Files
compute-runtime/shared/source/command_stream/tbx_command_stream_receiver_hw.h
Jack Myers 7f9fadc314 fix: regression caused by tbx fault mngr
Addresses regressions from the reverted merge
of the tbx fault manager for host memory.

Recursive locking of mutex caused deadlock.

To fix, separate tbx fault data from base
cpu fault data, allowing separate mutexes
for each, eliminating recursive locks on
the same mutex.

By separating, we also help ensure that tbx-related
changes don't affect the original cpu fault manager code
paths.

As an added safe guard preventing critical regressions
and avoiding another auto-revert, the tbx fault manager
is hidden behind a new debug flag which is disabled by default.

Related-To: NEO-12268
Signed-off-by: Jack Myers <jack.myers@intel.com>
2025-01-09 07:48:53 +01:00

118 lines
5.4 KiB
C++

/*
* Copyright (C) 2018-2025 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#pragma once
#include "shared/source/command_stream/command_stream_receiver_simulated_hw.h"
#include "shared/source/command_stream/tbx_command_stream_receiver.h"
#include "shared/source/command_stream/wait_status.h"
#include "shared/source/memory_manager/address_mapper.h"
#include "shared/source/memory_manager/page_table.h"
#include <set>
namespace NEO {
class AubSubCaptureManager;
class TbxStream;
class CpuPageFaultManager;
template <typename GfxFamily>
class TbxCommandStreamReceiverHw : public CommandStreamReceiverSimulatedHw<GfxFamily> {
protected:
typedef CommandStreamReceiverSimulatedHw<GfxFamily> BaseClass;
using AUB = typename AUBFamilyMapper<GfxFamily>::AUB;
using BaseClass::forceSkipResourceCleanupRequired;
using BaseClass::getParametersForMemory;
using BaseClass::osContext;
using BaseClass::pollForCompletion;
uint32_t getMaskAndValueForPollForCompletion() const;
bool getpollNotEqualValueForPollForCompletion() const;
void flushSubmissionsAndDownloadAllocations(TaskCountType taskCount, bool skipAllocationsDownload);
MOCKABLE_VIRTUAL uint64_t getNonBlockingDownloadTimeoutMs() const {
return 2000; // 2s
}
bool isAllocTbxFaultable(GraphicsAllocation *gfxAlloc);
void registerAllocationWithTbxFaultMngrIfTbxFaultable(GraphicsAllocation *gfxAllocation, void *cpuAddress, size_t size);
void allowCPUMemoryAccessIfTbxFaultable(GraphicsAllocation *gfxAllocation, void *cpuAddress, size_t size);
void protectCPUMemoryAccessIfTbxFaultable(GraphicsAllocation *gfxAllocation, void *cpuAddress, size_t size);
void protectCPUMemoryFromWritesIfTbxFaultable(GraphicsAllocation *gfxAllocation, void *cpuAddress, size_t size);
public:
using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::initAdditionalMMIO;
using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::aubManager;
using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::hardwareContextController;
using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::engineInfo;
using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::stream;
using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::peekExecutionEnvironment;
using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::writeMemory;
SubmissionStatus flush(BatchBuffer &batchBuffer, ResidencyContainer &allocationsForResidency) override;
WaitStatus waitForTaskCountWithKmdNotifyFallback(TaskCountType taskCountToWait, FlushStamp flushStampToWait, bool useQuickKmdSleep, QueueThrottle throttle) override;
WaitStatus waitForCompletionWithTimeout(const WaitParams &params, TaskCountType taskCountToWait) override;
void downloadAllocations(bool blockingWait, TaskCountType taskCount) override;
void downloadAllocationTbx(GraphicsAllocation &gfxAllocation);
void removeDownloadAllocation(GraphicsAllocation *alloc) override;
void processEviction() override;
SubmissionStatus processResidency(ResidencyContainer &allocationsForResidency, uint32_t handleId) override;
void writeMemory(uint64_t gpuAddress, void *cpuAddress, size_t size, uint32_t memoryBank, uint64_t entryBits) override;
bool writeMemory(GraphicsAllocation &gfxAllocation, bool isChunkCopy, uint64_t gpuVaChunkOffset, size_t chunkSize) override;
void writeMMIO(uint32_t offset, uint32_t value) override;
bool expectMemory(const void *gfxAddress, const void *srcAddress, size_t length, uint32_t compareOperation) override;
AubSubCaptureStatus checkAndActivateAubSubCapture(const std::string &kernelName) override;
// Family specific version
MOCKABLE_VIRTUAL void submitBatchBufferTbx(uint64_t batchBufferGpuAddress, const void *batchBuffer, size_t batchBufferSize, uint32_t memoryBank, uint64_t entryBits, bool overrideRingHead);
void pollForCompletion(bool skipTaskCountCheck) override;
void dumpAllocation(GraphicsAllocation &gfxAllocation) override;
static CommandStreamReceiver *create(const std::string &baseName,
bool withAubDump,
ExecutionEnvironment &executionEnvironment,
uint32_t rootDeviceIndex,
const DeviceBitfield deviceBitfield);
TbxCommandStreamReceiverHw(ExecutionEnvironment &executionEnvironment,
uint32_t rootDeviceIndex,
const DeviceBitfield deviceBitfield);
~TbxCommandStreamReceiverHw() override;
void initializeEngine() override;
MOCKABLE_VIRTUAL MemoryManager *getMemoryManager() {
return CommandStreamReceiver::getMemoryManager();
}
MOCKABLE_VIRTUAL CpuPageFaultManager *getTbxPageFaultManager();
TbxStream tbxStream;
std::unique_ptr<AubSubCaptureManager> subCaptureManager;
uint32_t aubDeviceId;
bool streamInitialized = false;
std::unique_ptr<PhysicalAddressAllocator> physicalAddressAllocator;
std::unique_ptr<std::conditional<is64bit, PML4, PDPE>::type> ppgtt;
std::unique_ptr<PDPE> ggtt;
// remap CPU VA -> GGTT VA
AddressMapper gttRemap;
std::set<GraphicsAllocation *> allocationsForDownload = {};
CommandStreamReceiverType getType() const override {
return CommandStreamReceiverType::tbx;
}
bool dumpTbxNonWritable = false;
bool isEngineInitialized = false;
};
} // namespace NEO