为什么read()的调用会永远阻塞

Question

我正面临一个最近才开始的令人费解的问题.

我有一个程序,它使用一个线程写入文件,另一个线程从该文件中读取.两个线程都使用不同的文件描述符.写入程序线程使用O_WRONLY标志打开文件,读取器线程以O_RDONLY模式打开文件.就逻辑而言,读者线程不知道编写器线程正在做什么,并且两者都可以使用不同的文件.

写入程序线程以固定间隔连续写入文件(数据来自设备流,速度高达20Mbit/s).

读者线程也定期读取文件.

这是读者循环:

while (tot < sz)
{
    LOG(VB_FILE, LOG_DEBUG, LOC +
        QString("read(%1) -- begin").arg(sz-tot));
    ret = read(fd2, (char *)data + tot, sz - tot);
    LOG(VB_FILE, LOG_DEBUG, LOC +
        QString("read(%1) -> %2 end").arg(sz).arg(ret));

    if ((sz - tot) != ret)
    {
        LOG(VB_FILE, LOG_DEBUG, LOC + QString("errno = %1").arg(errno));
    }

    if (ret < 0)
    {
        if (errno == EAGAIN)
        {
            LOG(VB_FILE, LOG_DEBUG, LOC +
                QString("read(%1) -> %2 EAGAIN").arg(sz).arg(ret));
            usleep(1000);
            continue;
        }

        LOG(VB_GENERAL, LOG_ERR,
            LOC + "File I/O problem in 'safe_read()'" + ENO);

        errcnt++;
        numfailures++;
        if (errcnt == 3)
            break;
    }
    else if (ret > 0)
    {
        tot += ret;
    }
    [...snipped...]
}

您可以看到我在调用read之前显示日志,并在返回之后立即显示.阅读将不时被调用,它永远不会回来......

2014-02-19 11:24:10.156417 D  TFW(/external/recordings/1001_20140219002351.mpg:64): write(65424) cnt 1 total 5076
2014-02-19 11:24:10.156466 D  TFW(/external/recordings/1001_20140219002351.mpg:64): total written so far: 26934760 bytes
2014-02-19 11:24:10.156514 D  FileRingBuf(/external/recordings/1001_20140219002351.mpg): read(65536) -- begin
2014-02-19 11:24:10.190769 D  FileRingBuf(/external/recordings/1001_20140219002351.mpg): read(65536) -> 60968 end
2014-02-19 11:24:10.190781 I  RingBuf(/external/recordings/1001_20140219002351.mpg): safe_read(...@1698944, 65536) -> 65536, took 60 ms (8.73813Mbps)
2014-02-19 11:24:10.190786 D  RingBuf(/external/recordings/1001_20140219002351.mpg): total read so far: 26930304 bytes
2014-02-19 11:24:10.190795 I  FileRingBuf(/external/recordings/1001_20140219002351.mpg): read(65536) -- begin
2014-02-19 11:24:10.195917 D  FileRingBuf(/external/recordings/1001_20140219002351.mpg): read(65536) -> 4456 end
2014-02-19 11:24:10.195927 D  FileRingBuf(/external/recordings/1001_20140219002351.mpg): errno = 0
2014-02-19 11:24:10.206445 D  TFW(/external/recordings/1001_20140219002351.mpg:64): write(65424) cnt 1 total 1692
2014-02-19 11:24:10.206489 D  TFW(/external/recordings/1001_20140219002351.mpg:64): total written so far: 27000184 bytes
2014-02-19 11:24:10.256103 D  FileRingBuf(/external/recordings/1001_20140219002351.mpg): read(61080) -- begin
2014-02-19 11:24:10.256499 D  TFW(/external/recordings/1001_20140219002351.mpg:64): write(47376) cnt 1 total 40984
2014-02-19 11:24:10.262073 D  TFW(/external/recordings/1001_20140219002351.mpg:64): total written so far: 27047560 bytes
2014-02-19 11:24:10.273385 D  TFW(/external/recordings/1001_20140219002351.mpg:64): write(65424) cnt 1 total 940
2014-02-19 11:24:10.385495 D  TFW(/external/recordings/1001_20140219002351.mpg:64): total written so far: 27112984 bytes

你可以在这里看到编写器已经向磁盘写了26934760个字节.到目前为止读取的读数为26930304字节,因此我们从EOF读取4456字节.然后尝试64kB读取,读取几乎立即返回4456字节.到现在为止还挺好.立即尝试另一次读取61080字节(65536-4456).

不久之后,编写器线程再次写入文件.64kB读取现在正在等待,并且不会再出现30秒.

所以关于为什么读取会突然阻塞的任何特定想法？

编辑:从查看行为开始,一旦读取达到EOF并且提前返回,如果在新写入发生之前立即重试读取,则阻塞似乎总是发生.在这种情况下,读取将不会退出几秒钟(通常为20+)

Answer 1

好的...

我已经发现了这个问题以及如何解决它。

正如原始问题中提到的，一旦读取达到 EOF、提前返回并立即重试读取（在文件发生新写入之前），阻塞似乎总是会发生。

在这种情况下，read()将在几秒钟内（通常超过20秒）不会退出

因此，解决方法是记录到目前为止我们已读取的字节数，以便它知道它在文件中的位置，并调用 fstat 来检查文件的大小。由此，如果我们已经到达文件末尾，请确保我们永远不会调用 read() ，或者要求 read() 检索比文件中更多的字节。

struct stat sb;
off_t current_pos = internalreadpos;

while (tot < sz)
{
    off_t toread = sz - tot;
    bool read_ok = true;

    // check that we have some data to read,
    // so we never attempt to read past the end of file
    // if fstat errored or isn't a regular file, default to previous behavior
    ret = fstat(fd2, &sb);
    if (ret == 0 && S_ISREG(sb.st_mode))
    {
        if (current_pos >= sb.st_size)
        {
            // We're at the end, don't attempt to read
            read_ok = false;
            LOG(VB_FILE, LOG_DEBUG, LOC + "not reading, reached EOF");
        }
        else
        {
            toread = min(sb.st_size - current_pos, toread);
            if (toread < (sz-tot))
            {
                LOG(VB_FILE, LOG_DEBUG,
                    LOC + QString("About to reach EOF, reading %1 wanted %2")
                    .arg(toread).arg(sz-tot));
            }
        }
    }

    if (read_ok)
    {
        LOG(VB_FILE, LOG_DEBUG, LOC +
            QString("read(%1) -- begin").arg(toread));
        ret = read(fd2, (char *)data + tot, toread);
        LOG(VB_FILE, LOG_DEBUG, LOC +
            QString("read(%1) -> %2 end").arg(toread).arg(ret));
    }
    if (ret < 0)
    {
        if (errno == EAGAIN)
            continue;

        LOG(VB_GENERAL, LOG_ERR,
            LOC + "File I/O problem in 'safe_read()'" + ENO);

        errcnt++;
        numfailures++;
        if (errcnt == 3)
            break;
    }
    else if (ret > 0)
    {
        tot += ret;
        current_pos += ret;
    }

    if (oldfile)
        break;

    if (ret == 0) // EOF returns 0
    {
        if (tot > 0)
            break;

        zerocnt++;

        // 0.36 second timeout for livetvchain with usleep(60000),
        // or 2.4 seconds if it's a new file less than 30 minutes old.
        if (zerocnt >= (livetvchain ? 6 : 40))
        {
            break;
        }
    }