fix: only increase fence/task count when submit task success

Related-To: NEO-7812

Signed-off-by: Pan Zhenjie <zhenjie.pan@intel.com>

shared/source/command_stream/command_stream_receiver.cpp

@@ -430,7 +430,9 @@ WaitStatus CommandStreamReceiver::baseWaitFunction(volatile TagAddressType *poll
 
     TaskCountType latestSentTaskCount = this->latestFlushedTaskCount;
     if (latestSentTaskCount < taskCountToWait) {
-        this->flushTagUpdate();
+        if (this->flushTagUpdate() != NEO::SubmissionStatus::SUCCESS) {
+            return WaitStatus::NotReady;
+        }
     }
     volatile TagAddressType *partitionAddress = pollAddress;
 

shared/source/command_stream/command_stream_receiver_hw.h

@@ -191,6 +191,7 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
     void handleFrontEndStateTransition(const DispatchFlags &dispatchFlags);
     void handlePipelineSelectStateTransition(const DispatchFlags &dispatchFlags);
     void handleStateBaseAddressStateTransition(const DispatchFlags &dispatchFlags, bool &isStateBaseAddressDirty);
+    void updateStreamTaskCount(LinearStream &stream, TaskCountType newTaskCount);
 
     bool bcsRelaxedOrderingAllowed(const BlitPropertiesContainer &blitPropertiesContainer, bool hasStallingCmds) const;
 

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -241,11 +241,11 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushBcsTask(LinearStream &c
                             streamToSubmit.getUsed(), &streamToSubmit, bbEndLocation, this->getNumClients(), (submitCSR || dispatchBcsFlags.hasStallingCmds),
                             dispatchBcsFlags.hasRelaxedOrderingDependencies};
 
-    streamToSubmit.getGraphicsAllocation()->updateTaskCount(this->taskCount + 1, this->osContext->getContextId());
+    updateStreamTaskCount(streamToSubmit, taskCount + 1);
-    streamToSubmit.getGraphicsAllocation()->updateResidencyTaskCount(this->taskCount + 1, this->osContext->getContextId());
 
     auto submissionStatus = flushHandler(batchBuffer, this->getResidencyAllocations());
     if (submissionStatus != SubmissionStatus::SUCCESS) {
+        updateStreamTaskCount(streamToSubmit, taskCount);
         CompletionStamp completionStamp = {CompletionStamp::getTaskCountFromSubmissionStatusError(submissionStatus)};
         return completionStamp;
     }
@@ -723,13 +723,14 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
                             dispatchFlags.requiresCoherency, dispatchFlags.lowPriority, dispatchFlags.throttle, dispatchFlags.sliceCount,
                             streamToSubmit.getUsed(), &streamToSubmit, bbEndLocation, this->getNumClients(), (submitCSR || dispatchFlags.hasStallingCmds || hasStallingCmdsOnTaskStream),
                             dispatchFlags.hasRelaxedOrderingDependencies};
-    streamToSubmit.getGraphicsAllocation()->updateTaskCount(this->taskCount + 1, this->osContext->getContextId());
+
-    streamToSubmit.getGraphicsAllocation()->updateResidencyTaskCount(this->taskCount + 1, this->osContext->getContextId());
+    updateStreamTaskCount(streamToSubmit, taskCount + 1);
 
     if (submitCSR || submitTask) {
         if (this->dispatchMode == DispatchMode::ImmediateDispatch) {
             auto submissionStatus = flushHandler(batchBuffer, this->getResidencyAllocations());
             if (submissionStatus != SubmissionStatus::SUCCESS) {
+                updateStreamTaskCount(streamToSubmit, taskCount);
                 CompletionStamp completionStamp = {CompletionStamp::getTaskCountFromSubmissionStatusError(submissionStatus)};
                 return completionStamp;
             }
@@ -1329,11 +1330,11 @@ TaskCountType CommandStreamReceiverHw<GfxFamily>::flushBcsTask(const BlitPropert
     BatchBuffer batchBuffer{commandStream.getGraphicsAllocation(), commandStreamStart, 0, taskStartAddress, nullptr, false, false, QueueThrottle::MEDIUM, QueueSliceCount::defaultSliceCount,
                             commandStream.getUsed(), &commandStream, endingCmdPtr, this->getNumClients(), hasStallingCmds, isRelaxedOrderingDispatch};
 
