fix: handle not aligned gtt size reported by i915

JablonskiMateusz · Compute-Runtime-Automation · commit 852107594943 · 2024-04-02T11:07:17.000+02:00
when i915 reports gtt size between 47 and 48 bits we consider it as 48 bit VA space Related-To: GSD-8215 Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com> Source: 420e139
diff --git a/shared/source/memory_manager/gfx_partition.cpp b/shared/source/memory_manager/gfx_partition.cpp
@@ -190,7 +190,7 @@ uint64_t GfxPartition::getHeapMinimalAddress(HeapIndex heapIndex) {
     }
 }
 
-bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize) {
+bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize, uint64_t gfxTop) {
 
     /*
      * I. 64-bit builds:
@@ -238,7 +238,6 @@ bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToRe
      *  0x0  0x100000000                                       gpuAddressSpace
      */
 
-    uint64_t gfxTop = gpuAddressSpace + 1;
     uint64_t gfxBase = 0x0ull;
     const uint64_t gfxHeap32Size = 4 * MemoryConstants::gigaByte;
 
diff --git a/shared/source/memory_manager/gfx_partition.h b/shared/source/memory_manager/gfx_partition.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2019-2023 Intel Corporation
+ * Copyright (C) 2019-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,7 +39,7 @@ class GfxPartition {
     GfxPartition(OSMemory::ReservedCpuAddressRange &reservedCpuAddressRangeForHeapSvm);
     MOCKABLE_VIRTUAL ~GfxPartition();
 
-    MOCKABLE_VIRTUAL bool init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize);
+    MOCKABLE_VIRTUAL bool init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize, uint64_t gfxTop);
 
     void heapInit(HeapIndex heapIndex, uint64_t base, uint64_t size) {
         getHeap(heapIndex).init(base, size, MemoryConstants::pageSize);
diff --git a/shared/source/memory_manager/os_agnostic_memory_manager.cpp b/shared/source/memory_manager/os_agnostic_memory_manager.cpp
@@ -46,7 +46,8 @@ void OsAgnosticMemoryManager::initialize(bool aubUsage) {
         this->enable64kbpages[rootDeviceIndex] = is64kbPagesEnabled(hwInfo);
         this->localMemorySupported.push_back(gfxCoreHelper.getEnableLocalMemory(*hwInfo));
         auto gpuAddressSpace = executionEnvironment.rootDeviceEnvironments[rootDeviceIndex]->getHardwareInfo()->capabilityTable.gpuAddressSpace;
-        if (!getGfxPartition(rootDeviceIndex)->init(gpuAddressSpace, reservedCpuAddressRangeSize, rootDeviceIndex, gfxPartitions.size(), heapAssigners[rootDeviceIndex]->apiAllowExternalHeapForSshAndDsh, OsAgnosticMemoryManager::getSystemSharedMemory(rootDeviceIndex))) {
+        auto gfxTop = gpuAddressSpace + 1;
+        if (!getGfxPartition(rootDeviceIndex)->init(gpuAddressSpace, reservedCpuAddressRangeSize, rootDeviceIndex, gfxPartitions.size(), heapAssigners[rootDeviceIndex]->apiAllowExternalHeapForSshAndDsh, OsAgnosticMemoryManager::getSystemSharedMemory(rootDeviceIndex), gfxTop)) {
             initialized = false;
             return;
         }
diff --git a/shared/source/os_interface/linux/drm_memory_manager.cpp b/shared/source/os_interface/linux/drm_memory_manager.cpp
@@ -99,7 +99,9 @@ void DrmMemoryManager::initialize(GemCloseWorkerMode mode) {
 
