[NativeAOT] Follow up changes for the suspension loop routine. (#73741)

* add timing

* do not hijack redirected threads

* tweak the suspend loop implementation

* remove instrumentation

Author: VSadov

src/coreclr/nativeaot/Runtime/StackFrameIterator.cpp

@@ -95,7 +95,7 @@ GVAL_IMPL_INIT(PTR_VOID, g_RhpRethrow2Addr, PointerToRhpRethrow2);
 StackFrameIterator::StackFrameIterator(Thread * pThreadToWalk, PInvokeTransitionFrame* pInitialTransitionFrame)
 {
     STRESS_LOG0(LF_STACKWALK, LL_INFO10000, "----Init---- [ GC ]\n");
-    ASSERT(!pThreadToWalk->DangerousCrossThreadIsHijacked());
+    ASSERT(!pThreadToWalk->IsHijacked());
 
     if (pInitialTransitionFrame == INTERRUPTED_THREAD_MARKER)
     {
@@ -1463,7 +1463,7 @@ void StackFrameIterator::NextInternal()
     else
     {
         // if the thread is safe to walk, it better not have a hijack in place.
-        ASSERT((ThreadStore::GetCurrentThread() == m_pThread) || !m_pThread->DangerousCrossThreadIsHijacked());
+        ASSERT(!m_pThread->IsHijacked());
 
         SetControlPC(dac_cast<PTR_VOID>(*(m_RegDisplay.GetAddrOfIP())));
 

src/coreclr/nativeaot/Runtime/thread.cpp

@@ -77,6 +77,10 @@ void Thread::WaitForGC(PInvokeTransitionFrame* pTransitionFrame)
         // set preemptive mode
         VolatileStoreWithoutBarrier(&m_pTransitionFrame, pTransitionFrame);
 
+#ifdef FEATURE_SUSPEND_REDIRECTION
+        ClearState(TSF_Redirected);
+#endif //FEATURE_SUSPEND_REDIRECTION
+
         RedhawkGCInterface::WaitForGCCompletion();
 
         // must be in cooperative mode when checking the trap flag
@@ -587,6 +591,14 @@ void Thread::Hijack()
         return;
     }
 
+#ifdef FEATURE_SUSPEND_REDIRECTION
+    // if the thread is redirected, leave it as-is.
+    if (IsStateSet(TSF_Redirected))
+    {
+        return;
+    }
+#endif //FEATURE_SUSPEND_REDIRECTION
+
     // PalHijack will call HijackCallback or make the target thread call it.
     // It may also do nothing if the target thread is in inconvenient state.
     PalHijack(m_hPalThread, this);
@@ -806,6 +818,8 @@ bool Thread::Redirect()
     if (!PalSetThreadContext(m_hPalThread, redirectionContext))
         return false;
 
+    // the thread will now inevitably try to suspend
+    SetState(TSF_Redirected);
     redirectionContext->SetIp(origIP);
 
     STRESS_LOG2(LF_STACKWALK, LL_INFO10000, "InternalRedirect: TgtThread = %llx, IP = %p\n",
@@ -887,55 +901,16 @@ void Thread::UnhijackWorker()
     m_uHijackedReturnValueFlags         = 0;
 }
 
-// @TODO: it would be very, very nice if we did not have to bleed knowledge of hijacking
-// and hijack state to other components in the runtime. For now, these are only used
-// when getting EH info during exception dispatch. We should find a better way to encapsulate
-// this.
 bool Thread::IsHijacked()
 {
-    // Note: this operation is only valid from the current thread. If one thread invokes
-    // this on another then it may be racing with other changes to the thread's hijack state.
-    ASSERT(ThreadStore::GetCurrentThread() == this);
-
-    return m_pvHijackedReturnAddress != NULL;
-}
+    ASSERT(((ThreadStore::GetCurrentThread() == this) && IsCurrentThreadInCooperativeMode()) ||
+        ThreadStore::GetCurrentThread()->IsGCSpecial() ||
+        ThreadStore::GetCurrentThread() == ThreadStore::GetSuspendingThread()
+    );
 