-    commandStream.getGraphicsAllocation()->updateTaskCount(newTaskCount, this->osContext->getContextId());
+    updateStreamTaskCount(commandStream, newTaskCount);
-    commandStream.getGraphicsAllocation()->updateResidencyTaskCount(newTaskCount, this->osContext->getContextId());
 
     auto flushSubmissionStatus = flush(batchBuffer, getResidencyAllocations());
     if (flushSubmissionStatus != SubmissionStatus::SUCCESS) {
+        updateStreamTaskCount(commandStream, taskCount);
         return CompletionStamp::getTaskCountFromSubmissionStatusError(flushSubmissionStatus);
     }
     makeSurfacePackNonResident(getResidencyAllocations(), true);
@@ -1447,7 +1448,9 @@ SubmissionStatus CommandStreamReceiverHw<GfxFamily>::flushSmallTask(LinearStream
 
     this->latestSentTaskCount = taskCount + 1;
     auto submissionStatus = flushHandler(batchBuffer, getResidencyAllocations());
-    taskCount++;
+    if (submissionStatus == SubmissionStatus::SUCCESS) {
+        taskCount++;
+    }
     return submissionStatus;
 }
 
@@ -1743,4 +1746,10 @@ inline void CommandStreamReceiverHw<GfxFamily>::handleStateBaseAddressStateTrans
     }
 }
 
+template <typename GfxFamily>
+void CommandStreamReceiverHw<GfxFamily>::updateStreamTaskCount(LinearStream &stream, TaskCountType newTaskCount) {
+    stream.getGraphicsAllocation()->updateTaskCount(newTaskCount, this->osContext->getContextId());
+    stream.getGraphicsAllocation()->updateResidencyTaskCount(newTaskCount, this->osContext->getContextId());
+}
+
 } // namespace NEO

shared/source/direct_submission/linux/drm_direct_submission.inl

@@ -126,7 +126,7 @@ bool DrmDirectSubmission<GfxFamily, Dispatcher>::submit(uint64_t gpuAddress, siz
     TaskCountType completionValue = 0u;
     uint64_t completionFenceGpuAddress = 0u;
     if (this->isCompletionFenceSupported()) {
-        completionValue = ++completionFenceValue;
+        completionValue = completionFenceValue + 1;
         completionFenceGpuAddress = this->completionFenceAllocation->getGpuAddress() + TagAllocationLayout::completionFenceOffset;
     }
 
@@ -155,6 +155,10 @@ bool DrmDirectSubmission<GfxFamily, Dispatcher>::submit(uint64_t gpuAddress, siz
         }
     }
 
+    if (this->isCompletionFenceSupported() && ret) {
+        completionFenceValue++;
+    }
+
     return ret;
 }
 

shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp

@@ -392,7 +392,7 @@ HWTEST_F(CommandStreamReceiverTest, givenFailingFlushSubmissionsAndNoGpuHangWhen
     EXPECT_EQ(WaitStatus::NotReady, waitStatus);
 }
 
-HWTEST_F(CommandStreamReceiverTest, givenGpuHangWhenWaititingForTaskCountThenGpuHangIsReturned) {
+HWTEST_F(CommandStreamReceiverTest, givenGpuHangWhenWaitingForTaskCountThenGpuHangIsReturned) {
     auto driverModelMock = std::make_unique<MockDriverModel>();
     driverModelMock->isGpuHangDetectedToReturn = true;
 
@@ -413,6 +413,32 @@ HWTEST_F(CommandStreamReceiverTest, givenGpuHangWhenWaititingForTaskCountThenGpu
     EXPECT_TRUE(csr.downloadAllocationCalled);
 }
 
+HWTEST_F(CommandStreamReceiverTest, givenFlushUnsuccessWhenWaitingForTaskCountThenNotReadyIsReturned) {
+    auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
+    csr.activePartitions = 1;
+
+    constexpr auto taskCountToWait = 1;
+    csr.flushReturnValue = SubmissionStatus::FAILED;
+    auto waitStatus = csr.waitForTaskCount(taskCountToWait);
+    EXPECT_EQ(WaitStatus::NotReady, waitStatus);
+
+    csr.flushReturnValue = SubmissionStatus::OUT_OF_MEMORY;
+    waitStatus = csr.waitForTaskCount(taskCountToWait);
+    EXPECT_EQ(WaitStatus::NotReady, waitStatus);
+
+    csr.flushReturnValue = SubmissionStatus::OUT_OF_HOST_MEMORY;
+    waitStatus = csr.waitForTaskCount(taskCountToWait);
+    EXPECT_EQ(WaitStatus::NotReady, waitStatus);
+
+    csr.flushReturnValue = SubmissionStatus::UNSUPPORTED;
+    waitStatus = csr.waitForTaskCount(taskCountToWait);
+    EXPECT_EQ(WaitStatus::NotReady, waitStatus);
+
+    csr.flushReturnValue = SubmissionStatus::DEVICE_UNINITIALIZED;
+    waitStatus = csr.waitForTaskCount(taskCountToWait);
+    EXPECT_EQ(WaitStatus::NotReady, waitStatus);
+}
+
 HWTEST_F(CommandStreamReceiverTest, whenDownloadTagAllocationThenDonwloadOnlyIfTagAllocationWasFlushed) {
     auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
     csr.activePartitions = 1;
@@ -2709,6 +2735,32 @@ HWTEST_F(CommandStreamReceiverHwTest, givenFailedFailureOnFlushWhenFlushingTaskT
     EXPECT_EQ(CompletionStamp::failed, completionStamp.taskCount);
 }
 
