performance: Allow power efficient waits with ULLS light

Resolves: HSD-18041633458
Related-To: NEO-13922

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>

shared/source/direct_submission/linux/drm_direct_submission.h

@@ -50,6 +50,7 @@ class DrmDirectSubmission : public DirectSubmissionHw<GfxFamily, Dispatcher> {
 
     constexpr static size_t ullsLightTimeout = 2'000'000;
     std::chrono::steady_clock::time_point lastUllsLightExecTimestamp{};
+    int boHandleForExec = 0;
 
     std::vector<BufferObject *> residency{};
     std::vector<ExecObject> execObjectsStorage{};

shared/source/direct_submission/linux/drm_direct_submission.inl

@@ -144,6 +144,7 @@ bool DrmDirectSubmission<GfxFamily, Dispatcher>::submit(uint64_t gpuAddress, siz
         this->handleResidency();
     } else {
         this->lastUllsLightExecTimestamp = std::chrono::steady_clock::now();
+        this->boHandleForExec = bb->peekHandle();
     }
 
     auto currentBase = this->ringCommandStream.getGraphicsAllocation()->getGpuAddress();
@@ -257,7 +258,7 @@ uint64_t DrmDirectSubmission<GfxFamily, Dispatcher>::updateTagValue(bool require
         this->currentTagData.tagValue++;
         this->ringBuffers[this->currentRingBuffer].completionFence = this->currentTagData.tagValue;
     }
-    return 0ull;
+    return boHandleForExec;
 }
 
 template <typename GfxFamily, typename Dispatcher>

shared/source/os_interface/linux/drm_neo.cpp

@@ -120,6 +120,9 @@ int Drm::ioctl(DrmIoctl request, void *arg) {
             end = std::chrono::steady_clock::now();
             long long elapsedTime = std::chrono::duration_cast<std::chrono::nanoseconds>(end - start).count();
 
+            static std::mutex mtx;
+            std::lock_guard lock(mtx);
+
             IoctlStatisticsEntry ioctlData{};
             auto ioctlDataIt = this->ioctlStatistics.find(request);
             if (ioctlDataIt != this->ioctlStatistics.end()) {

shared/source/xe_hpg_core/xe_lpg/linux/product_helper_xe_lpg_linux.inl

@@ -22,7 +22,7 @@ int ProductHelperHw<gfxProduct>::configureHardwareCustom(HardwareInfo *hwInfo, O
     kmdNotifyProperties.enableKmdNotify = true;
     kmdNotifyProperties.delayKmdNotifyMicroseconds = 150;
     kmdNotifyProperties.enableQuickKmdSleepForDirectSubmission = true;
-    kmdNotifyProperties.delayQuickKmdSleepForDirectSubmissionMicroseconds = 20;
+    kmdNotifyProperties.delayQuickKmdSleepForDirectSubmissionMicroseconds = 28000;
 
     return 0;
 }

shared/test/unit_test/os_interface/linux/drm_command_stream_tests_1.cpp

@@ -994,6 +994,7 @@ HWTEST_TEMPLATED_F(DrmCommandStreamDirectSubmissionTest, givenEnabledDirectSubmi
     batchBuffer.startOffset = 4;
     uint8_t bbStart[64];
     batchBuffer.endCmdPtr = &bbStart[0];
+    static_cast<DrmMockCustom *>(static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->drm)->isVmBindAvailableCall.callParent = false;
 
     auto flushStamp = csr->obtainCurrentFlushStamp();
     csr->flush(batchBuffer, csr->getResidencyAllocations());
@@ -1005,6 +1006,25 @@ HWTEST_TEMPLATED_F(DrmCommandStreamDirectSubmissionTest, givenEnabledDirectSubmi
     static_cast<MockDrmDirectSubmission<FamilyType> *>(directSubmission)->currentTagData.tagValue = 0u;
 }
 
