下面是我的代码，用于捕获子进程ASAP的每个输出，包括部分行。它同时泵浦，stdout和stderr的顺序几乎是正确的。

在Python 2.7 linux和windows上测试并正确工作。

#!/usr/bin/python
#
# Runner with stdout/stderr catcher
#
from sys import argv
from subprocess import Popen, PIPE
import os, io
from threading import Thread
import Queue
def __main__():
    if (len(argv) > 1) and (argv[-1] == "-sub-"):
        import time, sys
        print "Application runned!"
        time.sleep(2)
        print "Slept 2 second"
        time.sleep(1)
        print "Slept 1 additional second",
        time.sleep(2)
        sys.stderr.write("Stderr output after 5 seconds")
        print "Eol on stdin"
        sys.stderr.write("Eol on stderr\n")
        time.sleep(1)
        print "Wow, we have end of work!",
    else:
        os.environ["PYTHONUNBUFFERED"]="1"
        try:
            p = Popen( argv + ["-sub-"],
                       bufsize=0, # line-buffered
                       stdin=PIPE, stdout=PIPE, stderr=PIPE )
        except WindowsError, W:
            if W.winerror==193:
                p = Popen( argv + ["-sub-"],
                           shell=True, # Try to run via shell
                           bufsize=0, # line-buffered
                           stdin=PIPE, stdout=PIPE, stderr=PIPE )
            else:
                raise
        inp = Queue.Queue()
        sout = io.open(p.stdout.fileno(), 'rb', closefd=False)
        serr = io.open(p.stderr.fileno(), 'rb', closefd=False)
        def Pump(stream, category):
            queue = Queue.Queue()
            def rdr():
                while True:
                    buf = stream.read1(8192)
                    if len(buf)>0:
                        queue.put( buf )
                    else:
                        queue.put( None )
                        return
            def clct():
                active = True
                while active:
                    r = queue.get()
                    try:
                        while True:
                            r1 = queue.get(timeout=0.005)
                            if r1 is None:
                                active = False
                                break
                            else:
                                r += r1
                    except Queue.Empty:
                        pass
                    inp.put( (category, r) )
            for tgt in [rdr, clct]:
                th = Thread(target=tgt)
                th.setDaemon(True)
                th.start()
        Pump(sout, 'stdout')
        Pump(serr, 'stderr')

        while p.poll() is None:
            # App still working
            try:
                chan,line = inp.get(timeout = 1.0)
                if chan=='stdout':
                    print "STDOUT>>", line, "<?<"
                elif chan=='stderr':
                    print " ERROR==", line, "=?="
            except Queue.Empty:
                pass
        print "Finish"

if __name__ == '__main__':
    __main__()

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()

我有最初提问者的问题，但不希望调用线程。我将Jesse的解决方案与来自管道的直接read()和我自己的行读取缓冲处理程序混合在一起(然而，我的子进程- ping -总是写入<系统页面大小的整行)。我通过读入一个gobject-registered io watch来避免忙碌的等待。这些天我通常在gobject MainLoop中运行代码以避免线程。

def set_up_ping(ip, w):
    # run the sub-process
    # watch the resultant pipe
    p = subprocess.Popen(['/bin/ping', ip], stdout=subprocess.PIPE)
    # make stdout a non-blocking file
    fl = fcntl.fcntl(p.stdout, fcntl.F_GETFL)
    fcntl.fcntl(p.stdout, fcntl.F_SETFL, fl | os.O_NONBLOCK)
    stdout_gid = gobject.io_add_watch(p.stdout, gobject.IO_IN, w)
    return stdout_gid # for shutting down

观察者是

def watch(f, *other):
    print 'reading',f.read()
    return True

主程序建立一个ping，然后调用gobject邮件循环。

def main():
    set_up_ping('192.168.1.8', watch)
    # discard gid as unused here
    gobject.MainLoop().run()

任何其他工作都附加到gobject中的回调。

选择模块帮助您确定下一个有用的输入在哪里。

但是，您几乎总是更喜欢使用单独的线程。一个是阻塞读取stdin，另一个是阻塞你不想阻塞的地方。

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)

现有的解决方案不适合我(详情见下文)。最后成功的是使用read(1)实现readline(基于这个答案)。后者不阻塞:

from subprocess import Popen, PIPE
from threading import Thread
def process_output(myprocess): #output-consuming thread
    nextline = None
    buf = ''
    while True:
        #--- extract line using read(1)
        out = myprocess.stdout.read(1)
        if out == '' and myprocess.poll() != None: break
        if out != '':
            buf += out
            if out == '\n':
                nextline = buf
                buf = ''
        if not nextline: continue
        line = nextline
        nextline = None

        #--- do whatever you want with line here
        print 'Line is:', line
    myprocess.stdout.close()

myprocess = Popen('myprogram.exe', stdout=PIPE) #output-producing process
p1 = Thread(target=process_output, args=(myprocess,)) #output-consuming thread
p1.daemon = True
p1.start()

#--- do whatever here and then kill process and thread if needed
if myprocess.poll() == None: #kill process; will automatically stop thread
    myprocess.kill()
    myprocess.wait()
if p1 and p1.is_alive(): #wait for thread to finish
    p1.join()

为什么现有的解决方案不起作用:

Solutions that require readline (including the Queue based ones) always block. It is difficult (impossible?) to kill the thread that executes readline. It only gets killed when the process that created it finishes, but not when the output-producing process is killed. Mixing low-level fcntl with high-level readline calls may not work properly as anonnn has pointed out. Using select.poll() is neat, but doesn't work on Windows according to python docs. Using third-party libraries seems overkill for this task and adds additional dependencies.