Using Processes

  1. Overview
  2. Running Another Process
  3. Writing a ProcessProtocol
  4. Things that can happen to your ProcessProtocol
  5. Things you can do from your ProcessProtocol
  6. Verbose Example
  7. Doing it the Easy Way
  8. Mapping File Descriptors

Overview

Along with connection to servers across the internet, Twisted also connects to local processes with much the same API. The API is described in more detail in the documentation of:

Running Another Process

Processes are run through the reactor, using reactor.spawnProcess(). Pipes are created to the child process, and added to the reactor core so that the application will not block while sending data into or pulling data out of the new process. reactor.spawnProcess() requires two arguments, processProtocol and executable, and optionally takes six more: arguments, environment, path, userID, groupID, and usePTY.

from twisted.internet import reactor

mypp = MyProcessProtocol()
reactor.spawnProcess(processProtocol, executable, args=[program, arg1, arg2],
                     env={'HOME': os.environ['HOME']}, path,
                     uid, gid, usePTY, childFDs)

args and env have empty default values, but many programs depend upon them to be set correctly. At the very least, args[0] should probably be the same as executable. If you just provide os.environ for env, the child program will inherit the environment from the current process, which is usually the civilized thing to do (unless you want to explicitly clean the environment as a security precaution). The default is to give an empty env to the child.

reactor.spawnProcess() returns an instance that implements the twisted.internet.interfaces.IProcessTransport.

Writing a ProcessProtocol

The ProcessProtocol you pass to spawnProcess is your interaction with the process. It has a very similar signature to a regular Protocol, but it has several extra methods to deal with events specific to a process. In our example, we will interface with 'wc' to create a word count of user-given text. First, we'll start by importing the required modules, and writing the initialization for our ProcessProtocol.

from twisted.internet import protocol
class WCProcessProtocol(protocol.ProcessProtocol):

    def __init__(self, text):
        self.text = text

When the ProcessProtocol is connected to the protocol, it has the connectionMade method called. In our protocol, we will write our text to the standard input of our process and then close standard input, to the let the process know we are done writing to it.

def connectionMade(self):
        self.transport.write(self.text)
        self.transport.closeStdin()

At this point, the process has receieved the data, and it's time for us to read the results. Instead of being receieved in dataReceived, data from standard output is receieve in outReceived. This is to distinguish it from data on standard error.

def outReceived(self, data):
        fieldLength = len(data) / 3
        lines = int(data[:fieldLength])
        words = int(data[fieldLength:fieldLength*2])
        chars = int(data[fieldLength*2:])
        self.transport.loseConnection()
        self.receiveCounts(lines, words, chars)

Now, the process has parsed the output, and ended the connection to the process. Then it sends the results on to the final method, receiveCounts. This is for users of the class to override, so as to do other things with the data. For our demonstration, we will just print the results.

def receiveCounts(self, lines, words, chars):
        print 'Received counts from wc.'
        print 'Lines:', lines
        print 'Words:', words
        print 'Characters:', chars

We're done! To use our WCProcessProtocol, we create an instance, and pass it to spawnProcess.

from twisted.internet import reactor
wcProcess = WCProcessProtocol("accessing protocols through Twisted is fun!\n")
reactor.spawnProcess(wcProcess, 'wc', ['wc'])
reactor.run()

Things that can happen to your ProcessProtocol

These are the methods that you can usefully override in your subclass of ProcessProtocol:

The base-class definitions of these functions are all no-ops. This will result in all stdout and stderr being thrown away. Note that it is important for data you don't care about to be thrown away: if the pipe were not read, the child process would eventually block as it tried to write to a full pipe.

Things you can do from your ProcessProtocol

The following are the basic ways to control the child process:

Verbose Example

Here is an example that is rather verbose about exactly when all the methods are called. It writes a number of lines into the wc program and then parses the output.

