From 7357764cfcb0b75ff3714eb744d8908dab80e903 Mon Sep 17 00:00:00 2001
From: J Reece Wilson <me@reece.sx>
Date: Mon, 18 Sep 2023 18:21:46 +0100
Subject: [PATCH] [*] Fix abnormal ::UnlockWrite performance under heavyweight
 native WaitOnAddress platforms (Linux, BSD-like, etc)

---
 Source/Threading/Primitives/AuRWLock.cpp | 76 +++++++++++++++---------
 Source/Threading/Primitives/AuRWLock.hpp |  1 +
 2 files changed, 48 insertions(+), 29 deletions(-)

diff --git a/Source/Threading/Primitives/AuRWLock.cpp b/Source/Threading/Primitives/AuRWLock.cpp
index 88e94c44..a4ebb80d 100644
--- a/Source/Threading/Primitives/AuRWLock.cpp
+++ b/Source/Threading/Primitives/AuRWLock.cpp
@@ -150,15 +150,25 @@ namespace Aurora::Threading::Primitives
         return (AuUInt32 *)this->conditionVariableWriter_;
     }
 
+    template<bool bIsWriteRecursionAllowed>
+    AuUInt32 *RWLockImpl<bIsWriteRecursionAllowed>::GetReadSleepCounter()
+    {
+        return (AuUInt32 *)this->conditionVariable_;
+    }
+
     template<bool bIsWriteRecursionAllowed>
     bool RWLockImpl<bIsWriteRecursionAllowed>::LockReadNSAbs(AuUInt64 uTimeout)
     {
+        AuInt32 iCurState {};
+        bool bRet {};
+
         if (this->TryLockRead())
         {
             return true;
         }
 
-        AuInt32 iCurState {};
+        auto pCounter = this->GetReadSleepCounter();
+
         do
         {
             iCurState = this->state_;
@@ -167,7 +177,11 @@ namespace Aurora::Threading::Primitives
             {
                 if (gUseFutexRWLock)
                 {
-                    if (!InternalLTSWaitOnAddressHighRes((const void *)&this->state_, &iCurState, sizeof(iCurState), uTimeout))
+                    AuAtomicAdd(pCounter, 1u);
+                    bRet = InternalLTSWaitOnAddressHighRes((const void *)&this->state_, &iCurState, sizeof(iCurState), uTimeout);
+                    AuAtomicSub(pCounter, 1u);
+
+                    if (!bRet)
                     {
                         return false;
                     }
@@ -212,6 +226,9 @@ namespace Aurora::Threading::Primitives
     template<bool bIsWriteRecursionAllowed>
     bool RWLockImpl<bIsWriteRecursionAllowed>::LockReadNS(AuUInt64 uTimeout)
     {
+        AuInt32 iCurState {};
+        bool bRet {};
+
         if (this->TryLockReadNoSpin<true>())
         {
             return true;
@@ -224,7 +241,8 @@ namespace Aurora::Threading::Primitives
             return true;
         }
 
-        AuInt32 iCurState {};
+        auto pCounter = this->GetReadSleepCounter();
+
         do
         {
             iCurState = this->state_;
@@ -233,7 +251,11 @@ namespace Aurora::Threading::Primitives
             {
                 if (gUseFutexRWLock)
                 {
-                    if (!InternalLTSWaitOnAddressHighRes((const void *)&this->state_, &iCurState, sizeof(iCurState), uEndTime))
+                    AuAtomicAdd(pCounter, 1u);
+                    bRet = InternalLTSWaitOnAddressHighRes((const void *)&this->state_, &iCurState, sizeof(iCurState), uEndTime);
+                    AuAtomicSub(pCounter, 1u);
+
+                    if (!bRet)
                     {
                         return false;
                     }
@@ -331,15 +353,20 @@ namespace Aurora::Threading::Primitives
     bool RWLockImpl<bIsWriteRecursionAllowed>::LockWriteNSAbsSecondPath()
     { 
         auto uOld = this->state_;
-        if (uOld < 0)
+
+        if constexpr (bIsWriteRecursionAllowed)
         {
-            if (this->reentrantWriteLockHandle_ == GetThreadCookie())
+            if (uOld < 0)
             {
-                AuAtomicSub(&this->state_, 1);
-                return true;
+                if (this->reentrantWriteLockHandle_ == GetThreadCookie())
+                {
+                    AuAtomicSub(&this->state_, 1);
+                    return true;
+                }
             }
         }
-        else if (uOld == 0)
+        
+        if (uOld == 0)
         {
             if (AuAtomicCompareExchange(&this->state_, -1, uOld) == uOld)
             {
@@ -354,6 +381,9 @@ namespace Aurora::Threading::Primitives
     template<bool bIsWriteRecursionAllowed>
     bool RWLockImpl<bIsWriteRecursionAllowed>::LockWriteNS(AuUInt64 uTimeout)
     {
+        bool bRet {};
+        AuInt64 uEndTime {};
+
         if constexpr (!bIsWriteRecursionAllowed)
         {
             if (this->TryLockWrite())
@@ -363,27 +393,12 @@ namespace Aurora::Threading::Primitives
         }
         else
         {
-            auto uOld = AuAtomicLoad(&this->state_);
-            if (uOld < 0)
+            if (this->LockWriteNSAbsSecondPath())
             {
-                if (this->reentrantWriteLockHandle_ == GetThreadCookie())
-                {
-                    AuAtomicSub(&this->state_, 1);
-                    return true;
-                }
-            }
-            else if (uOld == 0)
-            {
-                if (AuAtomicCompareExchange(&this->state_, -1, uOld) == uOld)
-                {
-                    this->reentrantWriteLockHandle_ = GetThreadCookie();
-                    return true;
-                }
+                return true;
             }
         }
 
-        bool bRet {};
-        AuInt64 uEndTime {};
         AuAtomicAdd(&this->writersPending_, 1);
 
         if (gUseFutexRWLock)
@@ -411,17 +426,17 @@ namespace Aurora::Threading::Primitives
     {
         while (true)
         {
+            bool bStatus {};
             AuInt32 iCurState;
+            
             while ((iCurState = AuAtomicLoad(&this->state_)) != 0)
             {
                 AuInt64 uSecondTimeout = 0;
 
-                bool bStatus {};
                 if (gUseFutexRWLock)
                 {
                     auto pSemaphore = this->GetFutexConditionWriter();
 
-                    AuInt32 iCurState;
                     while ((iCurState = AuAtomicLoad(&this->state_)) != 0)
                     {
                         static const AuUInt32 kExpect { 0 };
@@ -532,7 +547,10 @@ namespace Aurora::Threading::Primitives
     {
         if (gUseFutexRWLock)
         {
-            InternalLTSWakeAll((const void *)&this->state_);
+            if (AuAtomicLoad(this->GetReadSleepCounter()))
+            {
+                InternalLTSWakeAll((const void *)&this->state_);
+            }
         }
         else
         {
diff --git a/Source/Threading/Primitives/AuRWLock.hpp b/Source/Threading/Primitives/AuRWLock.hpp
index 08a5e057..0a92c161 100644
--- a/Source/Threading/Primitives/AuRWLock.hpp
+++ b/Source/Threading/Primitives/AuRWLock.hpp
@@ -97,6 +97,7 @@ namespace Aurora::Threading::Primitives
         auline void SignalManyReader();
         auline void SignalManyWriter(int iBias = 0);
         
+        auline AuUInt32 *GetReadSleepCounter();
     private:
 
         RWLockAccessView<true, RWLockImpl>  read_;