多处理：我怎样才能 ʀᴇʟɪᴀʙʟʏ 从子进程重定向标准输出？

Question

NB. I have seen Log output of multiprocessing.Process - unfortunately, it doesn't answer this question.

I am creating a child process (on windows) via multiprocessing. I want all of the child process's stdout and stderr output to be redirected to a log file, rather than appearing at the console. The only suggestion I have seen is for the child process to set sys.stdout to a file. However, this does not effectively redirect all stdout output, due to the behaviour of stdout redirection on Windows.

To illustrate the problem, build a Windows DLL with the following code

#include <iostream>

extern "C"
{
    __declspec(dllexport) void writeToStdOut()
    {
        std::cout << "Writing to STDOUT from test DLL" << std::endl;
    }
}

Then create and run a python script like the following, which imports this DLL and calls the function:

from ctypes import *
import sys

print
print "Writing to STDOUT from python, before redirect"
print
sys.stdout = open("stdout_redirect_log.txt", "w")
print "Writing to STDOUT from python, after redirect"

testdll = CDLL("Release/stdout_test.dll")
testdll.writeToStdOut()

In order to see the same behaviour as me, it is probably necessary for the DLL to be built against a different C runtime than than the one Python uses. In my case, python is built with Visual Studio 2010, but my DLL is built with VS 2005.

The behaviour I see is that the console shows:

> stdout_test.py

Writing to STDOUT from python, before redirect

Writing to STDOUT from test DLL

While the file stdout_redirect_log.txt ends up containing:

Writing to STDOUT from python, after redirect

In other words, setting sys.stdout failed to redirect the stdout output generated by the DLL. This is unsurprising given the nature of the underlying APIs for stdout redirection in Windows. I have encountered this problem at the native/C++ level before and never found a way to reliably redirect stdout from within a process. It has to be done externally.

This is actually the very reason I am launching a child process - it's so that I can connect externally to its pipes and thus guarantee that I am intercepting all of its output. I can definitely do this by launching the process manually with pywin32, but I would very much like to be able to use the facilities of multiprocessing, in particular the ability to communicate with the child process via a multiprocessing Pipe object, in order to get progress updates. The question is whether there is any way to both use multiprocessing for its IPC facilities and to reliably redirect all of the child's stdout and stderr output to a file.

UPDATE: Looking at the source code for multiprocessing.Processs, it has a static member, _Popen, which looks like it can be used to override the class used to create the process. If it's set to None (default), it uses a multiprocessing.forking._Popen, but it looks like by saying

multiprocessing.Process._Popen = MyPopenClass

I could override the process creation. However, although I could derive this from multiprocessing.forking._Popen, it looks like I would have to copy a bunch of internal stuff into my implementation, which sounds flaky and not very future-proof. If that's the only choice I think I'd probably plump for doing the whole thing manually with pywin32 instead.

Answer 1

The solution you suggest is a good one: create your processes manually such that you have explicit access to their stdout/stderr file handles. You can then create a socket to communicate with the sub-process and use multiprocessing.connection over that socket (multiprocessing.Pipe creates the same type of connection object, so this should give you all the same IPC functionality).

Here's a two-file example.

master.py:

import multiprocessing.connection
import subprocess
import socket
import sys, os

## Listen for connection from remote process (and find free port number)
port = 10000
while True:
    try:
        l = multiprocessing.connection.Listener(('localhost', int(port)), authkey="secret")
        break
    except socket.error as ex:
        if ex.errno != 98:
            raise
        port += 1  ## if errno==98, then port is not available.

proc = subprocess.Popen((sys.executable, "subproc.py", str(port)), stdout=subprocess.PIPE, stderr=subprocess.PIPE)

## open connection for remote process
conn = l.accept()
conn.send([1, "asd", None])
print(proc.stdout.readline())

subproc.py:

import multiprocessing.connection
import subprocess
import sys, os, time

port = int(sys.argv[1])
conn = multiprocessing.connection.Client(('localhost', port), authkey="secret")

while True:
    try:
        obj = conn.recv()
        print("received: %s\n" % str(obj))
        sys.stdout.flush()
    except EOFError:  ## connection closed
        break

You may also want to see the first answer to this question to get non-blocking reads from the subprocess.

Answer 2

I don't think you have a better option than redirecting a subprocess to a file as you mentioned in your comment.

The way consoles stdin/out/err work in windows is each process when it's born has its std handles defined. You can change them with SetStdHandle . When you modify python's sys.stdout you only modify where python prints out stuff, not where other DLL's are printing stuff. Part of the CRT in your DLL is using GetStdHandle to find out where to print out to. If you want, you can do whatever piping you want in windows API in your DLL or in your python script with pywin32. Though I do think it'll be simpler with subprocess .

Answer 3

In my situation I changed sys.stdout.write to write to a PySide QTextEdit. I couldn't read from sys.stdout and I didn't know how to change sys.stdout to be readable. I created two Pipes. One for stdout and the other for stderr. In the separate process I redirect sys.stdout and sys.stderr to the child connection of the multiprocessing pipe. On the main process I created two threads to read the stdout and stderr parent pipe and redirect the pipe data to sys.stdout and sys.stderr .

import sys
import contextlib
import threading
import multiprocessing as mp
import multiprocessing.queues
from queue import Empty
import time


class PipeProcess(mp.Process):
    """Process to pipe the output of the sub process and redirect it to this sys.stdout and sys.stderr.

