fix: Return success for system address in setArg

This patch avoids returning error for system addresses in setArg

Related-To: GSD-3597

Signed-off-by: Joshua Santosh Ranjan <joshua.santosh.ranjan@intel.com>
Source: 91784a8

Author: joshuaranjan

level_zero/core/source/kernel/kernel_imp.cpp

@@ -701,10 +701,7 @@ ze_result_t KernelImp::setArgBuffer(uint32_t argIndex, size_t argSize, const voi
         allocData = svmAllocsManager->getSVMAlloc(requestedAddress);
     }
     NEO::SvmAllocationData *peerAllocData = nullptr;
-    if (driverHandle->isRemoteResourceNeeded(requestedAddress, alloc, allocData, device)) {
-        if (allocData == nullptr) {
-            return ZE_RESULT_ERROR_INVALID_ARGUMENT;
-        }
+    if (allocData && driverHandle->isRemoteResourceNeeded(requestedAddress, alloc, allocData, device)) {
 
         uint64_t pbase = allocData->gpuAllocations.getDefaultGraphicsAllocation()->getGpuAddress();
         uint64_t offset = (uint64_t)requestedAddress - pbase;
@@ -714,10 +711,23 @@ ze_result_t KernelImp::setArgBuffer(uint32_t argIndex, size_t argSize, const voi
         }
         gpuAddress += offset;
     }
-    const uint32_t allocId = allocData ? allocData->getAllocId() : 0u;
+
+    if (allocData == nullptr) {
+        if (NEO::DebugManager.flags.DisableSystemPointerKernelArgument.get() != 1) {
+            const auto &argAsPtr = kernelImmData->getDescriptor().payloadMappings.explicitArgs[argIndex].as<NEO::ArgDescPointer>();
+            auto patchLocation = ptrOffset(getCrossThreadData(), argAsPtr.stateless);
+            patchWithRequiredSize(const_cast<uint8_t *>(patchLocation), argAsPtr.pointerSize, reinterpret_cast<uintptr_t>(requestedAddress));
+            kernelArgInfos[argIndex] = KernelArgInfo{requestedAddress, 0, 0, false};
+            return ZE_RESULT_SUCCESS;
+        } else {
+            return ZE_RESULT_ERROR_INVALID_ARGUMENT;
+        }
+    }
+
+    const uint32_t allocId = allocData->getAllocId();
     kernelArgInfos[argIndex] = KernelArgInfo{requestedAddress, allocId, allocationsCounter, false};
 
-    if (allocData && allocData->virtualReservationData) {
+    if (allocData->virtualReservationData) {
         for (const auto &mappedAllocationData : allocData->virtualReservationData->mappedAllocations) {
             // Add additional allocations to the residency container if the virtual reservation spans multiple allocations.
             if (requestedAddress != mappedAllocationData.second->ptr) {

level_zero/core/test/unit_tests/sources/kernel/test_kernel.cpp

@@ -274,15 +274,17 @@ TEST_F(SetKernelArgCacheTest, givenValidBufferArgumentWhenSetMultipleTimesThenSe
     EXPECT_EQ(ZE_RESULT_SUCCESS, mockKernel.setArgBuffer(0, sizeof(secondSvmAllocation), &secondSvmAllocation));
     EXPECT_EQ(++callCounter, mockKernel.setArgBufferWithAllocCalled);
 
-    // same value but no svmData - ZE_RESULT_ERROR_INVALID_ARGUMENT
+    // same value but no svmData - ZE_RESULT_SUCCESS with allocId as 0
     svmAllocsManager->freeSVMAlloc(secondSvmAllocation);
     ++svmAllocsManager->allocationsCounter;
     ASSERT_GT(mockKernel.kernelArgInfos[0].allocId, 0u);
     ASSERT_LT(mockKernel.kernelArgInfos[0].allocId, SvmAllocationData::uninitializedAllocId);
     ASSERT_EQ(mockKernel.kernelArgInfos[0].value, secondSvmAllocation);
     ASSERT_GT(svmAllocsManager->allocationsCounter, 0u);
     ASSERT_NE(mockKernel.kernelArgInfos[0].allocIdMemoryManagerCounter, svmAllocsManager->allocationsCounter);
-    EXPECT_EQ(ZE_RESULT_ERROR_INVALID_ARGUMENT, mockKernel.setArgBuffer(0, sizeof(secondSvmAllocation), &secondSvmAllocation));
+    EXPECT_EQ(ZE_RESULT_SUCCESS, mockKernel.setArgBuffer(0, sizeof(secondSvmAllocation), &secondSvmAllocation));
+    ASSERT_EQ(mockKernel.kernelArgInfos[0].value, secondSvmAllocation);
+    ASSERT_EQ(mockKernel.kernelArgInfos[0].allocId, 0u);
     EXPECT_EQ(callCounter, mockKernel.setArgBufferWithAllocCalled);
 