-//
-// WARNING: This method must ONLY be called during stackwalks when we believe that all threads are
-// synchronized and there is no other thread racing with us trying to apply hijacks.
-//
-bool Thread::DangerousCrossThreadIsHijacked()
-{
-    // If we have a CachedTransitionFrame available, then we're in the proper state.  Otherwise, this method
-    // was called from an improper state.
-    ASSERT(GetTransitionFrame() != NULL);
     return m_pvHijackedReturnAddress != NULL;
 }
 
-void * Thread::GetHijackedReturnAddress()
-{
-    // Note: this operation is only valid from the current thread. If one thread invokes
-    // this on another then it may be racing with other changes to the thread's hijack state.
-    ASSERT(IsHijacked());
-    ASSERT(ThreadStore::GetCurrentThread() == this);
-
-    return m_pvHijackedReturnAddress;
-}
-
-void * Thread::GetUnhijackedReturnAddress(void ** ppvReturnAddressLocation)
-{
-    ASSERT(ThreadStore::GetCurrentThread() == this);
-
-    void * pvReturnAddress;
-    if (m_ppvHijackedReturnAddressLocation == ppvReturnAddressLocation)
-        pvReturnAddress = m_pvHijackedReturnAddress;
-    else
-        pvReturnAddress = *ppvReturnAddressLocation;
-
-    ASSERT(GetRuntimeInstance()->IsManaged(pvReturnAddress));
-    return pvReturnAddress;
-}
-
 void Thread::SetState(ThreadStateFlags flags)
 {
     PalInterlockedOr(&m_ThreadStateFlags, flags);

src/coreclr/nativeaot/Runtime/thread.h

@@ -136,6 +136,12 @@ class Thread : private ThreadBuffer
 #ifdef FEATURE_GC_STRESS
         TSF_IsRandSeedSet       = 0x00000040,       // set to indicate the random number generator for GCStress was inited
 #endif // FEATURE_GC_STRESS
+
+#ifdef FEATURE_SUSPEND_REDIRECTION
+        TSF_Redirected          = 0x00000080,       // Set to indicate the thread is redirected and will inevitably
+                                                    // suspend once resumed.
+                                                    // If we see this flag, we skip hijacking as an optimization.
+#endif //FEATURE_SUSPEND_REDIRECTION
     };
 private:
 
@@ -197,13 +203,11 @@ class Thread : private ThreadBuffer
 
     void                Hijack();
     void                Unhijack();
+    bool                IsHijacked();
+
 #ifdef FEATURE_GC_STRESS
     static void         HijackForGcStress(PAL_LIMITED_CONTEXT * pSuspendCtx);
 #endif // FEATURE_GC_STRESS
-    bool                IsHijacked();
-    void *              GetHijackedReturnAddress();
-    void *              GetUnhijackedReturnAddress(void** ppvReturnAddressLocation);
-    bool                DangerousCrossThreadIsHijacked();
 
     bool                IsSuppressGcStressSet();
     void                SetSuppressGcStress();
@@ -230,7 +234,7 @@ class Thread : private ThreadBuffer
 #endif // DACCESS_COMPILE
 #ifdef FEATURE_GC_STRESS
     void                SetRandomSeed(uint32_t seed);
-    uint32_t              NextRand();
+    uint32_t            NextRand();
     bool                IsRandInited();
 #endif // FEATURE_GC_STRESS
     PTR_ExInfo          GetCurExInfo();

src/coreclr/nativeaot/Runtime/threadstore.cpp

@@ -196,6 +196,35 @@ void ThreadStore::UnlockThreadStore()
     m_Lock.ReleaseReadLock();
 }
 