     for (uint32_t rootDeviceIndex = 0; rootDeviceIndex < gfxPartitions.size(); ++rootDeviceIndex) {
         auto gpuAddressSpace = executionEnvironment.rootDeviceEnvironments[rootDeviceIndex]->getHardwareInfo()->capabilityTable.gpuAddressSpace;
-        if (!getGfxPartition(rootDeviceIndex)->init(gpuAddressSpace, getSizeToReserve(), rootDeviceIndex, gfxPartitions.size(), heapAssigners[rootDeviceIndex]->apiAllowExternalHeapForSshAndDsh, DrmMemoryManager::getSystemSharedMemory(rootDeviceIndex))) {
+        uint64_t gfxTop{};
+        getDrm(rootDeviceIndex).queryGttSize(gfxTop, false);
+        if (!getGfxPartition(rootDeviceIndex)->init(gpuAddressSpace, getSizeToReserve(), rootDeviceIndex, gfxPartitions.size(), heapAssigners[rootDeviceIndex]->apiAllowExternalHeapForSshAndDsh, DrmMemoryManager::getSystemSharedMemory(rootDeviceIndex), gfxTop)) {
             initialized = false;
             return;
         }
@@ -1248,7 +1250,7 @@ uint64_t DrmMemoryManager::getSystemSharedMemory(uint32_t rootDeviceIndex) {
 
     uint64_t gpuMemorySize = 0u;
 
-    [[maybe_unused]] auto ret = getDrm(rootDeviceIndex).queryGttSize(gpuMemorySize);
+    [[maybe_unused]] auto ret = getDrm(rootDeviceIndex).queryGttSize(gpuMemorySize, false);
     DEBUG_BREAK_IF(ret != 0);
 
     return std::min(hostMemorySize, gpuMemorySize);
diff --git a/shared/source/os_interface/linux/drm_neo.cpp b/shared/source/os_interface/linux/drm_neo.cpp
@@ -232,13 +232,17 @@ bool Drm::readSysFsAsString(const std::string &relativeFilePath, std::string &re
     return true;
 }
 
-int Drm::queryGttSize(uint64_t &gttSizeOutput) {
+int Drm::queryGttSize(uint64_t &gttSizeOutput, bool alignUpToFullRange) {
     GemContextParam contextParam = {0};
     contextParam.param = ioctlHelper->getDrmParamValue(DrmParam::contextParamGttSize);
 
     int ret = ioctlHelper->ioctl(DrmIoctl::gemContextGetparam, &contextParam);
     if (ret == 0) {
-        gttSizeOutput = contextParam.value;
+        if (alignUpToFullRange) {
+            gttSizeOutput = Drm::alignUpGttSize(contextParam.value);
+        } else {
+            gttSizeOutput = contextParam.value;
+        }
     }
 
     return ret;
@@ -1579,6 +1583,17 @@ void Drm::waitOnUserFences(const OsContextLinux &osContext, uint64_t address, ui
 }
 const HardwareInfo *Drm::getHardwareInfo() const { return rootDeviceEnvironment.getHardwareInfo(); }
 
+uint64_t Drm::alignUpGttSize(uint64_t inputGttSize) {
+
+    constexpr uint64_t gttSize47bit = (1ull << 47);
+    constexpr uint64_t gttSize48bit = (1ull << 48);
+
+    if (inputGttSize > gttSize47bit && inputGttSize < gttSize48bit) {
+        return gttSize48bit;
+    }
+    return inputGttSize;
+}
+
 template std::vector<uint16_t> Drm::query<uint16_t>(uint32_t queryId, uint32_t queryItemFlags);
 template std::vector<uint32_t> Drm::query<uint32_t>(uint32_t queryId, uint32_t queryItemFlags);
 template std::vector<uint64_t> Drm::query<uint64_t>(uint32_t queryId, uint32_t queryItemFlags);
diff --git a/shared/source/os_interface/linux/drm_neo.h b/shared/source/os_interface/linux/drm_neo.h
@@ -92,7 +92,7 @@ class Drm : public DriverModel {
     int getTimestampFrequency(int &frequency);
     int getOaTimestampFrequency(int &frequency);
 