     svmAllocsManager->freeSVMAlloc(svmAllocation);
@@ -626,15 +628,22 @@ HWTEST2_F(SetKernelArg, givenSamplerAndKernelWhenSetArgSamplerThenCrossThreadDat
     EXPECT_EQ(static_cast<uint32_t>(SamplerPatchValues::NormalizedCoordsTrue), *pSamplerNormalizedCoords);
 }
 
-using ArgSupport = IsWithinProducts<IGFX_SKYLAKE, IGFX_TIGERLAKE_LP>;
-
-HWTEST2_F(SetKernelArg, givenBufferArgumentWhichHasNotBeenAllocatedByRuntimeThenInvalidArgumentIsReturned, ArgSupport) {
+TEST_F(SetKernelArg, givenBufferArgumentWhichHasNotBeenAllocatedByRuntimeThenSuccessIsReturned) {
     createKernel();
 
     uint64_t hostAddress = 0x1234;
-
     ze_result_t res = kernel->setArgBuffer(0, sizeof(hostAddress), &hostAddress);
+    EXPECT_EQ(ZE_RESULT_SUCCESS, res);
+}
 
+TEST_F(SetKernelArg, givenDisableSystemPointerKernelArgumentIsEnabledWhenBufferArgumentisNotAllocatedByRuntimeThenErrorIsReturned) {
+
+    DebugManagerStateRestore restorer;
+    NEO::DebugManager.flags.DisableSystemPointerKernelArgument.set(1);
+    createKernel();
+
+    uint64_t hostAddress = 0x1234;
+    ze_result_t res = kernel->setArgBuffer(0, sizeof(hostAddress), &hostAddress);
     EXPECT_EQ(ZE_RESULT_ERROR_INVALID_ARGUMENT, res);
 }
 

shared/source/debug_settings/debug_variables_base.inl

@@ -250,6 +250,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, SynchronizeEventBeforeReset, -1, "-1: default, 0
 DECLARE_DEBUG_VARIABLE(int32_t, TrackNumCsrClientsOnSyncPoints, -1, "-1: default, 0: Disabled, 1: If set, synchronization points like zeEventHostSynchronize will unregister CmdQ from CSR clients")
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideDriverVersion, -1, "-1: default, >=0: Use value as reported driver version")
 DECLARE_DEBUG_VARIABLE(int32_t, WaitForUserFenceOnEventHostSynchronize, -1, "-1: default, 0: Disabled, 1: Enabled. If enabled, use WaitUserFence KMD call for in-order Events instead of active polling on host.")
+DECLARE_DEBUG_VARIABLE(int32_t, DisableSystemPointerKernelArgument, -1, "-1: default, 0: Disabled, 1: using a system pointer for kernel argument returns an error.")
 DECLARE_DEBUG_VARIABLE(int32_t, ProgramUserInterruptOnResolvedDependency, -1, "-1: default, 0: Disabled, >=1: bitfield. 01b: program after semaphore, 10b: on signaling fence (non-walker append).")
 
 /*LOGGING FLAGS*/

shared/test/common/test_files/igdrcl.config

@@ -548,5 +548,6 @@ EnableAIL=1
 WaitForUserFenceOnEventHostSynchronize = -1
 ProgramUserInterruptOnResolvedDependency = -1
 ProgramBarrierInCommandStreamTask = -1
+DisableSystemPointerKernelArgument = -1
 DoNotValidateDriverPath = 0
 # Please don't edit below this line