+HWTEST_TEMPLATED_F(DrmCommandStreamDirectSubmissionTest, givenEnabledDirectSubmissionLightWhenFlushThenFlushStampIsUpdated) {
+    auto &cs = csr->getCS();
+    CommandStreamReceiverHw<FamilyType>::addBatchBufferEnd(cs, nullptr);
+    EncodeNoop<FamilyType>::alignToCacheLine(cs);
+    BatchBuffer batchBuffer = BatchBufferHelper::createDefaultBatchBuffer(cs.getGraphicsAllocation(), &cs, cs.getUsed());
+    batchBuffer.startOffset = 4;
+    uint8_t bbStart[64];
+    batchBuffer.endCmdPtr = &bbStart[0];
+
+    auto flushStamp = csr->obtainCurrentFlushStamp();
+    csr->flush(batchBuffer, csr->getResidencyAllocations());
+
+    EXPECT_NE(csr->obtainCurrentFlushStamp(), flushStamp);
+
+    auto directSubmission = static_cast<TestedDrmCommandStreamReceiver<FamilyType> *>(csr)->directSubmission.get();
+    ASSERT_NE(nullptr, directSubmission);
+    static_cast<MockDrmDirectSubmission<FamilyType> *>(directSubmission)->currentTagData.tagValue = 0u;
+}
+
 HWTEST_TEMPLATED_F(DrmCommandStreamDirectSubmissionTest, givenEnabledDirectSubmissionWhenFlushThenCommandBufferAllocationIsNotAddedToHandlerResidencySet) {
     mock->bindAvailable = true;
     auto &cs = csr->getCS();

shared/test/unit_test/xe_hpg_core/arl/linux/product_helper_tests_arl.cpp

@@ -36,7 +36,7 @@ ARLTEST_F(ArlProductHelperLinux, GivenArlWhenConfigureHardwareCustomThenKmdNotif
     EXPECT_TRUE(pInHwInfo.capabilityTable.kmdNotifyProperties.enableKmdNotify);
     EXPECT_EQ(150ll, pInHwInfo.capabilityTable.kmdNotifyProperties.delayKmdNotifyMicroseconds);
     EXPECT_TRUE(pInHwInfo.capabilityTable.kmdNotifyProperties.enableQuickKmdSleepForDirectSubmission);
-    EXPECT_EQ(20ll, pInHwInfo.capabilityTable.kmdNotifyProperties.delayQuickKmdSleepForDirectSubmissionMicroseconds);
+    EXPECT_EQ(28000ll, pInHwInfo.capabilityTable.kmdNotifyProperties.delayQuickKmdSleepForDirectSubmissionMicroseconds);
 }
 
 ARLTEST_F(ArlProductHelperLinux, givenArlWhenIsBlitterForImagesSupportedIsCalledThenTrueIsReturned) {

shared/test/unit_test/xe_hpg_core/mtl/linux/product_helper_tests_mtl_linux.cpp

@@ -36,7 +36,7 @@ MTLTEST_F(MtlProductHelperLinux, GivenMtlWhenConfigureHardwareCustomThenKmdNotif
     EXPECT_TRUE(pInHwInfo.capabilityTable.kmdNotifyProperties.enableKmdNotify);
     EXPECT_EQ(150ll, pInHwInfo.capabilityTable.kmdNotifyProperties.delayKmdNotifyMicroseconds);
     EXPECT_TRUE(pInHwInfo.capabilityTable.kmdNotifyProperties.enableQuickKmdSleepForDirectSubmission);
-    EXPECT_EQ(20ll, pInHwInfo.capabilityTable.kmdNotifyProperties.delayQuickKmdSleepForDirectSubmissionMicroseconds);
+    EXPECT_EQ(28000ll, pInHwInfo.capabilityTable.kmdNotifyProperties.delayQuickKmdSleepForDirectSubmissionMicroseconds);
 }
 
 MTLTEST_F(MtlProductHelperLinux, givenMtlWhenIsBlitterForImagesSupportedIsCalledThenTrueIsReturned) {