AuroraRuntime/Source/Threading/Primitives/AuMutex.NT.cpp

/***
    Copyright (C) 2021 J Reece Wilson (a/k/a "Reece"). All rights reserved.

    File: AuMutex.NT.cpp
    Date: 2021-6-12
    Author: Reece
***/
#include <Source/RuntimeInternal.hpp>
#include "AuMutex.Generic.hpp"
#include "SMTYield.hpp"
#include "../AuWakeInternal.hpp"

#if !defined(_AURUNTIME_GENERICMUTEX)
#include "AuMutex.NT.hpp"
#include "AuConditionMutex.NT.hpp"
#include <Time/Time.hpp>

#if !defined(NTSTATUS_TIMEOUT)
    #define NTSTATUS_TIMEOUT 0x102
#endif

namespace Aurora::Threading::Primitives
{
    MutexImpl::MutexImpl()
    {
        if (!pWaitOnAddress)
        {
        #if defined(AURORA_FORCE_SRW_LOCKS)
            ::InitializeSRWLock(&this->atomicHolder_);
            ::InitializeConditionVariable(&this->wakeup_);
        #endif
        }

        this->state_ = 0;
    }

    MutexImpl::~MutexImpl()
    {

    }

    bool MutexImpl::HasOSHandle(AuMach &mach)
    {
        return false;
    }

    bool MutexImpl::TryLockHeavy()
    {
        return DoTryIf([=]()
        {
            return this->TryLockNoSpin();
        });
    }

    bool MutexImpl::TryLock()
    {
        if (gRuntimeConfig.threadingConfig.bPreferNtMutexSpinTryLock)
        {
            return TryLockHeavy();
        }
        else
        {
            return TryLockNoSpin();
        }
    }

    bool MutexImpl::TryLockNoSpin()
    {
        return AuAtomicTestAndSet(&this->state_, 0) == 0;
    }

    bool MutexImpl::HasLockImplementation()
    {
        return true;
    }

    void MutexImpl::SlowLock()
    {
        auto status = LockNS(0);
        SysAssert(status, "Couldn't lock Mutex object");
    }

    bool MutexImpl::LockMS(AuUInt64 uTimeout)
    {
        if (this->TryLockNoSpin())
        {
            return true;
        }

        return this->LockNS(AuMSToNS<AuUInt64>(uTimeout));
    }

    bool MutexImpl::LockNS(AuUInt64 uTimeout)
    {
        bool returnValue = false;

        if (this->TryLockHeavy())
        {
            return true;
        }

        AuUInt64 uEndTime   = uTimeout ? Time::SteadyClockNS() + uTimeout : 0;
        int iYieldCounter {};

        if (gUseNativeWaitMutex)
        {
            while (!this->TryLockNoSpin())
            {
                auto &uValueRef = this->state_;
                auto uValue = uValueRef | 1;
                auto uNextValue = uValue + kFutexBitWait;

                if (AuAtomicCompareExchange(&uValueRef, uNextValue, uValue) == uValue)
                {
                    if (!InternalLTSWaitOnAddressHighRes((void *)&uValueRef, &uNextValue, sizeof(uNextValue), uEndTime))
                    {
                        return false;
                    }
                }
            }

            return true;
        }
        else
        {
        #if defined(AURORA_FORCE_SRW_LOCKS)
            ::AcquireSRWLockShared(&this->atomicHolder_);

            BOOL status = false;
            while (!this->TryLockNoSpin())
            {
                AuUInt32 uTimeoutMS = INFINITE;

                if (uTimeout != 0)
                {
                    auto uStartTime = Time::SteadyClockNS();
                    if (uStartTime >= uEndTime)
                    {
                        goto exitWin32;
                    }

                    uTimeoutMS = AuNSToMS<AuInt64>(uEndTime - uStartTime);
                }

                if (!uTimeoutMS)
                {
                    ::ReleaseSRWLockShared(&this->atomicHolder_);
                    SMPPause();
                    AuThreading::ContextYield();
                    ::AcquireSRWLockShared(&this->atomicHolder_);
                }
                else
                {
                    (void)SleepConditionVariableSRW(&this->wakeup_, &this->atomicHolder_, uTimeoutMS, CONDITION_VARIABLE_LOCKMODE_SHARED);
                }
            }

            returnValue = true;

        exitWin32:
            ::ReleaseSRWLockShared(&this->atomicHolder_);
        #else

            if (!uTimeout)
            {
                while (!this->TryLockNoSpin())
                {
                    auto &uValueRef = this->state_;
                    auto uValue = uValueRef | 1;

                    if (AuAtomicCompareExchange(&uValueRef, uValue + kFutexBitWait, uValue) == uValue)
                    {
                        pNtWaitForKeyedEvent(gKeyedEventHandle, (void *)&uValueRef, 0, NULL);
                        AuAtomicSub(&uValueRef, kFutexBitWake);
                    }
                }

                return true;
            }
            else
            {
                auto &uValueRef = this->state_;
                returnValue = true;

                auto uEndTimeSteady = AuTime::SteadyClockNS()  + uTimeout;
                auto uEndTimeWall   = AuTime::CurrentClockNS() + uTimeout;
                bool bFailed {};

                while (bFailed || (!this->TryLockNoSpin()))
                {
                    auto uValue = uValueRef | 1;

                    if (!bFailed &&
                        AuTime::SteadyClockNS() >= uEndTimeSteady)
                    {
                        returnValue = this->TryLock();
                        break;
                    }

                    if (bFailed || AuAtomicCompareExchange(&uValueRef, uValue + kFutexBitWait, uValue) == uValue)
                    {
                        auto uTargetTimeNt  = AuTime::ConvertTimestampNs(uEndTimeWall);