-    MOCKABLE_VIRTUAL int queryGttSize(uint64_t &gttSizeOutput);
+    MOCKABLE_VIRTUAL int queryGttSize(uint64_t &gttSizeOutput, bool alignUpToFullRange);
     bool isPreemptionSupported() const { return preemptionSupported; }
 
     MOCKABLE_VIRTUAL void checkPreemptionSupport();
@@ -269,6 +269,7 @@ class Drm : public DriverModel {
     void queryAndSetVmBindPatIndexProgrammingSupport();
     bool queryDeviceIdAndRevision();
     bool queryI915DeviceIdAndRevision();
+    static uint64_t alignUpGttSize(uint64_t inputGttSize);
 
 #pragma pack(1)
     struct PCIConfig {
diff --git a/shared/source/os_interface/linux/product_helper_drm.cpp b/shared/source/os_interface/linux/product_helper_drm.cpp
@@ -125,7 +125,7 @@ int ProductHelper::configureHwInfoDrm(const HardwareInfo *inHwInfo, HardwareInfo
     uint64_t gttSizeQuery = 0;
     featureTable->flags.ftrSVM = true;
 
-    ret = drm->queryGttSize(gttSizeQuery);
+    ret = drm->queryGttSize(gttSizeQuery, true);
 
     if (ret == 0) {
         featureTable->flags.ftrSVM = (gttSizeQuery > MemoryConstants::max64BitAppAddress);
diff --git a/shared/test/common/libult/linux/drm_mock.h b/shared/test/common/libult/linux/drm_mock.h
@@ -174,7 +174,7 @@ class DrmMock : public Drm {
         else
             return Drm::useVMBindImmediate();
     }
-    int queryGttSize(uint64_t &gttSizeOutput) override {
+    int queryGttSize(uint64_t &gttSizeOutput, bool alignUpToFullRange) override {
         gttSizeOutput = storedGTTSize;
         return storedRetValForGetGttSize;
     }
@@ -276,7 +276,7 @@ class DrmMock : public Drm {
     GemVmControl receivedGemVmControl{};
     uint32_t latestCreatedVmId = 0u;
 
-    uint64_t storedGTTSize = 1ull << 47;
+    uint64_t storedGTTSize = defaultHwInfo->capabilityTable.gpuAddressSpace + 1;
     uint64_t storedParamSseu = ULONG_MAX;
 
     Ioctls ioctlCount{};
diff --git a/shared/test/common/mocks/linux/mock_drm_memory_manager.cpp b/shared/test/common/mocks/linux/mock_drm_memory_manager.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2020-2023 Intel Corporation
+ * Copyright (C) 2020-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -69,7 +69,7 @@ void TestedDrmMemoryManager::injectPinBB(BufferObject *newPinBB, uint32_t rootDe
 DrmGemCloseWorker *TestedDrmMemoryManager::getgemCloseWorker() { return this->gemCloseWorker.get(); }
 void TestedDrmMemoryManager::forceLimitedRangeAllocator(uint64_t range) {
     for (auto &gfxPartition : gfxPartitions) {
-        gfxPartition->init(range, getSizeToReserve(), 0, 1, false, 0u);
+        gfxPartition->init(range, getSizeToReserve(), 0, 1, false, 0u, range + 1);
     }
 }
 void TestedDrmMemoryManager::overrideGfxPartition(GfxPartition *newGfxPartition) { gfxPartitions[0].reset(newGfxPartition); }
diff --git a/shared/test/common/mocks/mock_gfx_partition.h b/shared/test/common/mocks/mock_gfx_partition.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2019-2023 Intel Corporation
+ * Copyright (C) 2019-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -92,7 +92,7 @@ class MockGfxPartitionBasic : public GfxPartition {
 