    Note:
        The use_queue = True argument will pass data between processes using Queues instead of Pipes. Queues will
        give you the full output and read all of the data from the Queue. A pipe is more efficient, but may not
        redirect all of the output back to the main process.
    """
    def __init__(self, group=None, target=None, name=None, args=tuple(), kwargs={}, *_, daemon=None,
                 use_pipe=None, use_queue=None):
        self.read_out_th = None
        self.read_err_th = None
        self.pipe_target = target
        self.pipe_alive = mp.Event()

        if use_pipe or (use_pipe is None and not use_queue):  # Default
            self.parent_stdout, self.child_stdout = mp.Pipe(False)
            self.parent_stderr, self.child_stderr = mp.Pipe(False)
        else:
            self.parent_stdout = self.child_stdout = mp.Queue()
            self.parent_stderr = self.child_stderr = mp.Queue()

        args = (self.child_stdout, self.child_stderr, target) + tuple(args)
        target = self.run_pipe_out_target

        super(PipeProcess, self).__init__(group=group, target=target, name=name, args=args, kwargs=kwargs,
                                          daemon=daemon)

    def start(self):
        """Start the multiprocess and reading thread."""
        self.pipe_alive.set()
        super(PipeProcess, self).start()

        self.read_out_th = threading.Thread(target=self.read_pipe_out,
                                            args=(self.pipe_alive, self.parent_stdout, sys.stdout))
        self.read_err_th = threading.Thread(target=self.read_pipe_out,
                                            args=(self.pipe_alive, self.parent_stderr, sys.stderr))
        self.read_out_th.daemon = True
        self.read_err_th.daemon = True
        self.read_out_th.start()
        self.read_err_th.start()

    @classmethod
    def run_pipe_out_target(cls, pipe_stdout, pipe_stderr, pipe_target, *args, **kwargs):
        """The real multiprocessing target to redirect stdout and stderr to a pipe or queue."""
        sys.stdout.write = cls.redirect_write(pipe_stdout)  # , sys.__stdout__)  # Is redirected in main process
        sys.stderr.write = cls.redirect_write(pipe_stderr)  # , sys.__stderr__)  # Is redirected in main process

        pipe_target(*args, **kwargs)

    @staticmethod
    def redirect_write(child, out=None):
        """Create a function to write out a pipe and write out an additional out."""
        if isinstance(child, mp.queues.Queue):
            send = child.put
        else:
            send = child.send_bytes  # No need to pickle with child_conn.send(data)

        def write(data, *args):
            try:
                if isinstance(data, str):
                    data = data.encode('utf-8')

                send(data)
                if out is not None:
                    out.write(data)
            except:
                pass
        return write

    @classmethod
    def read_pipe_out(cls, pipe_alive, pipe_out, out):
        if isinstance(pipe_out, mp.queues.Queue):
            # Queue has better functionality to get all of the data
            def recv():
                return pipe_out.get(timeout=0.5)

            def is_alive():
                return pipe_alive.is_set() or pipe_out.qsize() > 0
        else:
            # Pipe is more efficient
            recv = pipe_out.recv_bytes  # No need to unpickle with data = pipe_out.recv()
            is_alive = pipe_alive.is_set

        # Loop through reading and redirecting data
        while is_alive():
            try:
                data = recv()
                if isinstance(data, bytes):
                    data = data.decode('utf-8')
                out.write(data)
            except EOFError:
                break
            except Empty:
                pass
            except:
                pass

    def join(self, *args):
        # Wait for process to finish (unless a timeout was given)
        super(PipeProcess, self).join(*args)

        # Trigger to stop the threads
        self.pipe_alive.clear()

        # Pipe must close to prevent blocking and waiting on recv forever
        if not isinstance(self.parent_stdout, mp.queues.Queue):
            with contextlib.suppress():
                self.parent_stdout.close()
            with contextlib.suppress():
                self.parent_stderr.close()

        # Close the pipes and threads
        with contextlib.suppress():
            self.read_out_th.join()
        with contextlib.suppress():
            self.read_err_th.join()

def run_long_print():
    for i in range(1000):
        print(i)
        print(i, file=sys.stderr)

    print('finished')


if __name__ == '__main__':
    # Example test write (My case was a QTextEdit)
    out = open('stdout.log', 'w')
    err = open('stderr.log', 'w')

    # Overwrite the write function and not the actual stdout object to prove this works
    sys.stdout.write = out.write
    sys.stderr.write = err.write

    # Create a process that uses pipes to read multiprocess output back into sys.stdout.write
    proc = PipeProcess(target=run_long_print, use_queue=True)  # If use_pipe=True Pipe may not write out all values
    # proc.daemon = True  # If daemon and use_queue Not all output may be redirected to stdout
    proc.start()

    # time.sleep(5)  # Not needed unless use_pipe or daemon and all of stdout/stderr is desired

    # Close the process
    proc.join()  # For some odd reason this blocks forever when use_queue=False

    # Close the output files for this test
    out.close()
    err.close()

Answer 4

Here is the simple and straightforward way for capturing stdout for multiprocessing.Process :

import app
import io
import sys
from multiprocessing import Process


def run_app(some_param):
    sys.stdout = io.TextIOWrapper(open(sys.stdout.fileno(), 'wb', 0), write_through=True)
    app.run()

app_process = Process(target=run_app, args=('some_param',))
app_process.start()
# Use app_process.termninate() for python <= 3.7.
app_process.kill() 

Answer 5

Alternatively - and I know this might be slightly off-topic, but helped in my case for the same problem -, this can be resolved with screen on Linux:

screen -L -Logfile './logfile_%Y-%m-%d.log' python my_multiproc_script.py

this way no need to implement all the master-child communication

Answer 6

I assume I'm off base and missing something, but for what it's worth here is what came to mind when I read your question.

If you can intercept all of the stdout and stderr (I got that impression from your question), then why not add or wrap that capture functionality around each of your processes? Then send what is captured through a queue to a consumer that can do whatever you want with all of the outputs?