                        LARGE_INTEGER word;
                        word.QuadPart = uTargetTimeNt;

                        auto uStatus = pNtWaitForKeyedEvent(gKeyedEventHandle, (void *)&this->state_, 0, &word);
                        if (uStatus == NTSTATUS_TIMEOUT)
                        {
                            auto uWWaiters = this->state_ & ~kFutexBitWake;
                            if (uWWaiters >= kFutexBitWait && AuAtomicCompareExchange(&this->state_, uWWaiters - kFutexBitWait, uWWaiters) == uWWaiters)
                            {
                                continue;
                            }
                            else
                            {
                                bFailed = true;
                                continue;
                            }
                        }
                        else
                        {
                            AuAtomicSub(&uValueRef, kFutexBitWake);
                            SysAssertDbg(uStatus == 0);
                        }
                    }

                    bFailed = false;
                }
            }

        #endif
            return returnValue;
        }
    }

    void MutexImpl::Unlock()
    {
    #if defined(AURORA_FORCE_SRW_LOCKS)

        if (gUseNativeWaitMutex)
        {
            auto &uValueRef = this->state_;

            *(AuUInt8 *)&uValueRef = 0;

            while (true)
            {
                auto uValue = uValueRef;

                if (uValue < kFutexBitWait)
                {
                    return;
                }

                if (AuAtomicCompareExchange(&uValueRef, uValue - kFutexBitWait, uValue) == uValue)
                {
                    pWakeByAddressSingle((void *)&this->state_);
                    return;
                }

                SMPPause();
            }

            return;
        }

        ::AcquireSRWLockExclusive(&this->atomicHolder_);
        this->state_ = 0;
        ::ReleaseSRWLockExclusive(&this->atomicHolder_);
        ::WakeAllConditionVariable(&this->wakeup_);
    #else

        auto &uValueRef = this->state_;

    //#if defined(AURORA_COMPILER_MSVC)
    //    #if defined(AURORA_ARCH_X86) || defined(AURORA_ARCH_X64)
    //        //  Intel 64 and IA - 32 Architectures Software Developer's Manual, Volume 3A: Section: 8.2.3.1
    //        *(AuUInt8 *)&uValueRef = 0;
    //
    //        // From this point onwards, our thread could be subject to StoreLoad re-ordering
    //        // ...but it should not matter.
    //            
    //        // Given the memory model of x86[64], we can only really expect to be out of order during an unfenced load operation, which in this class, can only be expected under this function before the CAS.
    //        // No other place reads.
    //
    //        // Re-ordering race condition 1: one thread wins an atomic bit set, that we dont catch until the CAS, resulting in: a slow implicit fence under the cas, a mm_pause stall, a compare, and a return
    //        //                                                             alt: uValueRef reads zero, resulting in a preemptive return while no threads need to be awoken
    //        // Re-ordering race condition 2: we unlock, multiple threads enter ::Lock(), we somehow read `uValue = uValueRef` as zero, and then the first atomic bitsetandtest winner thread signals the keyed mutex
    //        // I fail to see how:
    //        // *byte = 0; |                                    |
    //        //            | interlocked atomicbitset           | interlocked atomicbitset fail 
    //        //            | [logic]                            | interlocked atomic set kFutexBitWait
    //        //            | *byte = 0;                         | yield
    //        //            | auto uValue =[acquire]= uValueRef 
    //        // ...would result in the second thread missing the third threads atomic set kFutexBitWait (cst (?) on the account of 8.2.3.1, 8.2.3.8, etc)
    //
    //        // Also note: mfence is far too expensive and the _ReadWriteBarrier() intrinsics do absolutely nothing
    //        _ReadWriteBarrier();
    //    #else
    //        InterlockedAndRelease((volatile LONG *)&uValueRef, ~0xFF);
    //    #endif
    //#else
    //    __sync_lock_release((AuUInt8 *)&uValueRef); // __atomic_store_explicit((AuUInt8 *)&uValueRef, 0, __ATOMIC_RELEASE)
    //#endif
        // merged with ROXTL
        AuAtomicClearU8Lock(&uValueRef);

        while (true)
        {
            auto uValue = uValueRef;

            if (uValue < kFutexBitWait)
            {
                return;
            }

            // StoreLoad race-conditions here cannot result in a return
            // We should see StoreLoads of at least our *pByte = 0
            // or we should at least see the CST of kFutexBitWait being applied
            if (uValue & 1)
            {
                return;
            }

            if (gUseNativeWaitMutex)
            {
                if (AuAtomicCompareExchange(&uValueRef, uValue - kFutexBitWait, uValue) == uValue)
                {
                    pWakeByAddressSingle((void *)&this->state_);
                    return;
                }
            }
            else
            {
                if (uValue & kFutexBitWake)
                {
                    // StoreLoad paranoia
                    if (AuAtomicCompareExchange(&uValueRef, uValue, uValue) == uValue)
                    {
                        return;
                    }
                    else
                    {
                        SMPPause();
                        continue;
                    }
                }

                if (AuAtomicCompareExchange(&uValueRef, uValue - kFutexBitWait + kFutexBitWake, uValue) == uValue)
                {
                    pNtReleaseKeyedEvent(gKeyedEventHandle, (void *)&uValueRef, 0, NULL);
                    return;
                }
            }

            SMPPause();
        }

    #endif
    }
    
    AUKN_SYM IHyperWaitable *MutexNew()
    {
        return _new MutexImpl();
    }

    AUKN_SYM void MutexRelease(IHyperWaitable *pMutex)
    {
        AuSafeDelete<MutexImpl *>(pMutex);
    }

    AUROXTL_INTERFACE_SOO_SRC_EX(AURORA_SYMBOL_EXPORT, Mutex, MutexImpl)
}

#endif