 class FailedInitGfxPartition : public MockGfxPartition {
   public:
-    bool init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useFrontWindowPool, uint64_t systemMemorySize) override {
+    bool init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useFrontWindowPool, uint64_t systemMemorySize, uint64_t gfxTop) override {
         return false;
     }
 };
diff --git a/shared/test/common/mocks/mock_memory_manager.cpp b/shared/test/common/mocks/mock_memory_manager.cpp
@@ -197,7 +197,7 @@ GraphicsAllocation *MockMemoryManager::allocate32BitGraphicsMemoryImpl(const All
 }
 
 void MockMemoryManager::forceLimitedRangeAllocator(uint32_t rootDeviceIndex, uint64_t range) {
-    getGfxPartition(rootDeviceIndex)->init(range, 0, 0, gfxPartitions.size(), false, 0u);
+    getGfxPartition(rootDeviceIndex)->init(range, 0, 0, gfxPartitions.size(), false, 0u, range + 1);
 }
 
 bool MockMemoryManager::hasPageFaultsEnabled(const Device &neoDevice) {
diff --git a/shared/test/common/os_interface/linux/device_command_stream_fixture.h b/shared/test/common/os_interface/linux/device_command_stream_fixture.h
@@ -184,11 +184,11 @@ class DrmMockCustom : public Drm {
 
     virtual void execBufferExtensions(void *execbuf) {
     }
-    int queryGttSize(uint64_t &gttSizeOutput) override {
+    int queryGttSize(uint64_t &gttSizeOutput, bool alignUpToFullRange) override {
         if (callBaseQueryGttSize) {
-            return Drm::queryGttSize(gttSizeOutput);
+            return Drm::queryGttSize(gttSizeOutput, alignUpToFullRange);
         }
-        gttSizeOutput = 1;
+        gttSizeOutput = NEO::defaultHwInfo->capabilityTable.gpuAddressSpace + 1;
         return 0u;
     }
 
diff --git a/shared/test/unit_test/fixtures/front_window_fixture.cpp b/shared/test/unit_test/fixtures/front_window_fixture.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2021-2023 Intel Corporation
+ * Copyright (C) 2021-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -17,7 +17,7 @@
 using namespace NEO;
 
 MemManagerFixture::FrontWindowMemManagerMock::FrontWindowMemManagerMock(NEO::ExecutionEnvironment &executionEnvironment) : MockMemoryManager(executionEnvironment) {}
-void MemManagerFixture::FrontWindowMemManagerMock::forceLimitedRangeAllocator(uint32_t rootDeviceIndex, uint64_t range) { getGfxPartition(rootDeviceIndex)->init(range, 0, 0, gfxPartitions.size(), true, 0u); }
+void MemManagerFixture::FrontWindowMemManagerMock::forceLimitedRangeAllocator(uint32_t rootDeviceIndex, uint64_t range) { getGfxPartition(rootDeviceIndex)->init(range, 0, 0, gfxPartitions.size(), true, 0u, range + 1); }
 
 void MemManagerFixture::setUp() {
     DebugManagerStateRestore dbgRestorer;
diff --git a/shared/test/unit_test/memory_manager/gfx_partition_tests.cpp b/shared/test/unit_test/memory_manager/gfx_partition_tests.cpp
diff --git a/shared/test/unit_test/os_interface/linux/drm_memory_manager_localmem_prelim_tests.cpp b/shared/test/unit_test/os_interface/linux/drm_memory_manager_localmem_prelim_tests.cpp
diff --git a/shared/test/unit_test/os_interface/linux/drm_memory_manager_localmem_upstream_tests.cpp b/shared/test/unit_test/os_interface/linux/drm_memory_manager_localmem_upstream_tests.cpp
diff --git a/shared/test/unit_test/os_interface/linux/drm_memory_manager_tests.cpp b/shared/test/unit_test/os_interface/linux/drm_memory_manager_tests.cpp
diff --git a/shared/test/unit_test/os_interface/linux/drm_special_heap_test.cpp b/shared/test/unit_test/os_interface/linux/drm_special_heap_test.cpp
diff --git a/shared/test/unit_test/os_interface/linux/drm_tests.cpp b/shared/test/unit_test/os_interface/linux/drm_tests.cpp

