Add debug flag to read back command buffer pointer

Signed-off-by: Bartosz Dunajski <bartosz.dunajski@intel.com>

opencl/test/unit_test/os_interface/linux/drm_command_stream_tests_2.cpp

@@ -169,6 +169,42 @@ HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest, givenCommandStreamWithDuplicate
     mm->freeGraphicsMemory(commandBuffer);
 }
 
+HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest, givenDebugFlagSetWhenFlushingThenReadBackCommandBufferPointerIfRequired) {
+    auto commandBuffer = static_cast<DrmAllocation *>(mm->allocateGraphicsMemoryWithProperties(MockAllocationProperties{csr->getRootDeviceIndex(), MemoryConstants::pageSize}));
+    LinearStream cs(commandBuffer);
+
+    auto testedCsr = static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr);
+
+    CommandStreamReceiverHw<FamilyType>::addBatchBufferEnd(cs, nullptr);
+    EncodeNoop<FamilyType>::alignToCacheLine(cs);
+
+    BatchBuffer batchBuffer{cs.getGraphicsAllocation(), 0, 0, nullptr, false, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount, cs.getUsed(), &cs, nullptr, false};
+
+    {
+        DebugManager.flags.ReadBackCommandBufferAllocation.set(1);
+
+        csr->flush(batchBuffer, csr->getResidencyAllocations());
+
+        if (commandBuffer->isAllocatedInLocalMemoryPool()) {
+            EXPECT_EQ(commandBuffer->getUnderlyingBuffer(), testedCsr->latestReadBackAddress);
+        } else {
+            EXPECT_EQ(nullptr, testedCsr->latestReadBackAddress);
+        }
+
+        testedCsr->latestReadBackAddress = nullptr;
+    }
+
+    {
+        DebugManager.flags.ReadBackCommandBufferAllocation.set(2);
+
+        csr->flush(batchBuffer, csr->getResidencyAllocations());
+
+        EXPECT_EQ(commandBuffer->getUnderlyingBuffer(), testedCsr->latestReadBackAddress);
+    }
+
+    mm->freeGraphicsMemory(commandBuffer);
+}
+
 HWTEST_TEMPLATED_F(DrmCommandStreamEnhancedTest, givenDrmCsrCreatedWithInactiveGemCloseWorkerPolicyThenThreadIsNotCreated) {
     TestedDrmCommandStreamReceiver<FamilyType> testedCsr(gemCloseWorkerMode::gemCloseWorkerInactive,
                                                          *this->executionEnvironment,

opencl/test/unit_test/test_files/igdrcl.config

@@ -407,3 +407,4 @@ DirectSubmissionForceLocalMemoryStorageMode = -1
 EnableRingSwitchTagUpdateWa = -1
 DirectSubmissionReadBackCommandBuffer = -1
 DirectSubmissionReadBackRingBuffer = -1
+ReadBackCommandBufferAllocation = -1

shared/source/debug_settings/debug_variables_base.inl

@@ -197,6 +197,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, OverrideTimestampEvents, -1, "-1: default (based
 DECLARE_DEBUG_VARIABLE(int32_t, ForcePreParserEnabledForMiArbCheck, -1, "-1: default , 0: PreParser disabled, 1: PreParser enabled")
 DECLARE_DEBUG_VARIABLE(int32_t, BatchBufferStartPrepatchingWaEnabled, -1, "-1: default , 0: disabled, 1: enabled. WA applies valid VA pointing to 'self' instead of 0x0. This mitigates incorrect VA preparsing.")
 DECLARE_DEBUG_VARIABLE(int32_t, SetVmAdviseAtomicAttribute, -1, "-1: default - atomic system, 0: atomic none, 1: atomic device, 2: atomic system)")
+DECLARE_DEBUG_VARIABLE(int32_t, ReadBackCommandBufferAllocation, -1, "Read command buffer allocation back on the host side. -1: default, 0 - disabled, 1 - local memory only, 2 - local and system memory")
 DECLARE_DEBUG_VARIABLE(bool, DisableScratchPages, false, "Disable scratch pages during VM creations")
 /*LOGGING FLAGS*/
 DECLARE_DEBUG_VARIABLE(int32_t, PrintDriverDiagnostics, -1, "prints driver diagnostics messages to standard output, value corresponds to hint level")

