Fixes for AUBs

Change-Id: Iac55927eb96db8dd68b86d21e66392039ba1f058

Author: HoppeMateusz

runtime/command_stream/aub_command_stream_receiver_hw.inl

@@ -450,7 +450,7 @@ FlushStamp AUBCommandStreamReceiverHw<GfxFamily>::flush(BatchBuffer &batchBuffer
             physLRCA + 0x101c,
             &engineInfo.tailRingBuffer,
             sizeof(engineInfo.tailRingBuffer),
-            getAddressSpace(AubMemDump::DataTypeHintValues::TraceNotype));
+            getAddressSpace(getCsTraits(engineType).aubHintLRCA));
 
         DEBUG_BREAK_IF(engineInfo.tailRingBuffer >= engineInfo.sizeRingBuffer);
     }

runtime/memory_manager/page_table.cpp

@@ -19,16 +19,17 @@ uintptr_t PTE::map(uintptr_t vm, size_t size, uint64_t entryBits, uint32_t memor
     uintptr_t res = -1;
     bool updateEntryBits = entryBits != PageTableEntry::nonValidBits;
     uint64_t newEntryBits = entryBits & 0xfff;
+    auto entriesMask = std::numeric_limits<uintptr_t>::max() & ~MemoryConstants::pageMask;
     newEntryBits |= 0x1;
 
     for (size_t index = indexStart; index <= indexEnd; index++) {
         if (entries[index] == 0x0) {
             uint64_t tmp = allocator->reservePage(memoryBank);
             entries[index] = reinterpret_cast<void *>(tmp | newEntryBits);
         } else if (updateEntryBits) {
-            entries[index] = reinterpret_cast<void *>((reinterpret_cast<uintptr_t>(entries[index]) & 0xfffff000u) | newEntryBits);
+            entries[index] = reinterpret_cast<void *>((reinterpret_cast<uintptr_t>(entries[index]) & entriesMask) | newEntryBits);
         }
-        res = std::min(reinterpret_cast<uintptr_t>(entries[index]) & 0xfffff000u, res);
+        res = std::min(reinterpret_cast<uintptr_t>(entries[index]) & entriesMask, res);
     }
     return (res & ~newEntryBits) + (vm & (pageSize - 1));
 }
@@ -43,16 +44,17 @@ void PTE::pageWalk(uintptr_t vm, size_t size, size_t offset, uint64_t entryBits,
     uintptr_t rem = vm & (pageSize - 1);
     bool updateEntryBits = entryBits != PageTableEntry::nonValidBits;
     uint64_t newEntryBits = entryBits & 0xfff;
+    auto entriesMask = std::numeric_limits<uintptr_t>::max() & ~MemoryConstants::pageMask;
     newEntryBits |= 0x1;
 
     for (size_t index = indexStart; index <= indexEnd; index++) {
         if (entries[index] == 0x0) {
             uint64_t tmp = allocator->reservePage(memoryBank);
             entries[index] = reinterpret_cast<void *>(tmp | newEntryBits);
         } else if (updateEntryBits) {
-            entries[index] = reinterpret_cast<void *>((reinterpret_cast<uintptr_t>(entries[index]) & 0xfffff000u) | newEntryBits);
+            entries[index] = reinterpret_cast<void *>((reinterpret_cast<uintptr_t>(entries[index]) & entriesMask) | newEntryBits);
         }
-        res = reinterpret_cast<uintptr_t>(entries[index]) & 0xfffff000u;
+        res = reinterpret_cast<uintptr_t>(entries[index]) & entriesMask;
 
         size_t lSize = std::min(pageSize - rem, size);
         pageWalker((res & ~0x1) + rem, lSize, offset, reinterpret_cast<uintptr_t>(entries[index]) & 0xfffu);

unit_tests/aub_tests/command_queue/enqueue_write_copy_read_buffer_aub_tests.h

@@ -18,7 +18,7 @@ struct AubWriteCopyReadBuffer : public AUBFixture,
                                 public ::testing::Test {
 
     void SetUp() override {
-        AUBFixture::SetUp();
+        AUBFixture::SetUp(nullptr);
     }
 
     void TearDown() override {

unit_tests/aub_tests/fixtures/aub_fixture.h

@@ -24,8 +24,8 @@ namespace OCLRT {
 
 class AUBFixture : public CommandQueueHwFixture {
   public:
-    void SetUp() {
-        const HardwareInfo &hwInfo = *platformDevices[0];
+    void SetUp(const HardwareInfo *hardwareInfo) {
+        const HardwareInfo &hwInfo = hardwareInfo ? *hardwareInfo : *platformDevices[0];
         uint32_t deviceIndex = 0;
 
         const ::testing::TestInfo *const testInfo = ::testing::UnitTest::GetInstance()->current_test_info();
@@ -71,5 +71,8 @@ class AUBFixture : public CommandQueueHwFixture {
     std::unique_ptr<MockDevice> device;
 
     ExecutionEnvironment *executionEnvironment;
+
+  private:
+    using CommandQueueHwFixture::SetUp;
 };
 } // namespace OCLRT

unit_tests/memory_manager/page_table_tests.cpp

@@ -215,6 +215,38 @@ TEST_F(PageTableTests48, givenReservedPhysicalAddressWhenPageWalkIsCalledThenPag
     }
 }
 
+TEST_F(PageTableTests48, givenBigGpuAddressWhenPageWalkIsCalledThenPageTablesAreFilledWithProperAddresses) {
+    if (is64Bit) {
+        std::unique_ptr<MockPML4> pageTable(std::make_unique<MockPML4>(&allocator));
+
+        int shiftPML4 = is64Bit ? (47) : 0;
+        int shiftPDP = is64Bit ? (9 + 9 + 12) : 0;
+
+        uintptr_t gpuVa = (uintptr_t(0x1) << (shiftPML4)) | (uintptr_t(0x1) << (shiftPDP)) | (uintptr_t(0x1) << (9 + 12)) | 0x100;
+
+        size_t size = 10 * pageSize;
+
+        size_t walked = 0u;
+        auto address = allocator.mainAllocator.load();
+
+        PageWalker walker = [&](uint64_t physAddress, size_t size, size_t offset, uint64_t entryBits) {
+            walked += size;
+        };
+        pageTable->pageWalk(gpuVa, size, 0, 0, walker, MemoryBanks::MainBank);
+
+        EXPECT_EQ(size, walked);
+
+        ASSERT_NE(nullptr, pageTable->entries[0x100]);
+        ASSERT_NE(nullptr, pageTable->entries[0x100]->entries[1]);
+        ASSERT_NE(nullptr, pageTable->entries[0x100]->entries[1]->entries[1]);
+
+        for (uint32_t i = 0; i < 10; i++) {
+            EXPECT_EQ(reinterpret_cast<void *>(address | 0x1), pageTable->entries[0x100]->entries[1]->entries[1]->entries[i]);
+            address += pageSize;
+        }
+    }
+}
+
 TEST_F(PageTableTests48, givenZeroEntryBitsWhenPageWalkIsCalledThenPageTableEntryHasPresentBitSet) {
     std::unique_ptr<TypeSelector<MockPML4, MockPDPE, sizeof(void *) == 8>::type>
         pageTable(std::make_unique<TypeSelector<MockPML4, MockPDPE, sizeof(void *) == 8>::type>(&allocator));