#! /usr/bin/python

from twisted.internet import protocol
from twisted.internet import reactor
import re

class MyPP(protocol.ProcessProtocol):
    def __init__(self, verses):
        self.verses = verses
        self.data = ""
    def connectionMade(self):
        print "connectionMade!"
        for i in range(self.verses):
            self.transport.write("Aleph-null bottles of beer on the wall,\n" +
                                 "Aleph-null bottles of beer,\n" +
                                 "Take one down and pass it around,\n" +
                                 "Aleph-null bottles of beer on the wall.\n")
            self.transport.closeStdin() # tell them we're done
    def outReceived(self, data):
        print "outReceived! with %d bytes!" % len(data)
        self.data = self.data + data
    def errReceived(self, data):
        print "errReceived! with %d bytes!" % len(data)
    def inConnectionLost(self):
        print "inConnectionLost! stdin is closed! (we probably did it)"
    def outConnectionLost(self):
        print "outConnectionLost! The child closed their stdout!"
        # now is the time to examine what they wrote
        #print "I saw them write:", self.data
        (dummy, lines, words, chars, file) = re.split(r'\s+', self.data)
        print "I saw %s lines" % lines
    def errConnectionLost(self):
        print "errConnectionLost! The child closed their stderr."
    def processEnded(self, status_object):
        print "processEnded, status %d" % status_object.value.exitCode
        print "quitting"
        reactor.stop()

pp = MyPP(10)
reactor.spawnProcess(pp, "wc", ["wc"], {})
reactor.run()

The exact output of this program depends upon the relative timing of some un-synchronized events. In particular, the program may observe the child process close its stderr pipe before or after it reads data from the stdout pipe. One possible transcript would look like this:

% ./process.py 
connectionMade!
inConnectionLost! stdin is closed! (we probably did it)
errConnectionLost! The child closed their stderr.
outReceived! with 24 bytes!
outConnectionLost! The child closed their stdout!
I saw 40 lines
processEnded, status 0
quitting
Main loop terminated.
% 

Doing it the Easy Way

Frequently, one just needs a simple way to get all the output from a program. In the blocking world, you might use commands.getoutput from the standard library, but using that in an event-driven program will cause everything else to stall until the command finishes. (in addition, the SIGCHLD handler used by that function does not play well with Twisted's own signal handling). For these cases, the twisted.internet.utils.getProcessOutput function can be used. Here is a simple example:

from twisted.internet import protocol, utils, reactor
from twisted.python import failure
from cStringIO import StringIO

class FortuneQuoter(protocol.Protocol):

    fortune = '/usr/games/fortune'

    def connectionMade(self):
        output = utils.getProcessOutput(self.fortune)
        output.addCallbacks(self.writeResponse, self.noResponse)

    def writeResponse(self, resp):
        self.transport.write(resp)
        self.transport.loseConnection()

    def noResponse(self, err):
        self.transport.loseConnection()


if __name__ == '__main__':
    f = protocol.Factory()
    f.protocol = FortuneQuoter
    reactor.listenTCP(10999, f)
    reactor.run()

If you only need the final exit code (like commands.getstatusoutput(cmd)[0]), the twisted.internet.utils.getProcessValue function is useful. Here is an example:

from twisted.internet import utils, reactor

def printTrueValue(val):
    print "/bin/true exits with rc=%d" % val
    output = utils.getProcessValue('/bin/false')
    output.addCallback(printFalseValue)

def printFalseValue(val):
    print "/bin/false exits with rc=%d" % val
    reactor.stop()

output = utils.getProcessValue('/bin/true')
output.addCallback(printTrueValue)
reactor.run()

Mapping File Descriptors

stdin, stdout, and stderr are just conventions. Programs which operate as filters generally accept input on fd0, write their output on fd1, and emit error messages on fd2. This is common enough that the standard C library provides macros like stdin to mean fd0, and shells interpret the pipe character | to mean redirect fd1 from one command into fd0 of the next command.

But these are just conventions, and programs are free to use additional file descriptors or even ignore the standard three entirely. The childFDs argument allows you to specify exactly what kind of files descriptors the child process should be given.

