Use correct virtual addresses in TBX CSR makeCoherent method

- cpu virtual address was used instead of gpu va
- this caused incorrect behaviour of TBX server when
special heap allocator assigning GPU addresses was used

Change-Id: I2328cf2441be797311fd6a3c7b331b0fff79d4fc

runtime/command_queue/command_queue.cpp

@@ -142,7 +142,9 @@ volatile uint32_t *CommandQueue::getHwTagAddress() const {
     DEBUG_BREAK_IF(!this->device);
     auto &commandStreamReceiver = device->getCommandStreamReceiver();
     auto tag_address = commandStreamReceiver.getTagAddress();
-    commandStreamReceiver.makeCoherent((void *)tag_address, sizeof(tag_address));
+    auto allocation = commandStreamReceiver.getTagAllocation();
+    UNRECOVERABLE_IF(allocation == nullptr);
+    commandStreamReceiver.makeCoherent(*allocation);
     return tag_address;
 }
 

runtime/command_stream/command_stream_receiver.cpp

@@ -66,7 +66,7 @@ void CommandStreamReceiver::processEviction() {
 
 void CommandStreamReceiver::makeNonResident(GraphicsAllocation &gfxAllocation) {
     if (gfxAllocation.residencyTaskCount != ObjectNotResident) {
-        makeCoherent(gfxAllocation.getUnderlyingBuffer(), gfxAllocation.getUnderlyingBufferSize());
+        makeCoherent(gfxAllocation);
         getMemoryManager()->pushAllocationForEviction(&gfxAllocation);
     }
 

runtime/command_stream/command_stream_receiver.h

@@ -65,7 +65,7 @@ class CommandStreamReceiver {
 
     virtual void flushBatchedSubmissions() = 0;
 
-    virtual void makeCoherent(void *address, size_t length){};
+    virtual void makeCoherent(GraphicsAllocation &gfxAllocation){};
     virtual void makeResident(GraphicsAllocation &gfxAllocation);
     virtual void makeNonResident(GraphicsAllocation &gfxAllocation);
     void makeSurfacePackNonResident(ResidencyContainer *allocationsForResidency);
@@ -86,6 +86,9 @@ class CommandStreamReceiver {
     OSInterface *getOSInterface() { return osInterface.get(); };
 
     MOCKABLE_VIRTUAL void setTagAllocation(GraphicsAllocation *allocation);
+    GraphicsAllocation *getTagAllocation() const {
+        return tagAllocation;
+    }
     volatile uint32_t *getTagAddress() const { return tagAddress; }
 
     virtual bool waitForFlushStamp(FlushStamp &flushStampToWait) { return true; };

runtime/command_stream/tbx_command_stream_receiver_hw.h

@@ -46,7 +46,7 @@ class TbxCommandStreamReceiverHw : public CommandStreamReceiverHw<GfxFamily> {
 
   public:
     FlushStamp flush(BatchBuffer &batchBuffer, EngineType engineType, ResidencyContainer *allocationsForResidency) override;
-    void makeCoherent(void *address, size_t length) override;
+    void makeCoherent(GraphicsAllocation &gfxAllocation) override;
 
     void processResidency(ResidencyContainer *allocationsForResidency) override;
     bool writeMemory(GraphicsAllocation &gfxAllocation);

runtime/command_stream/tbx_command_stream_receiver_hw.inl

@@ -361,14 +361,18 @@ void TbxCommandStreamReceiverHw<GfxFamily>::processResidency(ResidencyContainer
 }
 
 template <typename GfxFamily>
-void TbxCommandStreamReceiverHw<GfxFamily>::makeCoherent(void *address, size_t length) {
+void TbxCommandStreamReceiverHw<GfxFamily>::makeCoherent(GraphicsAllocation &gfxAllocation) {
+    auto cpuAddress = gfxAllocation.getUnderlyingBuffer();
+    auto gpuAddress = gfxAllocation.getGpuAddress();
+    auto length = gfxAllocation.getUnderlyingBufferSize();
+
     if (length) {
         PageWalker walker = [&](uint64_t physAddress, size_t size, size_t offset) {
             DEBUG_BREAK_IF(offset > length);
-            stream.readMemory(physAddress, ptrOffset(address, offset), size);
+            stream.readMemory(physAddress, ptrOffset(cpuAddress, offset), size);
         };
 
-        ppgtt.pageWalk(reinterpret_cast<uintptr_t>(address), length, 0, walker);
+        ppgtt.pageWalk(static_cast<uintptr_t>(gpuAddress), length, 0, walker);
     }
 }
 } // namespace OCLRT

unit_tests/command_stream/command_stream_receiver_tests.cpp

@@ -257,3 +257,15 @@ HWTEST_F(CommandStreamReceiverTest, givenDefaultCommandStreamReceiverThenDefault
     auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
     EXPECT_EQ(CommandStreamReceiver::DispatchMode::ImmediateDispatch, csr.dispatchMode);
 }
+
+TEST(CommandStreamReceiverSimpleTest, givenCSRWithoutTagAllocationWhenGetTagAllocationIsCalledThenNullptrIsReturned) {
+    MockCommandStreamReceiver csr;
+    EXPECT_EQ(nullptr, csr.getTagAllocation());
+}
+
+TEST(CommandStreamReceiverSimpleTest, givenCSRWithTagAllocationSetWhenGetTagAllocationIsCalledThenCorrectAllocationIsReturned) {
+    MockCommandStreamReceiver csr;
+    GraphicsAllocation allocation(reinterpret_cast<void *>(0x1000), 0x1000);
+    csr.setTagAllocation(&allocation);
+    EXPECT_EQ(&allocation, csr.getTagAllocation());
+}