如何非阻塞地读取subprocess.Popen.stdout中的所有可用数据？

Question

我需要一种方法，可以读取通过 Popen 创建的流中所有当前可用的字符，或者找出缓冲区中还有多少字符。

背景： 我想用 Python 远程控制一个交互式应用程序。到目前为止，我使用 Popen 创建了一个新的子进程：

process=subprocess.Popen(["python"],shell=True,stdin=subprocess.PIPE,stdout=subprocess.PIPE,stderr=subprocess.PIPE, cwd=workingDir)

（我其实并不是在启动 Python，但实际的交互界面是类似的。） 现在我每次读取 1 个字节，直到我检测到进程已经到达命令提示符：

output = ""
while output[-6:]!="SCIP> ":
    output += process.stdout.read(1)
    sys.stdout.write(output[-1])
return output

然后我通过 process.stdin.write("command\n") 开始一个耗时的计算。 我的问题是，我无法检查计算是否完成，因为我无法判断流中的最后几个字符是否是提示符。read() 或 read(n) 会阻塞我的线程，直到达到文件结束符（EOF），但它永远不会到达，因为这个交互程序不会自己结束，直到我告诉它结束。像上面那样循环查找提示符也不行，因为提示符只会在计算完成后出现。

理想的解决方案是让我能够读取流中所有可用的字符，如果没有可读内容，则立即返回一个空字符串。

Answer 1

我尝试了很多方法，比如通过以下方式来创建一个非阻塞的 stdout：

fd = output.fileno()
fl = fcntl.fcntl(fd, fcntl.F_GETFL)
fcntl.fcntl(fd, fcntl.F_SETFL, fl | os.O_NONBLOCK)

但是唯一有效的解决方案在 这里 有描述：

master, slave = pty.openpty()

proc = subprocess.Popen(
    shlex.split(command), 
    stdout=slave, 
    stderr=slave, 
    close_fds=True, 
    bufsize=0
)

stdout = os.fdopen(master)

然后：

while True:
    out = stdout.readline()
    output_result = proc.poll()
    if out == '' and output_result is not None:
        break
    if out != '':
        print(out)

Answer 2

逐步解析Popen的标准输出其实并不复杂。你只需要在一个线程中插入一个管道，让它去处理输出，寻找分隔符。根据你的需求，可以把处理后的内容放到另一个管道或文件中，或者把解析出来的“块”放到“栈”里，以异步的方式进行处理。下面是一个基于自定义分隔符的标准输出异步“分块”的例子：

import cStringIO
import uuid
import threading
import os

class InputStreamChunker(threading.Thread):
    '''
    Threaded object / code that mediates reading output from a stream,
    detects "separation markers" in the stream and spits out chunks
    of original stream, split when ends of chunk are encountered.

    Results are made available as a list of filled file-like objects
    (your choice). Results are accessible either "asynchronously"
    (you can poll at will for results in a non-blocking way) or
    "synchronously" by exposing a "subscribe and wait" system based
    on threading.Event flags.

    Usage:
    - instantiate this object
    - give our input pipe as "stdout" to other subprocess and start it:
        Popen(..., stdout = th.input, ...)
    - (optional) subscribe to data_available event
    - pull resulting file-like objects off .data
      (if you are "messing" with .data from outside of the thread,
       be curteous and wrap the thread-unsafe manipulations between:
       obj.data_unoccupied.clear()
       ... mess with .data
       obj.data_unoccupied.set()
       The thread will not touch obj.data for the duration and will
       block reading.)

    License: Public domain
    Absolutely no warranty provided
    '''
    def __init__(self, delimiter = None, outputObjConstructor = None):
        '''
        delimiter - the string that will be considered a delimiter for the stream
        outputObjConstructor - instanses of these will be attached to self.data array
         (intantiator_pointer, args, kw)
        '''
        super(InputStreamChunker,self).__init__()

        self._data_available = threading.Event()
        self._data_available.clear() # parent will .wait() on this for results.
        self._data = []
        self._data_unoccupied = threading.Event()
        self._data_unoccupied.set() # parent will set this to true when self.results is being changed from outside
        self._r, self._w = os.pipe() # takes all inputs. self.input = public pipe in.
        self._stop = False
        if not delimiter: delimiter = str(uuid.uuid1())
        self._stream_delimiter = [l for l in delimiter]
        self._stream_roll_back_len = ( len(delimiter)-1 ) * -1
        if not outputObjConstructor:
            self._obj = (cStringIO.StringIO, (), {})
        else:
            self._obj = outputObjConstructor
    @property
    def data_available(self):
        '''returns a threading.Event instance pointer that is
        True (and non-blocking to .wait() ) when we attached a
        new IO obj to the .data array.
        Code consuming the array may decide to set it back to False
        if it's done with all chunks and wants to be blocked on .wait()'''
        return self._data_available
    @property
    def data_unoccupied(self):
        '''returns a threading.Event instance pointer that is normally
        True (and non-blocking to .wait() ) Set it to False with .clear()
        before you start non-thread-safe manipulations (changing) .data
        array. Set it back to True with .set() when you are done'''
        return self._data_unoccupied
    @property
    def data(self):
        '''returns a list of input chunkes (file-like objects) captured
        so far. This is a "stack" of sorts. Code consuming the chunks
        would be responsible for disposing of the file-like objects.
        By default, the file-like objects are instances of cStringIO'''
        return self._data
    @property
    def input(self):
        '''This is a file descriptor (not a file-like).
        It's the input end of our pipe which you give to other process
        to be used as stdout pipe for that process'''
        return self._w
    def flush(self):
        '''Normally a read on a pipe is blocking.
        To get things moving (make the subprocess yield the buffer,
        we inject our chunk delimiter into self.input