Each child FD can be put into one of three states:

Mapping the child FDs to the parent's is very commonly used to send the child's stderr output to the same place as the parent's. When you run a program from the shell, it will typically leave fds 0, 1, and 2 mapped to the shell's 0, 1, and 2, allowing you to see the child program's output on the same terminal you used to launch the child. Likewise, inetd will typically map both stdin and stdout to the network socket, and may map stderr to the same socket or to some kind of logging mechanism. This allows the child program to be implemented with no knowledge of the network: it merely speaks its protocol by doing reads on fd0 and writes on fd1.

Feeding into a parent's read pipe is used to gather output from the child, and is by far the most common way of interacting with child processes.

Feeding from a parent's write pipe allows the parent to control the child. Programs like bc or ftp can be controlled this way, by writing commands into their stdin stream.

The childFDs dictionary maps file descriptor numbers (as will be seen by the child process) to one of these three states. To map the fd to one of the parent's fds, simply provide the fd number as the value. To map it to a read pipe, use the string r as the value. To map it to a write pipe, use the string w.

For example, the default mapping sets up the standard stdin/stdout/stderr pipes. It is implemented with the following dictionary:

childFDs = { 0: "w", 1: "r", 2: "r" }

To launch a process which reads and writes to the same places that the parent python program does, use this:

childFDs = { 0: 0, 1: 1, 2: 2}

To write into an additional fd (say it is fd number 4), use this:

childFDs = { 0: "w", 1: "r", 2: "r" , 4: "w"}

ProcessProtocols with extra file descriptors

When you provide a childFDs dictionary with more than the normal three fds, you need addtional methods to access those pipes. These methods are more generalized than the .outReceived ones described above. In fact, those methods (outReceived and errReceived) are actually just wrappers left in for compatibility with older code, written before this generalized fd mapping was implemented. The new list of things that can happen to your ProcessProtocol is as follows:

In addition to those methods, there are other methods available to influence the child process:

Examples

GnuPG, the encryption program, can use additional file descriptors to accept a passphrase and emit status output. These are distinct from stdin (used to accept the crypttext), stdout (used to emit the plaintext), and stderr (used to emit human-readable status/warning messages). The passphrase FD reads until the pipe is closed and uses the resulting string to unlock the secret key that performs the actual decryption. The status FD emits machine-parseable status messages to indicate the validity of the signature, which key the message was encrypted to, etc.

gpg accepts command-line arguments to specify what these fds are, and then assumes that they have been opened by the parent before the gpg process is started. It simply performs reads and writes to these fd numbers.

To invoke gpg in decryption/verification mode, you would do something like the following:

class GPGProtocol(ProcessProtocol):
    def __init__(self, crypttext):
        self.crypttext = crypttext
        self.plaintext = ""
        self.status = ""
    def connectionMade(self):
        self.transport.writeToChild(3, self.passphrase)
        self.transport.closeChildFD(3)
        self.transport.writeToChild(0, self.crypttext)
        self.transport.closeChildFD(0)
    def childDataReceived(self, childFD, data):
        if childFD == 1: self.plaintext += data
        if childFD == 4: self.status += data
    def processEnded(self, status):
        rc = status.value.exitCode
        if rc == 0:
            self.deferred.callback(self)
        else:
            self.deferred.errback(rc)

def decrypt(crypttext):
    gp = GPGProtocol(crypttext)
    gp.deferred = Deferred()
    cmd = ["gpg", "--decrypt", "--passphrase-fd", "3", "--status-fd", "4",
           "--batch"]
    p = reactor.spawnProcess(gp, cmd[0], cmd, env=None,
                             childFDs={0:"w", 1:"r", 2:2, 3:"w", 4:"r"})
    return gp.deferred

In this example, the status output could be parsed after the fact. It could, of course, be parsed on the fly, as it is a simple line-oriented protocol. Methods from LineReceiver could be mixed in to make this parsing more convenient.

The stderr mapping (2:2) used will cause any GPG errors to be emitted by the parent program, just as if those errors had caused in the parent itself. This is sometimes desireable (it roughly corresponds to letting exceptions propagate upwards), especially if you do not expect to encounter errors in the child process and want them to be more visible to the end user. The alternative is to map stderr to a read-pipe and handle any such output from within the ProcessProtocol (roughly corresponding to catching the exception locally).

Index

Version: 2.5.0