Fix RecyclerVisited related crashes during marking

lib/Common/Core/FinalizableObject.h

@@ -15,7 +15,7 @@ class FinalizableObject : public IRecyclerVisitedObject
         Mark(static_cast<Recycler*>(recycler));
     }
 
-    void Trace(IRecyclerHeapMarkingContext* markingContext) final
+    void Trace(IRecyclerHeapMarkingContext* markingContext)
     {
         AssertMsg(false, "Trace called on object that isn't implemented by the host");
     }

lib/Common/Memory/LargeHeapBlock.cpp

@@ -553,6 +553,16 @@ LargeHeapBlock::AllocFreeListEntry(DECLSPEC_GUARD_OVERFLOW size_t size, ObjectIn
 #ifdef RECYCLER_WRITE_BARRIER
     header->hasWriteBarrier = (attributes & WithBarrierBit) == WithBarrierBit;
 #endif
+
+    if ((attributes & (FinalizeBit | TrackBit)) != 0)
+    {
+        // Make sure a valid vtable is installed as once the attributes have been set this allocation may be traced by background marking
+        allocObject = (char *)new (allocObject) DummyVTableObject();
+#if defined(_M_ARM32_OR_ARM64)
+        // On ARM, make sure the v-table write is performed before setting the attributes
+        MemoryBarrier();
+#endif
+    }
     header->SetAttributes(this->heapInfo->recycler->Cookie, (attributes & StoredObjectInfoBitMask));
     header->markOnOOMRescan = false;
     header->SetNext(this->heapInfo->recycler->Cookie, nullptr);
@@ -618,6 +628,15 @@ LargeHeapBlock::Alloc(DECLSPEC_GUARD_OVERFLOW size_t size, ObjectInfoBits attrib
 #ifdef RECYCLER_WRITE_BARRIER
     header->hasWriteBarrier = (attributes&WithBarrierBit) == WithBarrierBit;
 #endif
+    if ((attributes & (FinalizeBit | TrackBit)) != 0)
+    {
+        // Make sure a valid vtable is installed as once the attributes have been set this allocation may be traced by background marking
+        allocObject = (char *)new (allocObject) DummyVTableObject();
+#if defined(_M_ARM32_OR_ARM64)
+        // On ARM, make sure the v-table write is performed before setting the attributes
+        MemoryBarrier();
+#endif
+    }
     header->SetAttributes(recycler->Cookie, (attributes & StoredObjectInfoBitMask));
     HeaderList()[allocCount++] = header;
     finalizeCount += ((attributes & FinalizeBit) != 0);

lib/Common/Memory/MarkContext.inl

@@ -205,42 +205,47 @@ void MarkContext::ProcessMark()
     while (!markStack.IsEmpty() || !preciseStack.IsEmpty())
 #endif
     {
+        // It is possible that when the stacks were split, only one of them had any chunks to process.
+        // If that is the case, one of the stacks might not be initialized, so we must check !IsEmpty before popping.
+        if (!markStack.IsEmpty())
+        {
 #if defined(_M_IX86) || defined(_M_X64)
-        MarkCandidate current, next;
+            MarkCandidate current, next;
 
-        while (markStack.Pop(&current))
-        {
-            // Process entries and prefetch as we go.
-            while (markStack.Pop(&next))
+            while (markStack.Pop(&current))
             {
-                // Prefetch the next entry so it's ready when we need it.
-                _mm_prefetch((char *)next.obj, _MM_HINT_T0);
+                // Process entries and prefetch as we go.
+                while (markStack.Pop(&next))
+                {
+                    // Prefetch the next entry so it's ready when we need it.
+                    _mm_prefetch((char *)next.obj, _MM_HINT_T0);
 
-                // Process the previously retrieved entry.
-                ScanObject<parallel, interior>(current.obj, current.byteCount);
+                    // Process the previously retrieved entry.
+                    ScanObject<parallel, interior>(current.obj, current.byteCount);
 
-                _mm_prefetch((char *)*(next.obj), _MM_HINT_T0);
+                    _mm_prefetch((char *)*(next.obj), _MM_HINT_T0);
 
-                current = next;
-            }
+                    current = next;
+                }
 
-            // The stack is empty, but we still have a previously retrieved entry; process it now.
-            ScanObject<parallel, interior>(current.obj, current.byteCount);
+                // The stack is empty, but we still have a previously retrieved entry; process it now.
+                ScanObject<parallel, interior>(current.obj, current.byteCount);
 
-            // Processing that entry may have generated more entries in the mark stack, so continue the loop.
-        }
+                // Processing that entry may have generated more entries in the mark stack, so continue the loop.
+            }
 #else
-        // _mm_prefetch intrinsic is specific to Intel platforms.
-        // CONSIDER: There does seem to be a compiler intrinsic for prefetch on ARM,
-        // however, the information on this is scarce, so for now just don't do prefetch on ARM.
-        MarkCandidate current;
+            // _mm_prefetch intrinsic is specific to Intel platforms.
+            // CONSIDER: There does seem to be a compiler intrinsic for prefetch on ARM,
+            // however, the information on this is scarce, so for now just don't do prefetch on ARM.
+            MarkCandidate current;
 
-        while (markStack.Pop(&current))
-        {
-            ScanObject<parallel, interior>(current.obj, current.byteCount);
-        }
+            while (markStack.Pop(&current))
+            {
+                ScanObject<parallel, interior>(current.obj, current.byteCount);
+            }
 #endif
-
+        }
+
         Assert(markStack.IsEmpty());
 
