异步I/O回调方法并发

Question

我已经在我担心并发的情况下实现了异步I/O。我与您竞争，因为我一直在使用同一文件，并且OS文件 y i/o从根本上是同步操作，那么我在回调方法中不需要锁定机制 - 但是我很可能在这里错了 - 输入：o）我有一个缓冲管理器，该缓冲区管理器将读取数据放入其缓冲区缓存中后，并且根据EoverPappedStates枚举状态，每个重叠操作的状态发动机；" I/O未开始"，"成功"one_answers"错误"。您是否认为我需要锁定回调方法，以确保像我们这样的多线程程序中并发？

打开文件：

OS_FILE_HANDLE CUniformDiskInterface::OpenFile(const wchar_t *fileName, bool *fileExists, bool readData, bool writeData, bool overlap, 
bool disableDiskCache, bool disableOsCache, bool randomAccess, bool sequentalScan) {
// Set access method
DWORD desiredAccess = readData ? GENERIC_READ : 0;
desiredAccess |= writeData ? GENERIC_WRITE : 0;
// Set file flags
DWORD fileFlags = disableDiskCache ? FILE_FLAG_WRITE_THROUGH : 0;
fileFlags |= disableOsCache ? FILE_FLAG_NO_BUFFERING : 0;
fileFlags |= randomAccess ? FILE_FLAG_RANDOM_ACCESS : 0;
fileFlags |= sequentalScan ? FILE_FLAG_SEQUENTIAL_SCAN : 0;
fileFlags |= !fileFlags ? FILE_ATTRIBUTE_NORMAL : 0;
fileFlags |= overlap ? FILE_FLAG_OVERLAPPED : 0;
HANDLE hOutputFile = CreateFile(
    fileName,
    desiredAccess,
    0,
    NULL,
    OPEN_EXISTING,
    fileFlags,
    NULL);

阅读文件：

_UINT64 CUniformDiskInterface::ReadFromFile(OS_FILE_HANDLE hFile, void *outData, _UINT64 bytesToRead, OVERLAPPED *overlapped, LPOVERLAPPED_COMPLETION_ROUTINE completionRoutine) {
DWORD wBytesRead = 0;
BOOL result = completionRoutine ? 
    ReadFileEx(hFile, outData, (DWORD)(bytesToRead), overlapped, completionRoutine) : 
    ReadFile(hFile, outData, (DWORD)(bytesToRead), &wBytesRead, overlapped);
if (!result)
{
    int errorCode = GetLastError();
    if (errorCode != ERROR_IO_PENDING )
    {
        wstringstream err(wstringstream::in | wstringstream::out);
        err << L"Can't read sectors from file. [ReadFile] error #" << errorCode << L".";
        throw new FileIOException(L"CUniformDiskInterface", L"ReadFromFile", err.str().c_str(), GETDATE, GETFILE, GETLINE); 
    }
}
return (_UINT64)wBytesRead; }

扩展的重叠结构：

            /*!
        enum EOverlappedStates
        brief The different overlapped states
        details Used as inter-thread communication while waiting for the I/O operation to complete 
        */
        enum EOverlappedStates
        {
            /** The I/O operation has not started or in in-progress */
            EOverlappedNotStarted, 
            /** The I/O operation is done and was successful */
            EOverlappedSuccess, 
            /** The I/O operation is done but there was an error */
            EOverlappedError
        };
        /*!
        struct OverlappedEx
        brief Extended overlapped structure
        */
        struct OverlappedEx : OVERLAPPED
        {           
            /** The buffer manager that is designated to cache the record when it's loaded */
            CBufferManager *bufferManger;
            /** Transaction ID related to this disk I/O operation */
            _UINT64 transactionId;
            /** Start disk sector of the record */
            _UINT64 startDiskSector;
            /** Buffer */
            void *buffer;
            /** Number of bytes in c buffer */
            _UINT64 bufferSize;
            /** Current overlapped I/O state. Used for inter-thread communication while waiting for the I/O to complete */
            EOverlappedStates state;
            /** Error code, or c 0 if no error */
            _UINT32 errorCode;
        };

回调方法：

    /*! brief Callback routine after a overlapped read has completed
details Fills the buffer managers buffer cache with the read data
todo This callback method may be a bottleneck, so look into how to handle this better
*/
VOID WINAPI CompletedReadRoutine(DWORD dwErr, DWORD cbBytesRead, LPOVERLAPPED lpOverLap) 
{ 
    OverlappedEx *overlapped = (OverlappedEx*)lpOverLap;
    overlapped->errorCode = (_UINT32)dwErr;
    if (!dwErr && cbBytesRead) 
    { 
        overlapped->state = EOverlappedSuccess;
        overlapped->bufferManger->AddBuffer(overlapped->startDiskSector, overlapped->buffer, overlapped->bufferSize);
    }
    else
    {
        // An error occurred
        overlapped->state = EOverlappedError;
    }
} 

用法：

    _UINT64 startDiskSector = location / sectorByteSize;
void *buffer = bufferManager->GetBuffer(startDiskSector);
if (!buffer)
{
    /*
    The disk sector was not cached, so get the data from the disk and cache in internal memory with
    the buffer manager
    */
    buffer = new char[recordByteSize];
    // Create a overlapped structure to enable disk async I/O operations
    OverlappedEx *overlapped = new OverlappedEx;
    memset(overlapped, 0, sizeof(OverlappedEx));
    overlapped->Offset = (DWORD)(startDiskSector & 0xffffffffULL);
    overlapped->OffsetHigh = (DWORD)(startDiskSector >> 31ULL);
    overlapped->hEvent = CreateEvent(NULL, TRUE, FALSE, NULL);
    overlapped->bufferManger = bufferManager;
    overlapped->startDiskSector = startDiskSector;
    overlapped->buffer = buffer;
    overlapped->bufferSize = recordByteSize;
    overlapped->state = EOverlappedNotStarted;
    // Read from disk
    diskApi.ReadFromFile(fileHandle, buffer, sectorByteSize, overlapped, CompletedReadRoutine);
    return overlapped;
}

Answer 1

根据MSDN上的文档，回调函数只能在称为ReadFileEx函数的同一线程上调用，并且仅当线程在等待事件发生时。因此，保证呼叫ReadFileEx与回调的调用之间没有同步问题。

这意味着无需同步对OverlappedEx数据结构的访问，只要只有一个线程尝试读取该结构的特定实例，这相当于仅从一个线程中读取特定文件。如果您尝试从多个线程读取单个文件，则很可能会在Windows本身中遇到问题（我认为异步I/O本身是线程安全的），因此锁定Mutex不会帮助您在这种情况下。