernelSemaphore have already been incremented together... */
            /* ...therefore, the last kernel waker should always see bool(this->uAtomicSemaphore), unless stolen by another thread. */
            /*  if stolen, it's a race condition we dont care about; we avoided the kernel object and state entirely. We have to wait for another AddMany to wake us up. */
            /*  if not stolen, uAtomicSemaphore is read as non-zero, and is readded to the kernel semaphore */

            /* Most users just use the IO event objects LSAsync and LSLocalEvent. These are known good. */
        }
        AuAtomicAdd(&this->uAtomicKernelSemaphore, uCount);
    #else
        while (true)
        {
            auto uCurrentValue = AuAtomicLoad(&this->uAtomicKernelSemaphore);
            auto uNextValue    = uCurrentValue;
            bool bCanReturn    = false;

            if (uCurrentValue < uNext)
            {
                uNextValue = uNext;
            }
            else
            {
                bCanReturn = true;
            }

            if (AuAtomicCompareExchange(&this->uAtomicKernelSemaphore, uNextValue, uCurrentValue) == uCurrentValue)
            {
                if (bCanReturn)
                {
                    return true;
                }
                else
                {
                    uCount = uNext - uCurrentValue;
                    break;
                }
            }
        }
    #endif

        return LSSemaphore::AddMany(uCount);
    }

    bool LSLocalSemaphore::IsSignaled()
    {
        return this->TryTake();
    }

    bool LSLocalSemaphore::IsSignaledNoSpinIfUserland()
    {
        return this->TryTakeNoSpin();
    }

    ELoopSource LSLocalSemaphore::GetType()
    {
        return ELoopSource::eSourceFastSemaphore;
    }

    bool LSLocalSemaphore::TryTakeNoSpin()
    {
        AuUInt32 uOld {};

        while ((uOld = this->uAtomicSemaphore))
        {
            if (AuAtomicCompareExchange(&this->uAtomicSemaphore, uOld - 1, uOld) == uOld)
            {
                return true;
            }
        }

        return false;
    }

    bool LSLocalSemaphore::TryTakeSpin()
    {
        bool bRet = Threading::Primitives::DoTryIfAlderLake([&]
        {
            return this->TryTakeNoSpin();
        }, &this->uAtomicSemaphore);

    #if !defined(AU_NO_WAITMULTIPLELS_ALL_MS_PARITY)
        if (!bRet && gRuntimeConfig.ioConfig.bAimCloserForNTParityWaitALL)
        {
            AU_LOCK_GLOBAL_GUARD(gWaitForMultipleALLLock->AsReadable());
            bRet = this->TryTakeNoSpin();
        }
    #endif

        return bRet;
    }

    bool LSLocalSemaphore::TryTake()
    {
        return this->TryTakeSpin();
    }

    bool LSLocalSemaphore::TryTakeWaitNS(AuUInt64 uEndTime)
    {
        if (this->TryTakeSpin())
        {
            return true;
        }

        while (!this->TryTakeNoSpin())
        {
            if (!uEndTime)
            {
                if (LSSemaphore::WaitOn(0))
                {
                    return true;
                }
            }
            else
            {
                auto uStartTime = Time::SteadyClockNS();
                if (uStartTime >= uEndTime)
                {
                    return false;
                }

                auto uDeltaMs = AuNSToMS<AuInt64>(uEndTime - uStartTime);
                if (uDeltaMs &&
                    LSSemaphore::WaitOn(uDeltaMs))
                {
                    return true;
                }
                else if (!uDeltaMs)
                {
                    if (this->TryTakeSpin())
                    {
                        return true;
                    }
                }
            }
        }

        return true;
    }

    void LSLocalSemaphore::OnPresleep()
    {

    }

    void LSLocalSemaphore::OnFinishSleep()
    {

    }

    AUKN_SYM AuSPtr<ILSSemaphore> NewLSSemaphore(AuUInt32 uInitialCount)
    {
    #if defined(AURORA_IS_MODERNNT_DERIVED)
        if (gRuntimeConfig.ioConfig.bINeedPerfectAnds)
        {
            return NewLSSemaphoreSlow(uInitialCount);
        }
    #endif

        auto pMutex = AuMakeShared<LSLocalSemaphore>();
        if (!pMutex)
        {
            SysPushErrorGeneric();
            return {};
        }

        if (!pMutex->TryInit(uInitialCount))
        {
            SysPushErrorNested();
            return {};
        }

        return pMutex;
    }

    AUKN_SYM AuSPtr<ILSSemaphore>   NewLSSemaphoreEx(AuUInt32 uInitialCount, AuUInt32 uMaxCount)
    {
        auto pMutex = AuMakeShared<LSLocalSemaphore>();
        if (!pMutex)
        {
            SysPushErrorGeneric();
            return {};
        }

        if (!pMutex->TryInit(uInitialCount, uMaxCount))
        {
            SysPushErrorNested();
            return {};
        }

        return pMutex;
    }
}