 #ifdef RECYCLER_VISITED_HOST

lib/Common/Memory/Recycler.cpp

@@ -7300,6 +7300,21 @@ Recycler::FillCheckPad(void * address, size_t size, size_t alignedAllocSize, boo
             addressToVerify = ((char*) address + size);
             sizeToVerify = (alignedAllocSize - size);
         }
+        else
+        {
+            // It could be the case that an uninitialized object already has a dummy vtable installed
+            // at the beginning of the address. If that is the case, we can't verify the fill pattern
+            // on that memory, since it's already been initialized.
+            // Note that FillPadNoCheck will skip over the first sizeof(FreeObject) bytes, which
+            // prevents overwriting of the vtable.
+            static_assert(sizeof(DummyVTableObject) == sizeof(void*), "Incorrect size for a DummyVTableObject - it must contain a single v-table pointer");
+            DummyVTableObject dummy;
+            if ((*(void**)(&dummy)) == *((void**)address))
+            {
+                addressToVerify = (char*)address + sizeof(DummyVTableObject);
+                sizeToVerify = alignedAllocSize - sizeof(DummyVTableObject);
+            }
+        }
 
         // Actually this is filling the non-pad to zero
         VerifyCheckFill(addressToVerify, sizeToVerify - sizeof(size_t));

lib/Common/Memory/Recycler.inl

@@ -36,9 +36,10 @@ namespace Memory
 class DummyVTableObject : public FinalizableObject
 {
 public:
-    virtual void Finalize(bool isShutdown) {}
-    virtual void Dispose(bool isShutdown) {}
-    virtual void Mark(Recycler * recycler) {}
+    virtual void Finalize(bool isShutdown) final {}
+    virtual void Dispose(bool isShutdown) final {}
+    virtual void Mark(Recycler * recycler) final {}
+    virtual void Trace(IRecyclerHeapMarkingContext* markingContext) final {}
 };
 }
 
@@ -171,13 +172,6 @@ Recycler::AllocWithAttributesInlined(DECLSPEC_GUARD_OVERFLOW size_t size)
 #endif
 
 
-#pragma prefast(suppress:6313, "attributes is a template parameter and can be 0")
-    if (attributes & (FinalizeBit | TrackBit))
-    {
-        // Make sure a valid vtable is installed in case of OOM before the real vtable is set
-        memBlock = (char *)new (memBlock) DummyVTableObject();
-    }
-
 #ifdef RECYCLER_WRITE_BARRIER
     SwbVerboseTrace(this->GetRecyclerFlagsTable(), _u("Allocated SWB memory: 0x%p\n"), memBlock);
 

lib/Common/Memory/SmallHeapBlockAllocator.h

@@ -129,6 +129,15 @@ SmallHeapBlockAllocator<TBlockType>::InlinedAllocImpl(Recycler * recycler, DECLS
 
         if (NeedSetAttributes(attributes))
         {
+            if ((attributes & (FinalizeBit | TrackBit)) != 0)
+            {
+                // Make sure a valid vtable is installed as once the attributes have been set this allocation may be traced by background marking
+                memBlock = (char *)new (memBlock) DummyVTableObject();
+#if defined(_M_ARM32_OR_ARM64)
+                // On ARM, make sure the v-table write is performed before setting the attributes
+                MemoryBarrier();
+#endif
+            }
             heapBlock->SetAttributes(memBlock, (attributes & StoredObjectInfoBitMask));
         }
 
@@ -138,6 +147,11 @@ SmallHeapBlockAllocator<TBlockType>::InlinedAllocImpl(Recycler * recycler, DECLS
     if (memBlock != nullptr && endAddress == nullptr)
     {
         // Free list allocation
+        freeObjectList = ((FreeObject *)memBlock)->GetNext();
+#ifdef RECYCLER_MEMORY_VERIFY
+        ((FreeObject *)memBlock)->DebugFillNext();
+#endif
+
         Assert(!this->IsBumpAllocMode());
         if (NeedSetAttributes(attributes))
         {
@@ -149,13 +163,20 @@ SmallHeapBlockAllocator<TBlockType>::InlinedAllocImpl(Recycler * recycler, DECLS
                 Assert(allocationHeapBlock != nullptr);
                 Assert(!allocationHeapBlock->IsLargeHeapBlock());
             }
+
+            if ((attributes & (FinalizeBit | TrackBit)) != 0)
+            {
+                // Make sure a valid vtable is installed as once the attributes have been set this allocation may be traced by background marking
+                memBlock = (char *)new (memBlock) DummyVTableObject();
+#if defined(_M_ARM32_OR_ARM64)
+                // On ARM, make sure the v-table write is performed before setting the attributes
+                MemoryBarrier();
+#endif
+            }
             allocationHeapBlock->SetAttributes(memBlock, (attributes & StoredObjectInfoBitMask));
         }
-        freeObjectList = ((FreeObject *)memBlock)->GetNext();
 
 #ifdef RECYCLER_MEMORY_VERIFY
-        ((FreeObject *)memBlock)->DebugFillNext();
-
         if (this->IsExplicitFreeObjectListAllocMode())
         {
             HeapBlock* heapBlock = recycler->FindHeapBlock(memBlock);