shared/source/direct_submission/direct_submission_hw.h

@@ -153,7 +153,7 @@ class DirectSubmissionHw {
     uint32_t workloadModeOneExpectedValue = 0u;
     uint32_t activeTiles = 1u;
     uint32_t postSyncOffset = 0u;
-    uint32_t reserved = 0u;
+    volatile uint32_t reserved = 0u;
 
     bool ringStart = false;
     bool disableCpuCacheFlush = true;

shared/source/os_interface/linux/drm_command_stream.h

@@ -70,6 +70,7 @@ class DrmCommandStreamReceiver : public DeviceCommandStreamReceiver<GfxFamily> {
     MOCKABLE_VIRTUAL int flushInternal(const BatchBuffer &batchBuffer, const ResidencyContainer &allocationsForResidency);
     MOCKABLE_VIRTUAL int exec(const BatchBuffer &batchBuffer, uint32_t vmHandleId, uint32_t drmContextId, uint32_t index);
     MOCKABLE_VIRTUAL int waitUserFence(uint32_t waitValue);
+    MOCKABLE_VIRTUAL void readBackAllocation(void *source);
     bool isUserFenceWaitActive();
 
     std::vector<BufferObject *> residency;
@@ -77,6 +78,7 @@ class DrmCommandStreamReceiver : public DeviceCommandStreamReceiver<GfxFamily> {
     Drm *drm;
     gemCloseWorkerMode gemCloseWorkerOperationMode;
 
+    volatile uint32_t reserved = 0;
     int32_t kmdWaitTimeout = -1;
 
     bool useUserFenceWait = true;

shared/source/os_interface/linux/drm_command_stream.inl

@@ -153,6 +153,14 @@ SubmissionStatus DrmCommandStreamReceiver<GfxFamily>::flush(BatchBuffer &batchBu
     } else {
         this->flushStamp->setStamp(bb->peekHandle());
     }
+
+    auto readBackMode = DebugManager.flags.ReadBackCommandBufferAllocation.get();
+    bool readBackAllowed = ((batchBuffer.commandBufferAllocation->isAllocatedInLocalMemoryPool() && readBackMode == 1) || readBackMode == 2);
+
+    if (readBackAllowed) {
+        readBackAllocation(ptrOffset(batchBuffer.commandBufferAllocation->getUnderlyingBuffer(), batchBuffer.startOffset));
+    }
+
     auto ret = this->flushInternal(batchBuffer, allocationsForResidency);
 
     if (this->gemCloseWorkerOperationMode == gemCloseWorkerMode::gemCloseWorkerActive) {
@@ -167,6 +175,11 @@ SubmissionStatus DrmCommandStreamReceiver<GfxFamily>::flush(BatchBuffer &batchBu
     return SubmissionStatus::SUCCESS;
 }
 
+template <typename GfxFamily>
+void DrmCommandStreamReceiver<GfxFamily>::readBackAllocation(void *source) {
+    reserved = *reinterpret_cast<volatile uint32_t *>(source);
+}
+
 template <typename GfxFamily>
 void DrmCommandStreamReceiver<GfxFamily>::printBOsForSubmit(ResidencyContainer &allocationsForResidency, GraphicsAllocation &cmdBufferAllocation) {
     if (DebugManager.flags.PrintBOsForSubmit.get()) {

shared/test/common/mocks/linux/mock_drm_command_stream_receiver.h

@@ -106,14 +106,20 @@ class TestedDrmCommandStreamReceiver : public DrmCommandStreamReceiver<GfxFamily
         }
     }
 
-    bool callHwFlush = true;
-
     int flushInternal(const BatchBuffer &batchBuffer, const ResidencyContainer &allocationsForResidency) override {
         if (callHwFlush) {
             return DrmCommandStreamReceiver<GfxFamily>::flushInternal(batchBuffer, allocationsForResidency);
         }
         return 0;
     }
+
+    void readBackAllocation(void *source) override {
+        latestReadBackAddress = source;
+        DrmCommandStreamReceiver<GfxFamily>::readBackAllocation(source);
+    }
+
+    void *latestReadBackAddress = nullptr;
+    bool callHwFlush = true;
 };
 
 template <typename GfxFamily>