Check if KMD migration is supported

Related-To: NEO-3312

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

shared/source/debug_settings/debug_variables_base.inl

@@ -214,7 +214,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, OverrideThreadArbitrationPolicy, -1, "-1 (dont o
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideAubDeviceId, -1, "-1 dont override, any other: use this value for AUB generation device id")
 DECLARE_DEBUG_VARIABLE(int32_t, EnableTimestampPacket, -1, "-1: default, 0: disable, 1:enable. Write Timestamp Packet for each set of gpu walkers")
 DECLARE_DEBUG_VARIABLE(int32_t, AllocateSharedAllocationsWithCpuAndGpuStorage, -1, "When enabled driver creates cpu & gpu storage for shared unified memory allocations. (-1 - devices default mode, 0 - disable, 1 - enable)")
-DECLARE_DEBUG_VARIABLE(int32_t, UseKmdMigration, -1, "-1: devices default mode (currently - disabled), 0: disable - pagefault handling by UMD using handler for SIGSEGV, 1: enable - pagefault handling by KMD, GEM objects migrated by KMD upon access)")
+DECLARE_DEBUG_VARIABLE(int32_t, UseKmdMigration, 0, "-1: devices default mode, 0: disable - pagefault handling by UMD using handler for SIGSEGV, 1: enable - pagefault handling by KMD, GEM objects migrated by KMD upon access)")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceSemaphoreDelayBetweenWaits, -1, "Specifies the minimum number of microseconds allowed for command streamer to wait before re-fetching the data. 0 - poll interval will be equal to the memory latency of the read completion")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceLocalMemoryAccessMode, -1, "-1: don't override, 0: default rules apply, 1: CPU can access local memory, 3: CPU never accesses local memory")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceUserptrAlignment, -1, "-1: no force (4kb), >0: n kb alignment")

shared/source/helpers/hw_helper.h

@@ -129,6 +129,7 @@ class HwHelper {
     virtual bool isSipWANeeded(const HardwareInfo &hwInfo) const = 0;
     virtual bool additionalKernelExecInfoSupported(const HardwareInfo &hwInfo) const = 0;
     virtual bool isCpuImageTransferPreferred(const HardwareInfo &hwInfo) const = 0;
+    virtual bool isKmdMigrationSupported(const HardwareInfo &hwInfo) const = 0;
 
     static uint32_t getSubDevicesCount(const HardwareInfo *pHwInfo);
     static uint32_t getEnginesCount(const HardwareInfo &hwInfo);
@@ -318,6 +319,8 @@ class HwHelperHw : public HwHelper {
 
     bool isMediaBlockIOSupported(const HardwareInfo &hwInfo) const override;
 
+    bool isKmdMigrationSupported(const HardwareInfo &hwInfo) const override;
+
     bool isCopyOnlyEngineType(EngineGroupType type) const override;
 
     void adjustAddressWidthForCanonize(uint32_t &addressWidth) const override;

shared/source/helpers/hw_helper_base.inl

@@ -515,6 +515,11 @@ bool HwHelperHw<GfxFamily>::isMediaBlockIOSupported(const HardwareInfo &hwInfo)
     return hwInfo.capabilityTable.supportsImages;
 }
 
+template <typename GfxFamily>
+bool HwHelperHw<GfxFamily>::isKmdMigrationSupported(const HardwareInfo &hwInfo) const {
+    return false;
+}
+
 template <typename GfxFamily>
 bool HwHelperHw<GfxFamily>::isCopyOnlyEngineType(EngineGroupType type) const {
     return NEO::EngineGroupType::Copy == type;

shared/source/memory_manager/memory_manager.h

@@ -145,6 +145,8 @@ class MemoryManager {
     bool isLocalMemorySupported(uint32_t rootDeviceIndex) const;
     virtual bool isMemoryBudgetExhausted() const;
 
+    virtual bool isKmdMigrationAvailable(uint32_t rootDeviceIndex) { return false; }
+
     virtual AlignedMallocRestrictions *getAlignedMallocRestrictions() {
         return nullptr;
     }

shared/source/memory_manager/unified_memory_manager.cpp

@@ -220,13 +220,9 @@ void *SVMAllocsManager::createSharedUnifiedMemoryAllocation(size_t size,
     }
 
     if (supportDualStorageSharedMemory) {
-        bool useKmdMigration = false;
-
-        if (DebugManager.flags.UseKmdMigration.get() != -1) {
-            useKmdMigration = DebugManager.flags.UseKmdMigration.get();
-        }
-
+        bool useKmdMigration = memoryManager->isKmdMigrationAvailable(*memoryProperties.rootDeviceIndices.begin());
         void *unifiedMemoryPointer = nullptr;
+
         if (useKmdMigration) {
             unifiedMemoryPointer = createUnifiedKmdMigratedAllocation(size, {}, memoryProperties);
             if (!unifiedMemoryPointer) {

shared/source/os_interface/linux/drm_memory_manager.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2017-2020 Intel Corporation
+ * Copyright (C) 2017-2021 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -175,6 +175,20 @@ void DrmMemoryManager::releaseGpuRange(void *address, size_t unmapSize, uint32_t
     gfxPartition->freeGpuAddressRange(graphicsAddress, unmapSize);
 }
 
+bool DrmMemoryManager::isKmdMigrationAvailable(uint32_t rootDeviceIndex) {
+    auto hwInfo = executionEnvironment.rootDeviceEnvironments[rootDeviceIndex]->getHardwareInfo();
+    auto &hwHelper = NEO::HwHelper::get(hwInfo->platform.eRenderCoreFamily);
+
+    auto useKmdMigration = hwHelper.isKmdMigrationSupported(*hwInfo) &&
+                           this->getDrm(rootDeviceIndex).isVmBindAvailable();
+
+    if (DebugManager.flags.UseKmdMigration.get() != -1) {
+        useKmdMigration = DebugManager.flags.UseKmdMigration.get();
+    }
+
+    return useKmdMigration;
+}
+
 NEO::BufferObject *DrmMemoryManager::allocUserptr(uintptr_t address, size_t size, uint64_t flags, uint32_t rootDeviceIndex) {
     drm_i915_gem_userptr userptr = {};
     userptr.user_ptr = address;

shared/source/os_interface/linux/drm_memory_manager.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2017-2020 Intel Corporation
+ * Copyright (C) 2017-2021 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -62,6 +62,8 @@ class DrmMemoryManager : public MemoryManager {
     void freeGpuAddress(AddressRange addressRange, uint32_t rootDeviceIndex) override;
     MOCKABLE_VIRTUAL BufferObject *createBufferObjectInMemoryRegion(Drm *drm, uint64_t gpuAddress, size_t size, uint32_t memoryBanks, size_t maxOsContextCount);
 
+    bool isKmdMigrationAvailable(uint32_t rootDeviceIndex) override;
+
     std::unique_lock<std::mutex> acquireAllocLock();
     std::vector<GraphicsAllocation *> &getSysMemAllocs();
     std::vector<GraphicsAllocation *> &getLocalMemAllocs(uint32_t rootDeviceIndex);