在我的例子中，我需要一个日志模块来捕获后台应用程序的输出并对其进行扩充(添加时间戳、颜色等)。

我最终使用了一个后台线程来执行实际的I/O操作。以下代码仅适用于POSIX平台。我去掉了不重要的部分。

如果有人打算长期使用这个野兽，可以考虑管理开放描述符。对我来说，这不是什么大问题。

# -*- python -*-
import fcntl
import threading
import sys, os, errno
import subprocess

class Logger(threading.Thread):
    def __init__(self, *modules):
        threading.Thread.__init__(self)
        try:
            from select import epoll, EPOLLIN
            self.__poll = epoll()
            self.__evt = EPOLLIN
            self.__to = -1
        except:
            from select import poll, POLLIN
            print 'epoll is not available'
            self.__poll = poll()
            self.__evt = POLLIN
            self.__to = 100
        self.__fds = {}
        self.daemon = True
        self.start()

    def run(self):
        while True:
            events = self.__poll.poll(self.__to)
            for fd, ev in events:
                if (ev&self.__evt) != self.__evt:
                    continue
                try:
                    self.__fds[fd].run()
                except Exception, e:
                    print e

    def add(self, fd, log):
        assert not self.__fds.has_key(fd)
        self.__fds[fd] = log
        self.__poll.register(fd, self.__evt)

class log:
    logger = Logger()

    def __init__(self, name):
        self.__name = name
        self.__piped = False

    def fileno(self):
        if self.__piped:
            return self.write
        self.read, self.write = os.pipe()
        fl = fcntl.fcntl(self.read, fcntl.F_GETFL)
        fcntl.fcntl(self.read, fcntl.F_SETFL, fl | os.O_NONBLOCK)
        self.fdRead = os.fdopen(self.read)
        self.logger.add(self.read, self)
        self.__piped = True
        return self.write

    def __run(self, line):
        self.chat(line, nl=False)

    def run(self):
        while True:
            try: line = self.fdRead.readline()
            except IOError, exc:
                if exc.errno == errno.EAGAIN:
                    return
                raise
            self.__run(line)

    def chat(self, line, nl=True):
        if nl: nl = '\n'
        else: nl = ''
        sys.stdout.write('[%s] %s%s' % (self.__name, line, nl))

def system(command, param=[], cwd=None, env=None, input=None, output=None):
    args = [command] + param
    p = subprocess.Popen(args, cwd=cwd, stdout=output, stderr=output, stdin=input, env=env, bufsize=0)
    p.wait()

ls = log('ls')
ls.chat('go')
system("ls", ['-l', '/'], output=ls)

date = log('date')
date.chat('go')
system("date", output=date)

Python 3.4为异步IO引入了新的临时API——asyncio模块。

该方法类似于@Bryan Ward的twisted-based回答——定义一个协议，一旦数据准备好，就调用它的方法:

#!/usr/bin/env python3
import asyncio
import os

class SubprocessProtocol(asyncio.SubprocessProtocol):
    def pipe_data_received(self, fd, data):
        if fd == 1: # got stdout data (bytes)
            print(data)

    def connection_lost(self, exc):
        loop.stop() # end loop.run_forever()

if os.name == 'nt':
    loop = asyncio.ProactorEventLoop() # for subprocess' pipes on Windows
    asyncio.set_event_loop(loop)
else:
    loop = asyncio.get_event_loop()
try:
    loop.run_until_complete(loop.subprocess_exec(SubprocessProtocol, 
        "myprogram.exe", "arg1", "arg2"))
    loop.run_forever()
finally:
    loop.close()

请参阅文档中的“Subprocess”。

有一个高级接口asyncio.create_subprocess_exec()，它返回允许使用StreamReader.readline()协程异步读取一行的Process对象 (使用async/await Python 3.5+语法):

#!/usr/bin/env python3.5
import asyncio
import locale
import sys
from asyncio.subprocess import PIPE
from contextlib import closing

async def readline_and_kill(*args):
    # start child process
    process = await asyncio.create_subprocess_exec(*args, stdout=PIPE)

    # read line (sequence of bytes ending with b'\n') asynchronously
    async for line in process.stdout:
        print("got line:", line.decode(locale.getpreferredencoding(False)))
        break
    process.kill()
    return await process.wait() # wait for the child process to exit


if sys.platform == "win32":
    loop = asyncio.ProactorEventLoop()
    asyncio.set_event_loop(loop)
else:
    loop = asyncio.get_event_loop()

with closing(loop):
    sys.exit(loop.run_until_complete(readline_and_kill(
        "myprogram.exe", "arg1", "arg2")))

Readline_and_kill()执行以下任务:

启动子进程，将其标准输出重定向到管道 异步从子进程的stdout中读取一行 杀子流程 等待它退出

如果需要，每个步骤都可以被超时秒限制。

一种解决方案是让另一个进程执行对该进程的读取，或者创建一个带有超时的进程线程。

这是一个超时函数的线程版本:

http://code.activestate.com/recipes/473878/

但是，是否需要在stdout传入时读取它? 另一种解决方案可能是将输出转储到一个文件中，然后使用p.t wait()等待进程完成。

f = open('myprogram_output.txt','w')
p = subprocess.Popen('myprogram.exe', stdout=f)
p.wait()
f.close()


str = open('myprogram_output.txt','r').read()