Original file line number	Diff line number	Diff line change
`@@ -190,7 +190,7 @@ uint64_t GfxPartition::getHeapMinimalAddress(HeapIndex heapIndex) {`
`190`	`190`	`}`
`191`	`191`	`}`
`192`	`192`
`193`		`-bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize) {`
	`193`	`+bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize, uint64_t gfxTop) {`
`194`	`194`
`195`	`195`	`/*`
`196`	`196`	`* I. 64-bit builds:`
`@@ -238,7 +238,6 @@ bool GfxPartition::init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToRe`
`238`	`238`	`* 0x0 0x100000000 gpuAddressSpace`
`239`	`239`	`*/`
`240`	`240`
`241`		`- uint64_t gfxTop = gpuAddressSpace + 1;`
`242`	`241`	`uint64_t gfxBase = 0x0ull;`
`243`	`242`	`const uint64_t gfxHeap32Size = 4 * MemoryConstants::gigaByte;`
`244`	`243`
Original file line number	Diff line number	Diff line change
`@@ -1,5 +1,5 @@`
`1`	`1`	`/*`
`2`		`- * Copyright (C) 2019-2023 Intel Corporation`
	`2`	`+ * Copyright (C) 2019-2024 Intel Corporation`
`3`	`3`	`*`
`4`	`4`	`* SPDX-License-Identifier: MIT`
`5`	`5`	`*`
`@@ -39,7 +39,7 @@ class GfxPartition {`
`39`	`39`	`GfxPartition(OSMemory::ReservedCpuAddressRange &reservedCpuAddressRangeForHeapSvm);`
`40`	`40`	`MOCKABLE_VIRTUAL ~GfxPartition();`
`41`	`41`
`42`		`- MOCKABLE_VIRTUAL bool init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize);`
	`42`	`+ MOCKABLE_VIRTUAL bool init(uint64_t gpuAddressSpace, size_t cpuAddressRangeSizeToReserve, uint32_t rootDeviceIndex, size_t numRootDevices, bool useExternalFrontWindowPool, uint64_t systemMemorySize, uint64_t gfxTop);`
`43`	`43`
`44`	`44`	`void heapInit(HeapIndex heapIndex, uint64_t base, uint64_t size) {`
`45`	`45`	`getHeap(heapIndex).init(base, size, MemoryConstants::pageSize);`
Original file line number	Diff line number	Diff line change
`@@ -1,5 +1,5 @@`
`1`	`1`	`/*`
`2`		`- * Copyright (C) 2020-2023 Intel Corporation`
	`2`	`+ * Copyright (C) 2020-2024 Intel Corporation`
`3`	`3`	`*`
`4`	`4`	`* SPDX-License-Identifier: MIT`
`5`	`5`	`*`
`@@ -69,7 +69,7 @@ void TestedDrmMemoryManager::injectPinBB(BufferObject *newPinBB, uint32_t rootDe`
`69`	`69`	`DrmGemCloseWorker *TestedDrmMemoryManager::getgemCloseWorker() { return this->gemCloseWorker.get(); }`
`70`	`70`	`void TestedDrmMemoryManager::forceLimitedRangeAllocator(uint64_t range) {`
`71`	`71`	`for (auto &gfxPartition : gfxPartitions) {`
`72`		`- gfxPartition->init(range, getSizeToReserve(), 0, 1, false, 0u);`
	`72`	`+ gfxPartition->init(range, getSizeToReserve(), 0, 1, false, 0u, range + 1);`
`73`	`73`	`}`
`74`	`74`	`}`
`75`	`75`	`void TestedDrmMemoryManager::overrideGfxPartition(GfxPartition *newGfxPartition) { gfxPartitions[0].reset(newGfxPartition); }`