How kill child process without read() hanging in the parent process?

somanyquestions · 10-06-2021, 07:16 AM

There is an external program I'm calling from within my C/C++ program, by using fork() and execl(), and also redirecting stdio with dup2() to be able to read the output from the external program. It's based on this code: A simple popen2 implementation

In another thread just started, the timeout handling is done.
Then I make calls to read() on the out file handle, in a loop, writing what's read to a stream.
Now that external program waits for something on the system to happen - which might not happen, so after a while, I'd like to call it a timeout and end the child process (from the extra thread).
Whether sending SIGTERM or SIGKILL, changing the order of closing file handles w.r.t. to that, the result is always that the last call to read() hangs forever.

Added: Seems even if the process exits in time, the read() hangs. That did not happen when I used the regular popen() to start the process instead of popen2(), but since I need the PID to terminate the process, I'm using the latter.

I have found discussions about a problem with read() in a somewhat similar situation, and it was suggested that the SIGCHLD must be ignored in the parent process. That doesn't help, though.

Is there something wrong per se in doing things this way, that cause this read() to hang?

Code:

int exec_proc2(std::string cmd, std::ostream & outputStream, unsigned timeoutMilsecs)
{
    constexpr int bufsize = 128;
    std::array<char, bufsize> buffer;
    
    int inf, outf;
    auto pid = popen2( cmd.c_str(), &inf, &outf );
    close( inf ); // writing to process not used, we're only reading
    if (pid < 0)
    {   close( outf );
        return -1;
    }
    
    std::atomic<bool> notDone = true;
    std::atomic<bool> timedOut = false;
    thread timeoutTask( [&]
        {   TimeoutWatch tw(timeoutMilsecs, true);
            while (notDone && !tw.TimedOut())
            {   sleep(0.1);
            }
            if (tw.TimedOut())
            {   close( outf );
                int ret = kill( pid, SIGKILL );
                if (ret != 0)
                {   // this never failed so far
                }
                notDone = false;
                cout << "DBG: TIMEOUT@exec_proc2" << endl; // this is always printed after the timeout time
                timedOut = true;
            }
        } );
    
    size_t count;
    do
    {   if ((count = read( outf, buffer.data(), bufsize )) > 0)
  	    {   outputStream.write( buffer.data(), count );
  	    }
    } while(count > 0 && !timedOut);
    cout << "DBG: joining...@exec_proc2" << endl; // this line is never printed, so it hangs above already
    notDone = false; // for the case that the read() loop exits earlier than the timeout thread: tell the thread it's over
    timeoutTask.join();
    
    close( outf );
    return timedOut ? -2 : 0;
}

NevemTeve · 10-06-2021, 07:53 AM

This popen2 might be incomplete.
After fork, when pid!=0:

Code:

    close (p_stdin[READ]);
    if (infp == NULL)
        close(p_stdin[WRITE]);
    else
        *infp = p_stdin[WRITE];

    close (p_stdout[WRITE]);
    if (outfp == NULL)
        close(p_stdout[READ]);
    else
        *outfp = p_stdout[READ];

somanyquestions · 10-06-2021, 11:22 AM

Quote:

Originally Posted by NevemTeve

This popen2 might be incomplete.
After fork, when pid!=0:

Code:

    close (p_stdin[READ]);
    if (infp == NULL)
        close(p_stdin[WRITE]);
    else
        *infp = p_stdin[WRITE];

    close (p_stdout[WRITE]);
    if (outfp == NULL)
        close(p_stdout[READ]);
    else
        *outfp = p_stdout[READ];

Thanks! Yes you're right. I meanwhile found this here, and the accepted answer mentions the same:
example at SO