+// exponential spinwait with an approximate time limit for waiting in microsecond range.
+// when iteration == -1, only usecLimit is used
+void SpinWait(int iteration, int usecLimit)
+{
+    LARGE_INTEGER li;
+    PalQueryPerformanceCounter(&li);
+    int64_t startTicks = li.QuadPart;
+
+    PalQueryPerformanceFrequency(&li);
+    int64_t ticksPerSecond = li.QuadPart;
+    int64_t endTicks = startTicks + (usecLimit * ticksPerSecond) / 1000000;
+
+    int l = min((unsigned)iteration, 30);
+    for (int i = 0; i < l; i++)
+    {
+        for (int j = 0; j < (1 << i); j++)
+        {
+            System_YieldProcessor();
+        }
+
+        PalQueryPerformanceCounter(&li);
+        int64_t currentTicks = li.QuadPart;
+        if (currentTicks > endTicks)
+        {
+            break;
+        }
+    }
+}
+
 void ThreadStore::SuspendAllThreads(bool waitForGCEvent)
 {
     Thread * pThisThread = GetCurrentThreadIfAvailable();
@@ -216,50 +245,57 @@ void ThreadStore::SuspendAllThreads(bool waitForGCEvent)
     // reason for this is that we essentially implement Dekker's algorithm, which requires write ordering.
     PalFlushProcessWriteBuffers();
 
-    bool keepWaiting;
-    YieldProcessorNormalizationInfo normalizationInfo;
-    int waitCycles = 1;
-    do
+    int retries = 0;
+    int prevRemaining = 0;
+    int remaining = 0;
+    bool observeOnly = false;
+
+    while(true)
     {
-        keepWaiting = false;
+        prevRemaining = remaining;
+        remaining = 0;
+
         FOREACH_THREAD(pTargetThread)
         {
             if (pTargetThread == pThisThread)
                 continue;
 
             if (!pTargetThread->CacheTransitionFrameForSuspend())
             {
-                // We drive all threads to preemptive mode by hijacking them with return-address hijack.
-                keepWaiting = true;
-                pTargetThread->Hijack();
+                remaining++;
+                if (!observeOnly)
+                {
+                    pTargetThread->Hijack();
+                }
             }
         }
         END_FOREACH_THREAD
 
-        if (keepWaiting)
+        if (!remaining)
+            break;
+
+        // if we see progress or have just done a hijacking pass
+        // do not hijack in the next iteration
+        if (remaining < prevRemaining || !observeOnly)
         {
-            if (PalSwitchToThread() == 0 && g_RhNumberOfProcessors > 1)
+            // 5 usec delay, then check for more progress
+            SpinWait(-1, 5);
+            observeOnly = true;
+        }
+        else
+        {
+            SpinWait(retries++, 100);
+            observeOnly = false;
+
+            // make sure our spining is not starving other threads, but not too often,
+            // this can cause a 1-15 msec delay, depending on OS, and that is a lot while
+            // very rarely needed, since threads are supposed to be releasing their CPUs
+            if ((retries & 127) == 0)
             {
-                // No threads are scheduled on this processor.  Perhaps we're waiting for a thread
-                // that's scheduled on another processor.  If so, let's give it a little time
-                // to make forward progress.
-                // Note that we do not call Sleep, because the minimum granularity of Sleep is much
-                // too long (we probably don't need a 15ms wait here).  Instead, we'll just burn some
-                // cycles.
-    	        // @TODO: need tuning for spin
-                // @TODO: need tuning for this whole loop as well.
-                //        we are likley too aggressive with interruptions which may result in longer pauses.
-                YieldProcessorNormalizedForPreSkylakeCount(normalizationInfo, waitCycles);
-
-                // simplistic linear backoff for now
-                // we could be catching threads in restartable sequences such as LL/SC style interlocked on ARM64
-                // and forcing them to restart.
-                // if interrupt mechanism is fast, eagerness could be hurting our overall progress.
-                waitCycles += 10000;
+                PalSwitchToThread();
             }
         }
-
-    } while (keepWaiting);
+    }
 
 #if defined(TARGET_ARM) || defined(TARGET_ARM64)
     // Flush the store buffers on all CPUs, to ensure that all changes made so far are seen