You can do this really easily in Twisted. Depending upon your existing code base, this might not be that easy to use, but if you are building a twisted application, then things like this become almost trivial. You create a ProcessProtocol class, and override the outReceived() method. Twisted (depending upon the reactor used) is usually just a big select() loop with callbacks installed to handle data from different file descriptors (often network sockets). So the outReceived() method is simply installing a callback for handling data coming from STDOUT. A simple example demonstrating this behavior is as follows:

from twisted.internet import protocol, reactor

class MyProcessProtocol(protocol.ProcessProtocol):

    def outReceived(self, data):
        print data

proc = MyProcessProtocol()
reactor.spawnProcess(proc, './myprogram', ['./myprogram', 'arg1', 'arg2', 'arg3'])
reactor.run()

Twisted文档在这方面有一些很好的信息。

如果您围绕Twisted构建整个应用程序，它可以与其他进程(本地或远程)进行异步通信，就像这样非常优雅。另一方面，如果您的程序不是构建在Twisted之上，那么这真的不会有多大帮助。希望这能对其他读者有所帮助，即使它不适用于您的特定应用程序。

我基于J. F. Sebastian的解决方案创建了一个库。你可以使用它。

https://github.com/cenkalti/what

在这里添加这个答案，因为它提供了在Windows和Unix上设置非阻塞管道的能力。

所有的ctypes细节都要感谢@techtonik的回答。

有一个稍微修改过的版本，可以在Unix和Windows系统上使用。

兼容Python3(只需要一些小的修改)。 包括posix版本，并定义了用于这两个版本的异常。

这样你就可以在Unix和Windows代码中使用相同的函数和异常。

# pipe_non_blocking.py (module)
"""
Example use:

    p = subprocess.Popen(
            command,
            stdout=subprocess.PIPE,
            )

    pipe_non_blocking_set(p.stdout.fileno())

    try:
        data = os.read(p.stdout.fileno(), 1)
    except PortableBlockingIOError as ex:
        if not pipe_non_blocking_is_error_blocking(ex):
            raise ex
"""


__all__ = (
    "pipe_non_blocking_set",
    "pipe_non_blocking_is_error_blocking",
    "PortableBlockingIOError",
    )

import os


if os.name == "nt":
    def pipe_non_blocking_set(fd):
        # Constant could define globally but avoid polluting the name-space
        # thanks to: https://stackoverflow.com/questions/34504970
        import msvcrt

        from ctypes import windll, byref, wintypes, WinError, POINTER
        from ctypes.wintypes import HANDLE, DWORD, BOOL

        LPDWORD = POINTER(DWORD)

        PIPE_NOWAIT = wintypes.DWORD(0x00000001)

        def pipe_no_wait(pipefd):
            SetNamedPipeHandleState = windll.kernel32.SetNamedPipeHandleState
            SetNamedPipeHandleState.argtypes = [HANDLE, LPDWORD, LPDWORD, LPDWORD]
            SetNamedPipeHandleState.restype = BOOL

            h = msvcrt.get_osfhandle(pipefd)

            res = windll.kernel32.SetNamedPipeHandleState(h, byref(PIPE_NOWAIT), None, None)
            if res == 0:
                print(WinError())
                return False
            return True

        return pipe_no_wait(fd)

    def pipe_non_blocking_is_error_blocking(ex):
        if not isinstance(ex, PortableBlockingIOError):
            return False
        from ctypes import GetLastError
        ERROR_NO_DATA = 232

        return (GetLastError() == ERROR_NO_DATA)

    PortableBlockingIOError = OSError
else:
    def pipe_non_blocking_set(fd):
        import fcntl
        fl = fcntl.fcntl(fd, fcntl.F_GETFL)
        fcntl.fcntl(fd, fcntl.F_SETFL, fl | os.O_NONBLOCK)
        return True

    def pipe_non_blocking_is_error_blocking(ex):
        if not isinstance(ex, PortableBlockingIOError):
            return False
        return True

    PortableBlockingIOError = BlockingIOError

为了避免读取不完整的数据，我最终编写了自己的readline生成器(它为每一行返回字节字符串)。

它是一个发电机，所以你可以例如…

def non_blocking_readlines(f, chunk=1024):
    """
    Iterate over lines, yielding b'' when nothings left
    or when new data is not yet available.

    stdout_iter = iter(non_blocking_readlines(process.stdout))

    line = next(stdout_iter)  # will be a line or b''.
    """
    import os

    from .pipe_non_blocking import (
            pipe_non_blocking_set,
            pipe_non_blocking_is_error_blocking,
            PortableBlockingIOError,
            )

    fd = f.fileno()
    pipe_non_blocking_set(fd)

    blocks = []

    while True:
        try:
            data = os.read(fd, chunk)
            if not data:
                # case were reading finishes with no trailing newline
                yield b''.join(blocks)
                blocks.clear()
        except PortableBlockingIOError as ex:
            if not pipe_non_blocking_is_error_blocking(ex):
                raise ex

            yield b''
            continue

        while True:
            n = data.find(b'\n')
            if n == -1:
                break

            yield b''.join(blocks) + data[:n + 1]
            data = data[n + 1:]
            blocks.clear()
        blocks.append(data)