        This is useful when primary subprocess does not write anything
        to our in pipe, but we need to make internal pipe reader let go
        of the pipe and move on with things.
        '''
        os.write(self._w, ''.join(self._stream_delimiter))
    def stop(self):
        self._stop = True
        self.flush() # reader has its teeth on the pipe. This makes it let go for for a sec.
        os.close(self._w)
        self._data_available.set()
    def __del__(self):
        try:
            self.stop()
        except:
            pass
        try:
            del self._w
            del self._r
            del self._data
        except:
            pass
    def run(self):
        ''' Plan:
        - We read into a fresh instance of IO obj until marker encountered.
        - When marker is detected, we attach that IO obj to "results" array
          and signal the calling code (through threading.Event flag) that
          results are available
        - repeat until .stop() was called on the thread.
        '''
        marker = ['' for l in self._stream_delimiter] # '' is there on purpose
        tf = self._obj[0](*self._obj[1], **self._obj[2])
        while not self._stop:
            l = os.read(self._r, 1)
            print('Thread talking: Ordinal of char is:%s' %ord(l))
            trash_str = marker.pop(0)
            marker.append(l)
            if marker != self._stream_delimiter:
                tf.write(l)
            else:
                # chopping off the marker first
                tf.seek(self._stream_roll_back_len, 2)
                tf.truncate()
                tf.seek(0)
                self._data_unoccupied.wait(5) # seriously, how much time is needed to get your items off the stack?
                self._data.append(tf)
                self._data_available.set()
                tf = self._obj[0](*self._obj[1], **self._obj[2])
        os.close(self._r)
        tf.close()
        del tf

def waitforresults(ch, answers, expect):
    while len(answers) < expect:
        ch.data_available.wait(0.5); ch.data_unoccupied.clear()
        while ch.data:
            answers.append(ch.data.pop(0))
        ch.data_available.clear(); ch.data_unoccupied.set()
        print('Main talking: %s answers received, expecting %s\n' % ( len(answers), expect) )

def test():
    '''
    - set up chunker
    - set up Popen with chunker's output stream
    - push some data into proc.stdin
    - get results
    - cleanup
    '''

    import subprocess

    ch = InputStreamChunker('\n')
    ch.daemon = True
    ch.start()

    print('starting the subprocess\n')
    p = subprocess.Popen(
        ['cat'],
        stdin = subprocess.PIPE,
        stdout = ch.input,
        stderr = subprocess.PIPE)

    answers = []

    i = p.stdin
    i.write('line1 qwer\n') # will be in results
    i.write('line2 qwer\n') # will be in results
    i.write('line3 zxcv asdf') # will be in results only after a ch.flush(),
                                # prepended to other line or when the pipe is closed
    waitforresults(ch, answers, expect = 2)

    i.write('line4 tyui\n') # will be in results
    i.write('line5 hjkl\n') # will be in results
    i.write('line6 mnbv') # will be in results only after a ch.flush(),
                                # prepended to other line or when the pipe is closed
    waitforresults(ch, answers, expect = 4)

    ## now we will flush the rest of input (that last line did not have a delimiter)
    i.close()
    ch.flush()
    waitforresults(ch, answers, expect = 5)

    should_be = ['line1 qwer', 'line2 qwer',
        'line3 zxcv asdfline4 tyui', 'line5 hjkl', 'line6 mnbv']
    assert should_be == [i.read() for i in answers]

    # don't forget to stop the chunker. It it closes the pipes
    p.terminate()
    ch.stop()
    del p, ch

if __name__ == '__main__':
    test()

补充：删除了关于“写入进程的标准输入是一次性操作”的错误表述

Answer 3

我在网上找到了一个非常不错的解决方案。

持久化的Python子进程

这个方法通过使用 fcntl 来设置子进程管道的文件属性为非阻塞模式，完全避免了阻塞的问题，不需要额外的线程或轮询。我可能还有其他没想到的地方，但这个方法解决了我在交互式进程控制上的问题。