+HWTEST_F(CommandStreamReceiverHwTest, givenUnsuccessOnFlushWhenFlushingSmallTaskThenTaskCountIsNotIncreased) {
+    auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
+    auto &stream = commandStreamReceiver.getCS(4096u);
+
+    commandStreamReceiver.taskCount = 1u;
+    commandStreamReceiver.flushReturnValue = SubmissionStatus::FAILED;
+    commandStreamReceiver.flushSmallTask(stream, stream.getUsed());
+    EXPECT_EQ(1u, commandStreamReceiver.taskCount);
+
+    commandStreamReceiver.flushReturnValue = SubmissionStatus::OUT_OF_MEMORY;
+    commandStreamReceiver.flushSmallTask(stream, stream.getUsed());
+    EXPECT_EQ(1u, commandStreamReceiver.taskCount);
+
+    commandStreamReceiver.flushReturnValue = SubmissionStatus::OUT_OF_HOST_MEMORY;
+    commandStreamReceiver.flushSmallTask(stream, stream.getUsed());
+    EXPECT_EQ(1u, commandStreamReceiver.taskCount);
+
+    commandStreamReceiver.flushReturnValue = SubmissionStatus::UNSUPPORTED;
+    commandStreamReceiver.flushSmallTask(stream, stream.getUsed());
+    EXPECT_EQ(1u, commandStreamReceiver.taskCount);
+
+    commandStreamReceiver.flushReturnValue = SubmissionStatus::DEVICE_UNINITIALIZED;
+    commandStreamReceiver.flushSmallTask(stream, stream.getUsed());
+    EXPECT_EQ(1u, commandStreamReceiver.taskCount);
+}
+
 HWTEST_F(CommandStreamReceiverHwTest, givenOutOfMemoryFailureOnFlushWhenFlushingMiDWThenErrorIsPropagated) {
     auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
 

shared/test/unit_test/direct_submission/linux/drm_direct_submission_tests.cpp

@@ -399,6 +399,32 @@ HWTEST_F(DrmDirectSubmissionTest, givenNoCompletionFenceSupportAndExecFailureWhe
     ringBuffer->getBufferObjectToModify(0) = initialBO;
 }
 
+HWTEST_F(DrmDirectSubmissionTest, givenCompletionFenceSupportAndExecFailureWhenSubmittingThenCompletionFenceValueIsNotIncreased) {
+    uint64_t gpuAddress = 0x1000;
+    size_t size = 0x1000;
+    DebugManagerStateRestore restorer;
+    DebugManager.flags.EnableDrmCompletionFence.set(1);
+
+    auto &commandStreamReceiver = *device->getDefaultEngine().commandStreamReceiver;
+    auto drm = executionEnvironment.rootDeviceEnvironments[0]->osInterface->getDriverModel()->as<Drm>();
+
+    MockDrmDirectSubmission<FamilyType, RenderDispatcher<FamilyType>> drmDirectSubmission(commandStreamReceiver);
+    drmDirectSubmission.completionFenceAllocation = commandStreamReceiver.getTagAllocation();
+    EXPECT_TRUE(drmDirectSubmission.allocateResources());
+    auto ringBuffer = static_cast<DrmAllocation *>(drmDirectSubmission.ringBuffers[drmDirectSubmission.currentRingBuffer].ringBuffer);
+    auto initialBO = ringBuffer->getBufferObjectToModify(0);
+
+    MockBufferObject mockBO(drm);
+    ringBuffer->getBufferObjectToModify(0) = &mockBO;
+
+    mockBO.execReturnValue = 1;
+    drmDirectSubmission.completionFenceValue = 1u;
+    EXPECT_FALSE(drmDirectSubmission.submit(gpuAddress, size));
+    EXPECT_EQ(1u, drmDirectSubmission.completionFenceValue);
+
+    ringBuffer->getBufferObjectToModify(0) = initialBO;
+}
+
 HWTEST_F(DrmDirectSubmissionTest, givenTile0AndCompletionFenceSupportWhenSubmittingThenCompletionAddressAndValueArePassedToExec) {
     uint64_t gpuAddress = 0x1000;
     size_t size = 0x1000;